Automated segmentation of neurons and spinal cord structures in immunofluorescence images using SpineDL.

1. """
2. Configuration management for BiaPy.
4. This module defines the Config class, which encapsulates all configuration options
5. for BiaPy workflows using a YACS CfgNode. It provides default values and hierarchical
6. organization for system, data, augmentation, model, loss, training, inference,
7. post-processing, and logging parameters. Utility functions are included for updating
8. dependent configuration variables after merging user-provided configs.
9. """
10. import os
11. from yacs.config import CfgNode as CN
12. import copy
13. from typing import (
14. Dict,
15. )
17. class Config:
18. """
19. BiaPy configuration handler.
21. This class manages the hierarchical configuration for BiaPy experiments,
22. including system resources, problem specification, data loading, augmentation,
23. model architecture, loss functions, training, inference, post-processing, and
24. logging. It uses YACS CfgNode for flexible and robust configuration management.
26. Attributes
27. ----------
28. _C : CN
29. The root YACS configuration node containing all experiment parameters.
30. """
32. def __init__(self, job_dir: str, job_identifier: str):
33. """
34. Initialize the Config object with default values.
36. Parameters
37. ----------
38. job_dir : str
39. Directory where job outputs (results, checkpoints, logs) will be stored.
40. job_identifier : str
41. Unique identifier for the job (used in output paths).
42. """
43. if "/" in job_identifier:
44. raise ValueError("Job name can not contain / character. Provided: {}".format(job_identifier))
46. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
47. # Config definition
48. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
49. _C = CN()
51. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
52. # 1. System
53. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
54. _C.SYSTEM = CN()
55. # Maximum number of CPUs to use. Set it to "-1" to not set a limit.
56. _C.SYSTEM.NUM_CPUS = -1
57. # Maximum number of workers to use. You can disable this option by setting 0. With a -1 the workers are calculated automatically.
58. # They are the processes that prepare the data (reading, cropping and normalizing it) in parallel with the GPU work, one
59. # set of them per process, i.e. per GPU. In the "by chunks" inference each of them fills one tile of the workflow process
60. # at a time, so they multiply the RAM the tiles take (see 'TEST.BY_CHUNKS.WORKFLOW_PROCESS.PATCHES_PER_TILE'): lower this
61. # value, or raise it, to trade reading speed for the memory that bigger tiles need.
62. _C.SYSTEM.NUM_WORKERS = -1
63. # Do not set it as its value will be calculated based in --gpu input arg
64. _C.SYSTEM.NUM_GPUS = 0
65. # Device to be used when GPU is NOT selected. Most commonly "cpu", but also potentially "mps",
66. # "xpu", "xla" or "meta".
67. _C.SYSTEM.DEVICE = "cpu"
69. # Math seed to generate random numbers. Used to ensure reproducibility in the results.
70. _C.SYSTEM.SEED = 0
71. # Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.
72. _C.SYSTEM.PIN_MEM = True
74. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
75. # 2. Problem specification
76. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
77. _C.PROBLEM = CN()
78. # Whether to check what changed to adapt the given input configuration to the newest. Basically to now what
79. # biapy.engine.check_configuration file's convert_old_model_cfg_to_current_version() function changes
80. _C.PROBLEM.PRINT_OLD_KEY_CHANGES = True
81. # Possible options: 'SEMANTIC_SEG', 'INSTANCE_SEG', 'DETECTION', 'DENOISING', 'SUPER_RESOLUTION',
82. # 'SELF_SUPERVISED', 'CLASSIFICATION' and 'IMAGE_TO_IMAGE'
83. _C.PROBLEM.TYPE = "SEMANTIC_SEG"
84. # Possible options: '2D' and '3D'
85. _C.PROBLEM.NDIM = "2D"
87. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
88. # 2.1 Semantic segmentation problem specification
89. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
90. _C.PROBLEM.SEMANTIC_SEG = CN()
91. # Class id to ignore when DATA.N_CLASSES > 2
92. _C.PROBLEM.SEMANTIC_SEG.IGNORE_CLASS_ID = 0
94. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
95. # 2.2 Instance segmentation problem specification
96. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
97. _C.PROBLEM.INSTANCE_SEG = CN()
98. # Type of instances expected. Options are: ["regular", "synapses"]
99. _C.PROBLEM.INSTANCE_SEG.TYPE = "regular"
101. #### For "regular" type of instances ####
102. # This variable defines the channels to be used to represent the instances. The approach follows a bottom-up setting,
103. # where the intermediate representations of the instances are learned by the model and then fused to create the final
104. # instances. The variable needs to be a list with the representations to be extracted from the instances.
105. # These are the options available:
106. # - 'F' stands for 'Foreground', it is a binary representation of each instance
107. # - 'B' stands for 'Background', it is a binary representation of the background, i.e. everything that is not an instance.
108. # - 'P' stands for 'Central part', the extreme case of the F channel, where only the center of mass, or the skeleton is
109. # represented. It may also be expressed as a heatmap.
110. # - 'C' stands for 'Contour', it is a binary representation of the countours of each instance.
111. # - 'H' stands for 'Horizontal distance' (HoVer-Net style). For each instance, the pixel/voxel value is the signed
112. # horizontal displacement to the centroid. When 'norm' is True (default) values are rescaled to [-1, 1]: leftmost
113. # pixel of the cell = -1, rightmost = +1, centroid column = 0. Background pixels are always 0.
114. # - 'V' stands for 'Vertical distance' (HoVer-Net style), same as 'H' along the vertical axis:
115. # top pixel = -1, bottom pixel = +1, centroid row = 0. Background = 0.
116. # - 'Z' stands for 'Depth distance' (HoVer-Net style), same as 'H' along the depth (z) axis. Only relevant for 3D data.
117. # - 'Gh', 'Gv', and 'Gz' stand for the x/y/z components of a flow field derived from a heat-diffusion potential
118. # (Cellpose strategy) or the per-cell EDT gradient (Omnipose strategy). Each component is a unit vector in [-1, 1];
119. # background is (0, 0[, 0]). Inspired by "Cellpose: a generalist algorithm for cellular segmentation"
120. # (https://www.nature.com/articles/s41592-020-01018-x) and "Omnipose" (https://www.nature.com/articles/s41467-022-32267-2).
121. # - 'Db' stands for 'Distance to the boundary'. For each foreground pixel the value is its Euclidean distance to
122. # the nearest background pixel (i.e. intra-cell distance from the cell boundary inward). After per-cell normalization
123. # ('norm'=True), boundary pixels map to 0 and the innermost pixel maps to 1. Background is always 0.
124. # - 'Dc' stands for 'Distance to the center'. For each foreground pixel the value is its Euclidean distance to
125. # the centroid (or skeleton) of its cell. The centroid pixel therefore has Dc = 0 (the minimum, most important value).
126. # Background is also 0. When 'norm' is True, values are rescaled per-cell to [0, 1].
127. # - 'Dn' stands for 'Distance to the closest neighbor'. For each foreground pixel the value reflects the proximity to
128. # the nearest other instance; isolated cells (no neighbor) are assigned 0. Background is 0.
129. # - 'D' stands for 'signed Distance field'. Foreground pixels hold the positive Euclidean distance to the nearest
130. # background pixel; background pixels hold the negative distance to the nearest foreground pixel. The field spans
131. # from negative (deep background) through 0 (boundary) to positive (deep interior). When 'norm' is True (default)
132. # the field is passed through tanh, mapping all values to (-1, 1). Because both foreground (+) and background (-)
133. # carry meaningful signal, the loss is computed on all pixels without masking.
134. # - 'R' stands for 'Radial distance', distances to boundaries measure how far each pixel is from its object's edge along fixed
135. # radial directions, with StarDist being the most prominent approach.
136. # - 'T' stands for 'Touching area', where touching regions between instances are highlighted.
137. # - 'A' stands for 'Affinities' and contains the affinity values for each dimension. Affinities represent images as graphs for
138. # segmentation, where each voxel is a node connected to neighbors by edges indicating shared object membership. In 3D, nodes
139. # link along x-, y-, and z-axes, yielding multiple affinity channels, commonly used in connectomics.
140. # - 'E' stands for 'Embeddings' and contains the embedding vectors for each pixel/voxel. Embedding-based approaches map each
141. # pixel in an input image to a point in an n-dimensional feature space (typically n=24), referred to as the pixel embedding.
142. # The goal is to ensure that embeddings from the same instance are close together, while embeddings from different instances
143. # are well separated. We grouped here all the channels involved in this process
144. #
145. #### For "synapse" type of instances ####
146. # Possible options:
147. # * 'F_pre' and 'F_post': each one is a binary representation of the presynaptic and postsynaptic sites, respectively.
148. # * 'F_pre', 'H', 'V', 'Z': 'F_pre' is a binary representation of the presynaptic sites, while 'H', 'V' and 'Z' are the horizontal,
149. # vertical and depth distances to the closest postsynaptic site, respectively. This setting is inspired by the paper
150. # "Automatic detection of synaptic partners in a whole-brain Drosophila electron microscopy data set"
151. # (https://www.nature.com/articles/s41592-021-01183-7).
152. #
153. _C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS = ["B", "C"]
154. # Details for each channel. It must be a list with a unique element: a dict of dicts. The details can be only set for the following channels:
155. # - 'F' channel. Possible options:
156. # - 'erosion': int or list of ints, specifies the erosion size applied to the channel. Default: 0
157. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. Default: 0
158. # - 'B' channel. Possible options:
159. # - 'erosion': int or list of ints, specifies the erosion size applied to the channel. Default: 0
160. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. Default: 0
161. # - 'P' channel. Possible options:
162. # - 'type': str, specifies the reference used to define the channel. Options are 'centroid' or 'skeleton'. Default: 'centroid'
163. # - 'skeleton_mode': str, only used when 'type' is 'skeleton'. Controls how much of the skeleton is kept:
164. # - 'full': keep the whole skeleton, branching into every protrusion of the instance (default).
165. # - 'main': prune it down to the main body only, i.e. the longest path through the skeleton (its
166. # two farthest-apart points and the branch connecting them). Side branches coming from small
167. # protrusions (e.g. dendritic spines) are discarded, so the channel marks just the main shaft.
168. # Default: 'full'
169. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. Default: 1
170. # - 'erosion': int or list of ints, specifies the erosion size applied to the channel. Useful for thinning skeletons,
171. # which are typically 2 pixels thick. Default: 0
172. # - 'C' channel. Possible options:
173. # - 'mode': str, specifies how contours are generated. Corresponds to the 'mode' argument of the
174. # find_boundaries function from scikit-image.
175. # More information at: https://scikit-image.org/docs/stable/api/skimage.segmentation.html#skimage.segmentation.find_boundaries.
176. # Options are 'thick', 'inner', 'outer', 'subpixel', and 'dense'. The 'dense' option labels as contour
177. # every pixel that is not part of the foreground. Default: 'thick'
178. # - 'H', 'V' and 'Z' channels. Possible options:
179. # - 'norm': bool, specifies whether signed displacements are normalized to [-1, 1] (centroid = 0). Default: True
180. # - 'act': str, specifies the activation function used in the model's final layer when this channel is selected.
181. # These channels are signed, so the activation must allow negative values: 'linear' (unbounded) or 'tanh'
182. # (bounded to [-1, 1], only meaningful with 'norm'=True). Options are 'linear' and 'tanh'. Default: 'linear'
183. # Note: the loss for H/V/Z is computed over all pixels, not masked to the foreground, so the background
184. # target (0) also trains the model. That keeps the predictions bounded outside the instances, which the
185. # instance creation relies on (it min-max normalizes these channels over the whole image).
186. # - 'Gh', 'Gv' and 'Gz' channels. Possible options:
187. # - 'gradient_type': str, method to compute the gradients. Options are "cellpose" and "omnipose". Default: "cellpose"
188. # - 'Db' channel. Possible options:
189. # - 'val_type': str, to determine how to modify the distance values. Default: 'norm'. Options are:
190. # - 'raw': to leave the distances as they are calculated.
191. # - 'norm': to normalize the distances per cell to [0, 1] (boundary → 0, innermost pixel → 1).
192. # - 'discretize': to discretize the distances into bins.
193. # - 'act': str, activation function to be used in the last layer of the model when this channel is selected.
194. # Options are: '', 'linear' and 'sigmoid'. Default: ''.
195. # - 'mask_values': bool, kept for compatibility but superseded: the loss mask prefers a binary foreground channel
196. # (F/M/B) to avoid excluding boundary pixels (Db=0). Falls back to (Db > 0) if no binary channel is present. Default: True
197. # - 'Dc' channel. Possible options:
198. # - 'type': str, specifies the reference used to calculate the distance. Options are 'centroid' or 'skeleton'.
199. # Default: 'centroid'
200. # - 'norm': bool, specifies whether distances are normalized per cell to [0, 1]. Default: True
201. # - 'mask_values': bool, kept for compatibility but superseded: the loss mask prefers a binary foreground channel
202. # (F/M/B) to correctly include the centroid pixel (Dc=0). Falls back to (Dc > 0) if no binary channel is present,
203. # which still misses the centroid — prefer pairing Dc with F, M, or B. Default: True
204. # - 'Dn' channel. Possible options:
205. # - 'closing_size': int, specifies the size of the closing operation applied to the combined distance map.
206. # Default: 0
207. # - 'norm': bool, specifies whether distances are normalized between 0 and 1. Default: True
208. # - 'mask_values': bool, kept for compatibility but superseded: the loss mask prefers a binary foreground channel
209. # (F/M/B) to include isolated cells (Dn=0). Falls back to (Dn > 0) if no binary channel is present. Default: True
210. # - 'decline_power': int, specifies the power applied to distances to control the rate of decline. Default: 3
211. # - 'D' channel. Per-instance distance in [-1, 1]: each instance's distance to its own boundary,
212. # normalized per instance (boundary -> 0, innermost voxel -> 1), with the background set to -1.
213. # Measuring the distance per instance, instead of on the binary foreground, is what keeps
214. # touching instances apart, so each one keeps its own peak and they can be used as seeds.
215. # Possible options:
216. # - 'act': str, specifies the activation function used in the model's final layer when this channel is selected.
217. # Options are 'tanh' and 'linear'. Default: 'tanh'
218. # - 'R' channel. Possible options:
219. # - 'nrays': int, specifies the number of rays used to represent radial distances.
220. # Default: 32 for 2D data and 96 for 3D data
221. # - 'norm': bool, specifies whether distances are normalized between 0 and 1. Default: True
222. # - 'mask_values': bool, kept for compatibility but superseded: the loss mask prefers a binary foreground channel
223. # (F/M/B) to avoid excluding near-boundary ray pixels (R≈0). Falls back to (R > 0) if no binary channel is present. Default: True
224. # - 'T' channel. Possible options:
225. # - 'thickness': int, specifies the thickness (in pixels) of the touching area. Default: 2
226. # - 'A' channel. Possible options:
227. # - 'z_affinities': list of int, specifies offsets along the z axis. Default: [1]
228. # - 'y_affinities': list of int, specifies offsets along the y axis. Default: [1]
229. # - 'x_affinities': list of int, specifies offsets along the x axis. Default: [1]
230. # - 'widen_borders': int, specifies the number of pixels used to widen affinity borders. Default: 1
231. # Note that 'z_affinities', 'y_affinities', and 'x_affinities' must be lists of equal length: they are
232. # paired by index, so entry k of each describes one set of neighbour offsets. The resulting channels are
233. # stored as one (z, y, x) triple per offset index, i.e. interleaved (Az_1, Ay_1, Ax_1, Az_2, Ay_2, Ax_2,
234. # ...) rather than grouped by axis.
235. # To reproduce the paper "Superhuman Accuracy on the SNEMI3D Connectomics Challenge", use:
236. # 'z_affinities': [1, 2, 3, 4], 'y_affinities': [1, 3, 9, 27], and 'x_affinities': [1, 3, 9, 27]
237. # - 'E' channel. Possible options:
238. # - 'center_mode': str, specifies the center definition used for Gaussian seeds.
239. # Options are 'centroid' and 'medoid'. Default: 'centroid'
240. # - 'medoid_max_points': int, specifies the maximum number of points used when computing the medoid.
241. # Default: 10000
242. # For synapse type of instances, the options are:
243. # - 'F_pre' channel. Possible options:
244. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. Default: [1,3,3]
245. # - 'F_post' channel. Possible options:
246. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. Default: [1,3,3]
247. # - 'F_cleft' channel. Possible options:
248. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. Default: [1,3,3]
249. # - 'H', 'V' and 'Z' channels (synapse mode). Possible options:
250. # - 'norm': bool, if True signed displacements are normalized to [-1, 1]; if False raw physical distances are kept.
251. # Default: False (raw distances in physical units for synapse partner offsets)
252. # - 'act': str, specifies the activation function used in the model's final layer when this channel is selected.
253. # Options are '', 'linear', and 'sigmoid'. Default: ''
254. # - 'dilation': int or list of ints, specifies the dilation size applied to the channel. In the case of the synapses this
255. # will represent the area that will define the distance channels. Default: [3,25,25]
256. # Note: the loss mask for H/V/Z is derived automatically from a binary foreground channel (F_pre/F_post/F_cleft)
257. # when one is present; otherwise all pixels are included.
258. #
259. # For example:
260. # DATA_CHANNELS = ['F', 'C']
261. # DATA_CHANNELS_EXTRA_OPTS = [{'F': {'erosion': 2, 'dilation': 0}, 'C': {'mode': 'inner'}}]
262. _C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS_EXTRA_OPTS = [{}]
263. # Losses to be applied to each channel. If not provided, the losses will be set automatically depending on the channel.
264. # The options are:
265. # - "bce"/"ce": binary cross entropy. Ref: https://pytorch.org/docs/stable/generated/torch.nn.BCEWithLogitsLoss.html
266. # - "l1"/"mae": mean absolute error. Ref: https://pytorch.org/docs/stable/generated/torch.nn.L1Loss.html
267. # - "mse": mean squared error. Ref: https://pytorch.org/docs/stable/generated/torch.nn.MSELoss.html
268. # - "triplet": triplet loss. Ref: https://pytorch.org/docs/stable/generated/torch.nn.TripletMarginLoss.html
269. _C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS_LOSSES = []
270. # Whether to apply a rebalancing strategy to the loss function to give more importance to underrepresented pixels within the channels.
271. # The weights are calculated automatically based on the number of pixels of each class per batch and directly in the loss computation.
272. # For example, in the case of the "C" channel, which represents the contours of the instances, there are usually much less pixels representing
273. # contours than pixels representing the background. With this option activated, the loss will give more importance to contour pixels to help
274. # the model learn better to predict them.
275. _C.PROBLEM.INSTANCE_SEG.CLASS_REBALANCE_WITHIN_CHANNELS = True
276. # Information on how the channels are distributed in the model's output heads. It must be a list of list of ints, where each
277. # inner list contains the number of channels that are going to be predicted in the same head. If not provided, it will be set automatically
278. # with all the channels in one head, i.e. [DATA_CHANNELS]. For example, if "V", "H" and "Z" channels are in the same head, it is likely
279. # that they are representing distance maps and therefore the model will be able to learn them together. On the other hand, if "C" channel
280. # is in a different head than "H", "V" and "Z", it is likely that the model will be able to learn better the contours of the instances
281. # separately from the distance maps.
282. _C.PROBLEM.INSTANCE_SEG.CHANNELS_PER_HEAD_INFO = []
283. # Whether to use a different decoder for each head in the model. The heads are defined by 'CHANNELS_PER_HEAD_INFO',
284. # so at least two entries are required there. Only supported by: ['unet', 'resunet', 'resunet++', 'seunet',
285. # 'resunet_se', 'attention_unet', 'unext_v1', 'unext_v2'].
286. _C.PROBLEM.INSTANCE_SEG.SEPARATED_DECODERS_PER_HEAD = False
287. # Whether to divide 'MODEL.FEATURE_MAPS' by the number of decoders created when 'SEPARATED_DECODERS_PER_HEAD'
288. # is enabled. This way the model keeps a number of parameters closer to the one built with just one decoder.
289. # If False each decoder is built with 'MODEL.FEATURE_MAPS' as they are, so the model will be as many times
290. # bigger (in its decoder part) as decoders are created. Only used when 'SEPARATED_DECODERS_PER_HEAD' is True.
291. _C.PROBLEM.INSTANCE_SEG.SEPARATED_DECODERS_DIVIDE_FEATURE_MAPS = False
292. # Weights to be applied to the channels. Notice that these weights are not applied directly to the loss, but to the predicted channels before
293. # calculating the loss. The length of the list must be equal to the number of channels.
294. # Notice that this is different from LOSS.WEIGHTS, which are used to apply weights to different losses.
295. _C.PROBLEM.INSTANCE_SEG.DATA_CHANNEL_WEIGHTS = (1, 1)
296. # Whether to add an extra weight map to the loss calculation to focus on the borders between instances. Different weighting maps
297. # can be defined. Options: ["unet-like", ""]
298. _C.PROBLEM.INSTANCE_SEG.BORDER_EXTRA_WEIGHTS = ""
299. # Defines how the instances are created. Options:
300. # - "watershed" to use watershed algorithm
301. # - "gradient-flow" to use gradient flow algorithm (Cellpose/Omnipose)
302. # - "stardist" to use stardist algorithm
303. # - "embeddings" to use embedding-based clustering algorithms
304. # - "agglomeration" to oversegment predicted affinities and greedily merge fragments by
305. # affinity (requires 'PROBLEM.INSTANCE_SEG.DATA_CHANNELS' == ['A']). See
306. # 'PROBLEM.INSTANCE_SEG.AGGLOMERATION' for its options.
307. _C.PROBLEM.INSTANCE_SEG.INSTANCE_CREATION_PROCESS = ""
309. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
310. # 2.2.1 Watershed options for instance segmentation
311. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
312. # Options for marker-controlled watershed
313. _C.PROBLEM.INSTANCE_SEG.WATERSHED = CN()
314. # List of the channels to be used for seed creation. If not provided will be automatically set based on
315. # 'PROBLEM.INSTANCE_SEG.DATA_CHANNELS'.
316. _C.PROBLEM.INSTANCE_SEG.WATERSHED.SEED_CHANNELS = []
317. # Thresholds for the seed channels. If not provided will be automatically set with "auto" for each channel.
318. # If set it must have same length than 'PROBLEM.INSTANCE_SEG.WATERSHED.SEED_CHANNELS'. Options:
319. # - A float between 0 and 1 to threshold each channel
320. # - "auto" to automatically define a value by applying an Otsu thresholding
321. _C.PROBLEM.INSTANCE_SEG.WATERSHED.SEED_CHANNELS_THRESH = []
322. # Defines the topographic surface to grow the seeds. If not provided will be automatically set based on
323. # 'PROBLEM.INSTANCE_SEG.DATA_CHANNELS'. If defined it must be a channel name, e.g. "B" or "C".
324. _C.PROBLEM.INSTANCE_SEG.WATERSHED.TOPOGRAPHIC_SURFACE_CHANNEL = ""
325. # Channel to be used for growth mask creation. If not provided will be automatically set based on
326. # 'PROBLEM.INSTANCE_SEG.DATA_CHANNELS'. Options:
327. # - A list of channel names, e.g. ["B"], ["C"] or ["B", "C"]. The channels must be in 'PROBLEM.INSTANCE_SEG.DATA_CHANNELS'
328. # - "auto" to automatically create it based on the selected channels ('PROBLEM.INSTANCE_SEG.DATA_CHANNELS')
329. _C.PROBLEM.INSTANCE_SEG.WATERSHED.GROWTH_MASK_CHANNELS = []
330. # Thresholds for the growth mask channels. If not provided will be automatically set with "auto" for each channel.
331. # If set it must have same length than 'PROBLEM.INSTANCE_SEG.WATERSHED.GROWTH_MASK_CHANNELS'. Options:
332. # - A float between 0 and 1 to threshold each channel
333. # - "auto" to automatically define a value depending on the channels provided
334. _C.PROBLEM.INSTANCE_SEG.WATERSHED.GROWTH_MASK_CHANNELS_THRESH = []
335. # Sequence of string to determine the morphological filters to apply to instance seeds. They will be done in that order.
336. # Possible options 'dilate' and 'erode'. E.g. ['erode','dilate'] to erode first and dilate later.
337. _C.PROBLEM.INSTANCE_SEG.WATERSHED.SEED_MORPH_SEQUENCE = []
338. # Sequence of ints to determine the radius of the erosion or dilation for instance seeds
339. _C.PROBLEM.INSTANCE_SEG.WATERSHED.SEED_MORPH_RADIUS = []
340. # To erode and dilate the foreground mask before using marker controlled watershed. The idea is to remove the small holes
341. # that may be produced so the instances grow without them
342. _C.PROBLEM.INSTANCE_SEG.WATERSHED.ERODE_AND_DILATE_GROWTH_MASK = False
343. # Radius to erode the foreground mask
344. _C.PROBLEM.INSTANCE_SEG.WATERSHED.FORE_EROSION_RADIUS = 5
345. # Radius to dilate the foreground mask
346. _C.PROBLEM.INSTANCE_SEG.WATERSHED.FORE_DILATION_RADIUS = 5
347. # Whether to save watershed check files
348. _C.PROBLEM.INSTANCE_SEG.WATERSHED.DATA_CHECK_MW = False
349. # Size of small objects to be removed before doing watershed
350. _C.PROBLEM.INSTANCE_SEG.WATERSHED.DATA_REMOVE_SMALL_OBJ_BEFORE = 10
351. # Whether to remove objects before watershed
352. _C.PROBLEM.INSTANCE_SEG.WATERSHED.DATA_REMOVE_BEFORE_MW = False
353. # Whether to apply or not the watershed to create instances slice by slice in a 3D problem. This can solve instances invading
354. # others if the objects in Z axis overlap too much.
355. _C.PROBLEM.INSTANCE_SEG.WATERSHED.BY_2D_SLICES = False
357. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
358. # 2.2.1.1 Affinity agglomeration options for instance segmentation
359. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
360. # Used when 'PROBLEM.INSTANCE_SEG.INSTANCE_CREATION_PROCESS' is 'agglomeration' (requires
361. # 'PROBLEM.INSTANCE_SEG.DATA_CHANNELS' == ['A']): oversegment the predicted affinities into
362. # small fragments via watershed, then greedily merge fragment pairs by a quantile of their
363. # affinity histogram until the merge threshold is reached. Matches waterz's
364. # (https://github.com/funkey/waterz) canonical behaviour: only the first short-range
365. # (z, y, x)/(y, x) affinity triple is used, by both the fragment watershed and the merge
366. # scoring; any further, longer-range offsets are ignored. See
367. # biapy/data/post_processing/affinity_agglomeration.py.
368. _C.PROBLEM.INSTANCE_SEG.AGGLOMERATION = CN()
369. # Seed threshold for the initial oversegmented fragments (high, so fragments never straddle a
370. # real instance boundary).
371. _C.PROBLEM.INSTANCE_SEG.AGGLOMERATION.FRAGMENT_SEED_TH = 0.9
372. # Growth-mask threshold for those fragments (low, so they cover all foreground with no gaps).
373. _C.PROBLEM.INSTANCE_SEG.AGGLOMERATION.FRAGMENT_GROWTH_TH = 0.1
374. # Two fragments merge while their MERGE_QUANTILE-th percentile affinity (short-range triple
375. # only, see above) is >= this value.
376. _C.PROBLEM.INSTANCE_SEG.AGGLOMERATION.MERGE_TH = 0.5
377. # Percentile (0-100) of each fragment pair's affinity histogram used as its merge score.
378. # 50 = median.
379. _C.PROBLEM.INSTANCE_SEG.AGGLOMERATION.MERGE_QUANTILE = 50.0
380. # Minimum supporting voxel-pairs an edge needs before it can trigger a merge. 1 effectively
381. # disables this.
382. _C.PROBLEM.INSTANCE_SEG.AGGLOMERATION.MIN_EDGE_VOXELS = 5
384. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
385. # 2.2.2 Stardist-like post-processing options for instance segmentation
386. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
387. # Options for stardist-kind instance creation
388. _C.PROBLEM.INSTANCE_SEG.STARDIST = CN()
389. # Probability threshold to consider a pixel/voxel as a potential instance center
390. _C.PROBLEM.INSTANCE_SEG.STARDIST.PROB_THRESH = 0.4
391. # Non-maximum suppression IoU threshold to filter overlapping instance candidates
392. _C.PROBLEM.INSTANCE_SEG.STARDIST.NMS_IOU_THRESH = 0.3
393. # Output stride of the StarDist head per axis (StarDist's 'grid'); post-processing scales candidate
394. # coordinates by it. BiaPy's U-Net predicts at full resolution, so keep it all-ones; only set >1 if
395. # the model is changed to emit the maps on a coarser lattice. Empty -> all ones for the given ndim.
396. _C.PROBLEM.INSTANCE_SEG.STARDIST.GRID = []
399. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
400. # 2.2.3 Cellpose/Omnipose-like post-processing options for instance segmentation
401. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
402. # Options for flow-field-based instance creation from Gv/Gh/[Gz] channels.
403. # Inspired by the work:
404. # References:
405. # "Cellpose: a generalist algorithm for cellular segmentation"
406. # [link]: https://www.nature.com/articles/s41592-020-01018-x
407. # "Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation"
408. # [link]: https://www.nature.com/articles/s41592-022-01639-4
409. # Cellpose vs Omnipose post-processing is selected by the flow channels' gradient strategy
410. # (DATA_CHANNELS_EXTRA_OPTS[0]['Gv']['gradient_type'], 'cellpose' or 'omnipose'); parameters for
411. # each live in the CELLPOSE and OMNIPOSE nodes below.
412. #
413. # Cellpose post-processing: Euler integration followed by histogram peak detection and 3x3
414. # expansion, with an optional flow-error check that removes instances whose flow (regenerated
415. # from the mask) disagrees with the network. Suited for Cellpose-style flows.
416. _C.PROBLEM.INSTANCE_SEG.CELLPOSE = CN()
417. # Foreground probability threshold. Pixels/voxels with a predicted foreground probability
418. # above this value are considered foreground and will be traced through the flow field.
419. # Applies when a dedicated binary channel ('F', 'M', or 'B') is present in DATA_CHANNELS.
420. # Matches Cellpose default (cellprob_threshold=0.0 in logit space = sigmoid 0.5).
421. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.FG_THRESH = 0.5
422. # Flow-error threshold to discard spurious instances: each mask's flow is regenerated by
423. # diffusion and compared (mean squared error) against the network flow, removing the fragments
424. # left by over-segmentation. 0 disables the check. Default 0.4.
425. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.FLOW_THRESHOLD = 0.4
426. # Number of Euler integration steps (overridden internally by niter = (DIAMETER / DIAM_MEAN) * 200).
427. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.N_STEPS = 200
428. # Expected cell diameter (px) in the TEST data (test time only; training always rescales per
429. # image). The input is rescaled by DIAM_MEAN / DIAMETER, predicted, and resized back, with
430. # niter = (DIAMETER / DIAM_MEAN) * 200. DIAMETER == 0 estimates it per image
431. # (TEST_DOUBLE_INFERENCE), else uses the training-set median. Default: 0.0.
432. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.DIAMETER = 0.0
433. # Reference cell diameter (px) the flow model is trained at (30 'cyto', 17 'nuclei'). Both
434. # training and test rescale cells to this size. Default: 30.0.
435. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.DIAM_MEAN = 30.0
436. # Amount of random scale jitter applied on top of the per-image diameter rescale during training
437. # (data augmentation), mirroring Cellpose's 'scale_range'. Each training patch is scaled by a
438. # factor drawn as (1 - SCALE_RANGE/2) + SCALE_RANGE * U[0, 1), i.e. uniformly in
439. # [1 - SCALE_RANGE/2, 1 + SCALE_RANGE/2], and this factor multiplies the DIAM_MEAN / diameter
440. # normalization. Only applied when training with flow channels; validation/test use the plain
441. # diameter rescale with no jitter. 0.0 disables the jitter. Matches Cellpose's default for the
442. # rescale training path (cellpose/train.py: scale_range0 = 0.5 when rescale=True), giving the
443. # range [0.75, 1.25]. Default: 0.5.
444. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.SCALE_RANGE = 0.5
445. # When DIAMETER == 0, estimate each test image's diameter with a cheap first inference pass (on
446. # one central patch) instead of the training-set median, then rescale the input for the real
447. # pass. Ignored when DIAMETER > 0, for the by-chunks/Zarr path, and for 'torchvision' models.
448. # Default: True.
449. _C.PROBLEM.INSTANCE_SEG.CELLPOSE.TEST_DOUBLE_INFERENCE = True
451. # Omnipose post-processing: divergence-rescaled, suppressed (step-damped) Euler flow-following
452. # with the step count derived from the predicted distance field, then DBSCAN clustering of the
453. # convergence points (a faithful port of Omnipose's compute_masks). Requires a predicted distance
454. # channel ('Db' with val_type 'omnipose'). Omnipose is diameter-agnostic, so the Cellpose
455. # DIAMETER/DIAM_MEAN/SCALE_RANGE rescaling does not apply.
456. _C.PROBLEM.INSTANCE_SEG.OMNIPOSE = CN()
457. # Foreground threshold on the predicted distance field ('Db'): pixels with distance above this
458. # are traced (Omnipose's mask_threshold, applied via hysteresis at [MASK_THRESHOLD-1,
459. # MASK_THRESHOLD]). Default: 0.0.
460. _C.PROBLEM.INSTANCE_SEG.OMNIPOSE.MASK_THRESHOLD = 0.0
461. # Flow-error threshold to discard masks whose flow (regenerated from the mask) disagrees with the
462. # network. 0 disables the check. Matches Omnipose default 0.4.
463. _C.PROBLEM.INSTANCE_SEG.OMNIPOSE.FLOW_THRESHOLD = 0.4
464. # Number of Euler integration steps. 0 = automatic (Omnipose derives it from the predicted
465. # distance field: niter = 2*(ndim+1)*mean(distance)). A positive value overrides it, mirroring
466. # Omnipose's --niter. Default: 0.
467. _C.PROBLEM.INSTANCE_SEG.OMNIPOSE.NITER = 0
469. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
470. # 2.2.4 EmbedSeg-like post-processing options for instance segmentation
471. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
472. # Options for embedding-based clustering instance creation. They are inspired by the work:
473. # Reference:
474. # "EmbedSeg: Embedding-based Instance Segmentation for Biomedical Microscopy Data"
475. # [link]: https://www.sciencedirect.com/science/article/pii/S1361841522001700
476. # Code adapted from:
477. # Embedseg: https://github.com/juglab/EmbedSeg
478. _C.PROBLEM.INSTANCE_SEG.EMBEDSEG = CN()
479. # Seediness threshold defining the foreground pixels that get clustered into objects.
480. _C.PROBLEM.INSTANCE_SEG.EMBEDSEG.FG_THRESH = 0.5
481. # Seediness threshold a pixel must reach to seed a new object; clustering stops once no
482. # remaining foreground pixel exceeds it.
483. _C.PROBLEM.INSTANCE_SEG.EMBEDSEG.SEED_THRESH = 0.9
484. # Minimum number of foreground pixels required to perform clustering.
485. _C.PROBLEM.INSTANCE_SEG.EMBEDSEG.MIN_MASK_SUM = 0
486. # Minimum number of unclustered foreground pixels to continue clustering.
487. _C.PROBLEM.INSTANCE_SEG.EMBEDSEG.MIN_UNCLUSTERED_SUM = 0
488. # Size (in pixels) of the canonical coordinate grid used to build the spatial-embedding
489. # coordinate map, i.e. the per-pixel coordinate step is ``1 / (GRID_SIZE - 1)``. This mirrors
490. # EmbedSeg's ``n_x = n_y`` (the dataset's max image size) and is decoupled from
491. # ``DATA.PATCH_SIZE`` on purpose: training patches are treated as slices of this grid so the
492. # coordinate scale (and hence the sigma initialisation ``s = exp(10)``) matches the original
493. # regardless of the patch size. The SAME value is used by the loss (training) and the clustering
494. # (inference). Leave it at -1 to compute it automatically from the dataset's max image size
495. # (cached to a JSON next to the instance-channel folder, like the Cellpose diameter); set a
496. # positive value to force it.
497. _C.PROBLEM.INSTANCE_SEG.EMBEDSEG.GRID_SIZE = -1
499. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
500. # 2.2.5 Synapse-specific options for instance segmentation
501. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
502. #### For "synapses" type of instances (only available for 3D H5/Zarr data) ####
503. _C.PROBLEM.INSTANCE_SEG.SYNAPSES = CN()
504. # Method to create the points from the synapse prediction (in "F_pre" + "F_post" setting). Options are:
505. # - 'peak_local_max' to use the skimage.feature.peak_local_max function to create the points
506. # - 'blob_log' to use the skimage.feature.blob_log function to create the points
507. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.POINT_CREATION_FUNCTION = "peak_local_max"
508. # The minimal allowed distance separating peaks. To find the maximum number of peaks, use min_distance=1.
509. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.PEAK_LOCAL_MAX_MIN_DISTANCE = 1
510. # Whether the threshold are going to be set as automaticaly (with Otsu thresholding) or manually.
511. # Options available:
512. # * 'auto' to decide the threshold to be applied by measuring it with Otsu
513. # * 'manual' to set a fixed threshold defined by 'PROBLEM.INSTANCE_SEG.SYNAPSES.MIN_TH_TO_BE_PEAK'
514. # * 'relative_by_patch' to use 'PROBLEM.INSTANCE_SEG.SYNAPSES.MIN_TH_TO_BE_PEAK' but relative to the maximum value in the predicted patch
515. # data. More info in https://scikit-image.org/docs/0.25.x/api/skimage.feature.html#skimage.feature.peak_local_max and
516. # https://scikit-image.org/docs/0.23.x/api/skimage.feature.html#skimage.feature.blob_log (see 'threshold_rel' argument description)
517. # * 'relative' to use 'PROBLEM.INSTANCE_SEG.SYNAPSES.MIN_TH_TO_BE_PEAK' but relative to the maximum value in the whole predicted data.
518. # More info in https://scikit-image.org/docs/0.25.x/api/skimage.feature.html#skimage.feature.peak_local_max and
519. # https://scikit-image.org/docs/0.23.x/api/skimage.feature.html#skimage.feature.blob_log (see 'threshold_rel' argument description)
520. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.TH_TYPE = "auto"
521. # Minimun value to consider a point as a peak. Corresponds to 'threshold_abs' argument of the function
522. # 'peak_local_max' of skimage.feature
523. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.MIN_TH_TO_BE_PEAK = 0.2
524. # Corresponds to 'exclude_border' argument of 'peak_local_max' or 'blob_log' function of skimage. If True it will exclude
525. # peaks from the border of the image to avoid partial detection.
526. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.EXCLUDE_BORDER = False
527. # Corresponds to 'min_sigma' argument of 'blob_log' function. It is the minimum standard deviation for Gaussian kernel.
528. # Keep this low to detect smaller blobs. The standard deviations of the Gaussian filter are given for each axis as a
529. # sequence, or as a single number, in which case it is equal for all axes.
530. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.BLOB_LOG_MIN_SIGMA = 5
531. # Corresponds to 'max_sigma' argument of 'blob_log' function. It is the maximum standard deviation for Gaussian kernel.
532. # Keep this high to detect larger blobs. The standard deviations of the Gaussian filter are given for each axis as a
533. # sequence, or as a single number, in which case it is equal for all axes.
534. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.BLOB_LOG_MAX_SIGMA = 10
535. # Corresponds to 'num_sigma' argument of 'blob_log' function. The number of intermediate values of standard deviations
536. # to consider between min_sigma and max_sigma.
537. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.BLOB_LOG_NUM_SIGMA = 2
538. # For removing close points detected during synapse prediction
539. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.REMOVE_CLOSE_PRE_POINTS_RADIUS = 0
540. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.REMOVE_CLOSE_POST_POINTS_RADIUS = 0
541. # If this is activated the post points will be remove not only looking the distance between them but the must lay in the same mask.
542. # For creating this mask there can be a few options:
543. # * Implemented: By creating instances binarizing the predictions (using PROBLEM.INSTANCE_SEG.SYNAPSES.MIN_TH_TO_BE_PEAK) and applying
544. # a connected-components
545. # * TODO: by looking at external neuron segmentation
546. _C.PROBLEM.INSTANCE_SEG.SYNAPSES.REMOVE_CLOSE_POINTS_RADIUS_BY_MASK = False
548. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
549. # 2.3 Detection problem specification
550. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
551. _C.PROBLEM.DETECTION = CN()
552. # Shape of the ellipse that will be used to dilate the central point created from the CSV file. 0 to not dilate and only create a 3x3 square.
553. # The value is the radius of the ellipse in pixels. If an integer is given, the shape will be a ball with the given side length.
554. # If a list is given, the shape will be a hyperball with the given side lengths. List order is (y,x) or (z,y,x) for 2D and 3D respectively.
555. # For example [1, 2, 3] will result in an ellipse with a radius of 1 in the first dimension, 2 in the second and 3 in the third.
556. _C.PROBLEM.DETECTION.CENTRAL_POINT_DILATION = [2]
557. _C.PROBLEM.DETECTION.CHECK_POINTS_CREATED = True
558. # Whether to save watershed check files
559. _C.PROBLEM.DETECTION.DATA_CHECK_MW = False
560. # Whether to use a different decoder for each head in the model. Only applicable when 'DATA.N_CLASSES' > 2
561. # (multi-class detection), where the model has two heads: one for the detection points and one for the class
562. # predictions. Only supported by: ['unet', 'resunet', 'resunet++', 'seunet', 'resunet_se', 'attention_unet',
563. # 'unext_v1', 'unext_v2'].
564. _C.PROBLEM.DETECTION.SEPARATED_DECODERS_PER_HEAD = False
565. # Whether to divide 'MODEL.FEATURE_MAPS' by the number of decoders created when 'SEPARATED_DECODERS_PER_HEAD'
566. # is enabled. This way the model keeps a number of parameters closer to the one built with just one decoder.
567. # If False each decoder is built with 'MODEL.FEATURE_MAPS' as they are, so the model will be as many times
568. # bigger (in its decoder part) as decoders are created. Only used when 'SEPARATED_DECODERS_PER_HEAD' is True.
569. _C.PROBLEM.DETECTION.SEPARATED_DECODERS_DIVIDE_FEATURE_MAPS = False
570. # Whether to apply a rebalancing strategy to the loss function to give more importance to underrepresented pixels within the channels.
571. # The weights are calculated automatically based on the number of pixels of each class per batch and directly in the loss computation.
572. # In the specific case of detection, where there are usually much less pixels representing the center of the objects to detect than
573. # background pixels, with this option activated, the loss will give more importance to the pixels representing the center of the objects
574. # to help the model learn better to predict them.
575. _C.PROBLEM.DETECTION.CLASS_REBALANCE_WITHIN_CHANNELS = True
576. # Weights to be applied to the channels when doing detection with classes. Notice that these weights are not applied directly to the loss,
577. # but to the predicted channels before calculating the loss. The length of the list must be equal to the number of channels.
578. # Notice that this is different from LOSS.WEIGHTS, which are used to apply weights to different losses.
579. _C.PROBLEM.DETECTION.DATA_CHANNEL_WEIGHTS = (1, 1)
581. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
582. # 2.4 Denoising problem specification
583. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
584. # Based Noise2Void paper: https://arxiv.org/abs/1811.10980
585. _C.PROBLEM.DENOISING = CN()
586. # This variable corresponds to n2v_perc_pix from Noise2Void. It explanation is as follows: for faster training multiple
587. # pixels per input patch can be manipulated. In our experiments we manipulated about 0.198% of the input pixels per
588. # patch. For a patch size of 64 by 64 pixels this corresponds to about 8 pixels. This fraction can be tuned via this variable
589. _C.PROBLEM.DENOISING.N2V_PERC_PIX = 0.198
590. # This variable corresponds to n2v_manipulator from Noise2Void. Most pixel manipulators will compute the replacement value based
591. # on a neighborhood and this variable controls how to do that
592. _C.PROBLEM.DENOISING.N2V_MANIPULATOR = "uniform_withCP"
593. # This variable corresponds to n2v_neighborhood_radius from Noise2Void. Size of the neighborhood to compute the replacement
594. _C.PROBLEM.DENOISING.N2V_NEIGHBORHOOD_RADIUS = 5
595. # To apply a structured mask as is proposed in Noise2Void to alleviate the limitation of the method of not removing effectively
596. # the structured noise (section 4.4 of their paper).
597. _C.PROBLEM.DENOISING.N2V_STRUCTMASK = False
598. # Whether to load ground truth data in denoising
599. _C.PROBLEM.DENOISING.LOAD_GT_DATA = False
601. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
602. # 2.4 Super-resolution problem specification
603. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
604. _C.PROBLEM.SUPER_RESOLUTION = CN()
605. # Upscaling to be done to the input images on every dimension. Examples: (2,2) in 2D or (2,2,2) in 3D.
606. _C.PROBLEM.SUPER_RESOLUTION.UPSCALING = ()
608. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
609. # 2.5 Self-supervised problem specification
610. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
611. _C.PROBLEM.SELF_SUPERVISED = CN()
612. # Pretext task to do. Options are as follows:
613. # - 'crappify': crappifies input image by adding Gaussian noise and downsampling and upsampling it so the resolution
614. # gets worsen. Then, the model is trained to recover the original images.
615. # - 'masking': mask input image and the model needs to recover the original image. This option can only
616. # be done with 'mae' transformer. This strategy follows the one proposed in:
617. # Masked Autoencoders Are Scalable Vision Learners (https://arxiv.org/pdf/2111.06377.pdf)
618. _C.PROBLEM.SELF_SUPERVISED.PRETEXT_TASK = "crappify"
619. # Downsizing factor to reshape the image. It will be downsampled and upsampled again by this factor so the
620. # quality of the image is worsens
621. _C.PROBLEM.SELF_SUPERVISED.RESIZING_FACTOR = 4
622. # Number between [0, 1] indicating the std of the Gaussian noise N(0,std).
623. _C.PROBLEM.SELF_SUPERVISED.NOISE = 0.2
625. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
626. # 2.6 Image-to-image problem specification
627. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
628. _C.PROBLEM.IMAGE_TO_IMAGE = CN()
629. # Information on how the channels are distributed in the model's output heads. It must be a list of list of ints, where each
630. # inner list contains the number of channels that are going to be predicted in the same head.
631. _C.PROBLEM.IMAGE_TO_IMAGE.CHANNELS_PER_HEAD_INFO = []
632. # Whether to use a different decoder for each head in the model. The heads are defined by 'CHANNELS_PER_HEAD_INFO',
633. # so at least two entries are required there. Only supported by: ['unet', 'resunet', 'resunet++', 'seunet',
634. # 'resunet_se', 'attention_unet', 'unext_v1', 'unext_v2'].
635. _C.PROBLEM.IMAGE_TO_IMAGE.SEPARATED_DECODERS_PER_HEAD = False
636. # Whether to divide 'MODEL.FEATURE_MAPS' by the number of decoders created when 'SEPARATED_DECODERS_PER_HEAD'
637. # is enabled. This way the model keeps a number of parameters closer to the one built with just one decoder.
638. # If False each decoder is built with 'MODEL.FEATURE_MAPS' as they are, so the model will be as many times
639. # bigger (in its decoder part) as decoders are created. Only used when 'SEPARATED_DECODERS_PER_HEAD' is True.
640. _C.PROBLEM.IMAGE_TO_IMAGE.SEPARATED_DECODERS_DIVIDE_FEATURE_MAPS = False
641. # Expected output channels to be predicted by the model.
642. _C.PROBLEM.IMAGE_TO_IMAGE.OUTPUT_CHANNELS = 1
643. # To use a custom data loader to load a random image from each image sample folder. The data needs to be structured
644. # in an special way, that is, instead of having images in the training/val folder a folder for each sample is expected,
645. # where in each of those different versions of the same data sample will be placed. Visit the following tutorial
646. # for a real use case and a more detailed description:
647. # - https://biapy.readthedocs.io/en/latest/tutorials/image-to-image/lightmycells.html
648. _C.PROBLEM.IMAGE_TO_IMAGE.MULTIPLE_RAW_ONE_TARGET_LOADER = False
649. # Activation function to be applied to the output channels of the model. It can be a string or a list of strings if
650. # different activation functions are desired for different channels. Leave empty to use linear activation.
651. _C.PROBLEM.IMAGE_TO_IMAGE.OUTPUT_CHANNEL_ACT = []
653. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
654. # 2.6.1 Membrane repair sub-problem (IMAGE_TO_IMAGE)
655. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
656. # Trains a dataset-agnostic network that repairs membrane-segmentation errors (gaps and
657. # spurious fragments) coming out of an upstream foundation-model + GMM pipeline. Model input
658. # is EXCLUSIVELY DERIVED_CHANNELS, computed from SOURCE_CHANNELS (never fed to the model
659. # directly). Output is affinities learned from GT instance labels. Uses a dedicated data
660. # generator (Membrane2DRepairDataGenerator/Membrane3DRepairDataGenerator) instead of the
661. # shared instance-seg/Cellpose/N2V generator.
662. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR = CN()
663. # Master switch. When True, PROBLEM.TYPE == "IMAGE_TO_IMAGE" is routed through the dedicated
664. # membrane-repair data generator and Membrane_Repair_Workflow instead of the regular
665. # IMAGE_TO_IMAGE_Workflow path.
666. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ENABLE = False
667. # Ordered list of the raw, on-disk input channels. Never fed to the model directly -- only
668. # used to compute DERIVED_CHANNELS (see below). Channels are resolved by name, not
669. # position, so any order/subset works, e.g. ["membrane"], ["raw"], ["membrane", "raw"] or
670. # ["raw", "membrane"] ("raw" required whenever 'meijering' is in DERIVED_CHANNELS;
671. # "membrane" required whenever 'skeleton_dt'/'hessian_blob' are). Any entry that is neither
672. # "membrane" nor "raw" (e.g. unsupervised per-slice GMM class maps, whose cluster-to-channel
673. # assignment is not stable across slices) is treated as a "class" channel, folded into the
674. # single 'class_union' derived channel below -- their names/count/order don't matter beyond
675. # that, since 'class_union' only takes their per-pixel union.
676. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SOURCE_CHANNELS = ["membrane", "raw"]
677. # Ordered list of channels derived on the fly from SOURCE_CHANNELS -- the model's actual
678. # input; must be non-empty. Options:
679. # - 'skeleton_dt': clamped Euclidean distance transform of the per-slice membrane skeleton.
680. # Derived from 'membrane' (required in SOURCE_CHANNELS).
681. # - 'hessian_blob': Hessian-eigenvalue-based "dense blob" response (mito/synapse/vesicle cue).
682. # Derived from 'membrane' (required in SOURCE_CHANNELS).
683. # - 'meijering': standardised multi-scale Meijering ridge filter. Derived from 'raw'
684. # (required in SOURCE_CHANNELS).
685. # - 'class_union': per-pixel union (max) of every SOURCE_CHANNELS entry that is not
686. # 'membrane'/'raw' (at least one such entry required in SOURCE_CHANNELS). Order-
687. # invariant by construction -- use this instead of feeding per-class channels directly
688. # whenever the class identity behind each channel isn't stable slice to slice.
689. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.DERIVED_CHANNELS = ["skeleton_dt", "hessian_blob", "meijering"]
690. # Per-channel options for DERIVED_CHANNELS. Must be a list with a unique element: a dict of
691. # dicts, keyed by channel name. Possible options:
692. # - 'skeleton_dt' channel. Possible options:
693. # - 'clamp_px': int, clamp distance (in canonical pixels) applied to the DT. Default: 10
694. # - 'per_slice': bool, whether to skeletonize/derive the DT per z-slice (2D skeleton in a
695. # 3D stack) rather than on the full 3D volume. Default: True
696. # - 'hessian_blob' channel. Possible options:
697. # - 'sigma_range': list of 2 floats, [min, max] Gaussian scales probed for the Hessian.
698. # Default: [1.0, 3.0]
699. # - 'meijering' channel. Possible options:
700. # - 'sigma_range': list of 2 floats, [min, max] Gaussian scales probed for the ridge
701. # filter. Default: [1.0, 4.0]
702. # - 'standardize': bool, whether to z-score/percentile-normalize the response for
703. # cross-dataset comparability. Default: True
704. # - 'class_union' channel. Possible options:
705. # - 'threshold': float, threshold applied to each class map before taking the union.
706. # Default: 0.5
707. # - identity-passthrough channels (any name matching a SOURCE_CHANNELS entry): no options.
708. # For example:
709. # DERIVED_CHANNELS = ['skeleton_dt', 'hessian_blob']
710. # DERIVED_CHANNELS_EXTRA_OPTS = [{'skeleton_dt': {'clamp_px': 8}, 'hessian_blob': {'sigma_range': [1.0, 2.5]}}]
711. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.DERIVED_CHANNELS_EXTRA_OPTS = [{}]
712. # Subset of DERIVED_CHANNELS naming an "exchangeable" group of identity-passthrough class
713. # channels whose channel-to-semantic mapping isn't stable slice to slice (e.g. per-slice
714. # GMM cluster ids). When non-empty, this is forwarded to MODEL.STUNET.CLASS_SET_CHANNEL_IDXS
715. # (translated from names to DERIVED_CHANNELS positions) so the model routes this group
716. # through a permutation-invariant encoder instead of its regular first layer -- see
717. # MODEL.STUNET.CLASS_SET_CHANNEL_IDXS's docstring. Only meaningful with MODEL.ARCHITECTURE
718. # == 'stunet' and MODEL.STUNET.VARIANT == 'custom'. Left empty (default), this is a no-op.
719. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.CLASS_SET_CHANNELS = []
720. # Output width and pooling of the CLASS_SET_CHANNELS group's pooled encoding -- forwarded to
721. # MODEL.STUNET.CLASS_SET_OUT_CHANNELS / CLASS_SET_POOLING. Only used when CLASS_SET_CHANNELS
722. # is non-empty.
723. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.CLASS_SET_ENCODER = CN()
724. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.CLASS_SET_ENCODER.OUT_CHANNELS = 8
725. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.CLASS_SET_ENCODER.POOLING = 'max'
727. # Y-side: GT target channels, generated offline from the raw GT instance-label folder via the
728. # same 'labels_into_channels' machinery INSTANCE_SEG uses (see PROBLEM.INSTANCE_SEG.DATA_CHANNELS
729. # for the full list of channel letters), then regenerated online after each augmentation warp so
730. # directional channels never get corrupted by interpolation. Must include 'I' (the virtual raw
731. # instance-label channel), the regeneration source that is dropped before the batch reaches the
732. # model -- exactly as INSTANCE_SEG does for its own directional channels. 'A' (affinities) is the
733. # actual training target.
734. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.DATA_CHANNELS = ["A", "I"]
735. # Same shape/semantics as PROBLEM.INSTANCE_SEG.DATA_CHANNELS_EXTRA_OPTS's 'A' entry:
736. # 'z_affinities'/'y_affinities'/'x_affinities' (paired-by-index neighbour offset lists) and
737. # 'widen_borders'. E.g. [{'A': {'z_affinities': [1], 'y_affinities': [1], 'x_affinities': [1]}}]
738. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.DATA_CHANNELS_EXTRA_OPTS = [{}]
739. # Like PROBLEM.INSTANCE_SEG.CLASS_REBALANCE_WITHIN_CHANNELS, but per affinity offset channel
740. # instead of pooled across the whole 'A' stack. False = old pooled behavior (reproducibility).
741. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.CLASS_REBALANCE_WITHIN_CHANNELS = True
743. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
744. # 2.6.2 Membrane repair test-time post-processing (affinities -> instances)
745. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
746. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS = CN()
747. # How to turn predicted affinities into instance labels at test time: 'watershed' or
748. # 'agglomeration' (see biapy/data/post_processing/affinity_agglomeration.py).
749. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.METHOD = "agglomeration"
750. # 'watershed' only: seed threshold for watershed_by_channels's 'A'-only branch.
751. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.WATERSHED_SEED_TH = 0.5
752. # 'watershed' only: growth-mask threshold. Equal to WATERSHED_SEED_TH gives plain
753. # threshold + 3D connected components (no seeded growth).
754. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.WATERSHED_GROWTH_TH = 0.5
755. # 'agglomeration' only: seed threshold for the initial oversegmented fragments (high, so
756. # fragments never straddle a real instance boundary).
757. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.FRAGMENT_SEED_TH = 0.9
758. # 'agglomeration' only: growth-mask threshold for those fragments (low, so they cover all
759. # foreground with no gaps).
760. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.FRAGMENT_GROWTH_TH = 0.1
761. # 'agglomeration' only: two fragments merge while their MERGE_QUANTILE-th percentile
762. # affinity (short-range triple only, see METHOD above) is >= this value.
763. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.MERGE_TH = 0.5
764. # 'agglomeration' only: percentile (0-100) of each fragment pair's affinity histogram used
765. # as its merge score. 50 = median.
766. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.MERGE_QUANTILE = 50.0
767. # 'agglomeration' only: minimum supporting voxel-pairs an edge needs before it can trigger a
768. # merge. 1 effectively disables this.
769. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.POSTPROCESS.MIN_EDGE_VOXELS = 5
771. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
772. # 2.6.3 Membrane repair corruption augmentors
773. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
774. # These synthesize the upstream pipeline's failure modes directly on the membrane source
775. # channel, so the network learns the repair operation from corrupted -> clean pairs.
776. # Applied after the geometric warps, before the derived channels are computed.
777. # PROB is a per-z-slice probability, independently rolled for every slice (2D: probability of
778. # augmenting the whole image), not a single roll for the whole sample.
779. #
780. # Blacks out random bands across the membrane channel (merge error). LENGTH_RANGE: fraction
781. # (0-1) of border-to-border extent (1.0 = full). THICKNESS_RANGE: pixels. N_LINES: bands
782. # per slice.
783. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.GAP_AUG = CN()
784. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.GAP_AUG.ENABLE = False
785. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.GAP_AUG.PROB = 0.5
786. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.GAP_AUG.LENGTH_RANGE = (0.3, 1.0)
787. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.GAP_AUG.THICKNESS_RANGE = (4, 9)
788. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.GAP_AUG.N_LINES = (1, 3)
789. # Inverse of GAP_AUG: paints bands instead of erasing (split error). Same options.
790. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.BRIDGE_AUG = CN()
791. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.BRIDGE_AUG.ENABLE = False
792. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.BRIDGE_AUG.PROB = 0.3
793. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.BRIDGE_AUG.LENGTH_RANGE = (0.3, 1.0)
794. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.BRIDGE_AUG.THICKNESS_RANGE = (4, 9)
795. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.BRIDGE_AUG.N_LINES = (1, 3)
796. # Heavy acquisition artifact. With probability BAND_PROB: a border-to-border band
797. # (BAND_THICKNESS_RANGE px) -- membrane channel set to 1 inside it, every other channel
798. # blacked out; rest of the image untouched. Otherwise: BLOB_N_RANGE ink-blot blobs
799. # (BLOB_SIZE_RANGE radius, fraction of min(h, w)) blacked out in every channel.
800. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG = CN()
801. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG.ENABLE = False
802. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG.PROB = 0.1
803. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG.BAND_PROB = 0.5
804. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG.BAND_THICKNESS_RANGE = (50, 70)
805. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG.BLOB_SIZE_RANGE = (0.1, 0.3)
806. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ARTIFACT_AUG.BLOB_N_RANGE = (1, 3)
807. # Applies a small random dilation/erosion/spur injection to the membrane channel before the
808. # derived channels are computed, so the model doesn't over-trust exact skeleton geometry.
809. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SKELETON_PERTURB_AUG = CN()
810. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SKELETON_PERTURB_AUG.ENABLE = False
811. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SKELETON_PERTURB_AUG.PROB = 0.3
812. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SKELETON_PERTURB_AUG.RADIUS_RANGE = (1, 2)
813. # Zeroes z-slices of each raw source channel (SOURCE_CHANNELS, e.g. membrane or raw)
814. # independently with probability PROB (see 'slice_dropout' in
815. # biapy/data/generators/membrane_augmentors.py).
816. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SLICE_DROPOUT_AUG = CN()
817. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SLICE_DROPOUT_AUG.ENABLE = False
818. _C.PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.SLICE_DROPOUT_AUG.PROB = 0.3
820. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
821. # 3. Dataset
822. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
823. _C.DATA = CN()
824. # Save all data of a generator in the given path.
825. _C.DATA.CHECK_GENERATORS = False
826. # _C.PROBLEM.NDIM='2D' -> _C.DATA.PATCH_SIZE=(y,x,c) ; _C.PROBLEM.NDIM='3D' -> _C.DATA.PATCH_SIZE=(z,y,x,c)
827. _C.DATA.PATCH_SIZE = (256, 256, 1)
828. # Number of classes including the background class (that should be using 0 label)
829. _C.DATA.N_CLASSES = 2
830. # Whether to reshape the dimensions that does not satisfy the patch shape selected by padding it with reflect.
831. _C.DATA.REFLECT_TO_COMPLETE_SHAPE = True
832. # If 'DATA.PATCH_SIZE' selected has 3 channels, e.g. RGB images are expected, so will force grayscale images to be
833. # converted into RGB (e.g. in ImageNet some of the images are grayscale)
834. _C.DATA.FORCE_RGB = False
835. # If filtering is done, with any of DATA.*.FILTER_SAMPLES.* variables, this will decide how this filtering will be done:
836. # * True: apply filter image by image.
837. # * False: apply filtering sample by sample. Each sample represents a patch within an image.
838. _C.DATA.FILTER_BY_IMAGE = False
839. # Determines whether to save filtered images. If 'DATA.FILTER_BY_IMAGE' is enabled, two subfolders will be created: one for
840. # filtered images and another for non-filtered images. Otherwise, no subfolders will be created, and the images will
841. # display filtered patches as black (all zero values) while retaining original patch values in non-filtered areas.
842. _C.DATA.SAVE_FILTERED_IMAGES = False
843. # Number of filtered images to save. Only work when 'DATA.SAVE_FILTERED_IMAGES' is True
844. _C.DATA.SAVE_FILTERED_IMAGES_NUM = 3
846. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
847. # 3.1 Normalization options for the data
848. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
849. _C.DATA.NORMALIZATION = CN()
850. # Whether to apply or not a percentile clipping before normalizing the data
851. _C.DATA.NORMALIZATION.PERC_CLIP = CN()
852. _C.DATA.NORMALIZATION.PERC_CLIP.ENABLE = False
853. # Lower and upper bound for percentile clip. Must be set when DATA.NORMALIZATION.PERC_CLIP.ENABLE = 'True'
854. _C.DATA.NORMALIZATION.PERC_CLIP.LOWER_PERC = -1.0
855. _C.DATA.NORMALIZATION.PERC_CLIP.UPPER_PERC = -1.0
856. # Lower and upper values to clip. If these are provided the percentiles are not calculated based on
857. # 'DATA.NORMALIZATION.PERC_CLIP.LOWER_PERC' and 'DATA.NORMALIZATION.PERC_CLIP.UPPER_PERC'
858. _C.DATA.NORMALIZATION.PERC_CLIP.LOWER_VALUE = [-1.0]
859. _C.DATA.NORMALIZATION.PERC_CLIP.UPPER_VALUE = [-1.0]
860. # Normalization type to use. Possible options:
861. # 'div' to divide values from 0/255 (or 0/65535 if uint16) in [0,1] range
862. # 'scale_range' same as 'div' but scaling the range to [0-max] and then dividing by the maximum value of the data
863. # and not by 255 or 65535
864. # 'zero_mean_unit_variance' to substract the mean and divide by std.
865. _C.DATA.NORMALIZATION.TYPE = "zero_mean_unit_variance"
866. # Custom normalization variables: mean and std (they are calculated if not provided)
867. _C.DATA.NORMALIZATION.ZERO_MEAN_UNIT_VAR = CN()
868. _C.DATA.NORMALIZATION.ZERO_MEAN_UNIT_VAR.MEAN_VAL = [-1.0]
869. _C.DATA.NORMALIZATION.ZERO_MEAN_UNIT_VAR.STD_VAL = [-1.0]
871. # Target/GT-specific normalization (image-to-image style targets only). Off by default = fully
872. # retrocompatible (target normalized like the input, as before). When enabled, the target is
873. # normalized with its own FIXED type/mean/std/clip below, and predictions are un-normalized with
874. # those same fixed values at test time instead of the input's - needed because the input's stats
875. # are computable at test time (the input always exists) but the target's are not.
876. _C.DATA.NORMALIZATION.TARGET = CN()
877. _C.DATA.NORMALIZATION.TARGET.ENABLE = False
878. # '' reuses 'DATA.NORMALIZATION.TYPE'. Only 'zero_mean_unit_variance' is supported here.
879. _C.DATA.NORMALIZATION.TARGET.TYPE = ""
880. _C.DATA.NORMALIZATION.TARGET.PERC_CLIP = CN()
881. _C.DATA.NORMALIZATION.TARGET.PERC_CLIP.ENABLE = False
882. # Fixed clip values only (no percentiles - those would need the target image at test time).
883. _C.DATA.NORMALIZATION.TARGET.PERC_CLIP.LOWER_VALUE = [-1.0]
884. _C.DATA.NORMALIZATION.TARGET.PERC_CLIP.UPPER_VALUE = [-1.0]
885. _C.DATA.NORMALIZATION.TARGET.ZERO_MEAN_UNIT_VAR = CN()
886. # Required (not -1) when 'DATA.NORMALIZATION.TARGET.ENABLE' is True.
887. _C.DATA.NORMALIZATION.TARGET.ZERO_MEAN_UNIT_VAR.MEAN_VAL = [-1.0]
888. _C.DATA.NORMALIZATION.TARGET.ZERO_MEAN_UNIT_VAR.STD_VAL = [-1.0]
890. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
891. # 3.2 Training data options
892. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
893. _C.DATA.TRAIN = CN()
894. # Whether to check if the data mask contains correct values, e.g. same classes as defined
895. _C.DATA.TRAIN.CHECK_DATA = True
896. _C.DATA.TRAIN.IN_MEMORY = True
897. _C.DATA.TRAIN.PATH = os.path.join("user_data", "train", "x")
898. _C.DATA.TRAIN.GT_PATH = os.path.join("user_data", "train", "y")
899. # Whether your input Zarr contains the raw images and labels together or not. Use 'DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_RAW_PATH'
900. # and 'DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_GT_PATH' to determine the tag to find within the Zarr
901. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA = False
902. # Paths to the raw and gt within the Zarr file. Only used when 'DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA' is True.
903. # E.g. 'volumes.raw' for raw and 'volumes.labels.neuron_ids' for GT path.
904. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_RAW_PATH = ""
905. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_GT_PATH = ""
906. # For synapse detection. The information must be stored as CREMI dataset (https://cremi.org/data/)
907. # Path within the file where the ``ids`` are stored. Reference in CREMI: ``annotations/ids``
908. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_ID_PATH = "annotations.ids"
909. # Path within the file where the ``types`` are stored (not used). Reference in CREMI: ``annotations/types``
910. # _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_TYPES_PATH = "annotations.types"
911. # Path within the file where the ``partners`` are stored. Reference in CREMI: ``annotations/partners``
912. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_PARTNERS_PATH = "annotations.presynaptic_site.partners"
913. # Path within the file where the ``locations`` are stored. Reference in CREMI: ``annotations/locations``
914. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_LOCATIONS_PATH = "annotations.locations"
915. # Path within the file where the ``resolution`` is stored. Reference in CREMI: ``["volumes/raw"].attrs["offset"]``
916. _C.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA_RESOLUTION_PATH = 'volumes.raw'
917. # File to load/save data prepared with the appropiate channels in a instance segmentation problem.
918. # E.g. _C.PROBLEM.TYPE ='INSTANCE_SEG' and _C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS != 'B'
919. _C.DATA.TRAIN.INSTANCE_CHANNELS_MASK_DIR = os.path.join(
920. "user_data", "train", "y_" + "".join(_C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS)
921. )
922. # Path to load/save detection masks prepared.
923. _C.DATA.TRAIN.DETECTION_MASK_DIR = os.path.join("user_data", "train", "y_detection_masks")
924. # Path to load/save SSL target prepared.
925. _C.DATA.TRAIN.SSL_SOURCE_DIR = os.path.join("user_data", "train", "x_ssl_source")
926. # Extract random patches during data augmentation (DA)
927. _C.DATA.TRAIN.EXTRACT_RANDOM_PATCH = False
928. # Create a probability map so the patches extracted will have a high probability of having an object in the middle
929. # of it. Useful to avoid extracting patches which no foreground class information. Use it only when
930. # 'PROBLEM.TYPE' is 'SEMANTIC_SEG', 'INSTANCE_SEG' or 'DETECTION'
931. _C.DATA.TRAIN.PROBABILITY_MAP = False # Used when _C.DATA.TRAIN.EXTRACT_RANDOM_PATCH=True
932. _C.DATA.TRAIN.W_FOREGROUND = 0.94 # Used when _C.DATA.TRAIN.PROBABILITY_MAP=True
933. _C.DATA.TRAIN.W_BACKGROUND = 0.06 # Used when _C.DATA.TRAIN.PROBABILITY_MAP=True
934. # Extra train data generation: number of times to duplicate the train data. Useful when
935. # _C.DATA.TRAIN.EXTRACT_RANDOM_PATCH=True is made, as more original train data can be cover on each epoch
936. _C.DATA.TRAIN.REPLICATE = 0
937. # Percentage of overlap in (y,x)/(z,y,x) when cropping validation. Set to 0 to calculate the minimun overlap.
938. # The values must be floats between range [0, 1). It needs to be a 2D tuple when using _C.PROBLEM.NDIM='2D' and
939. # 3D tuple when using _C.PROBLEM.NDIM='3D'
940. _C.DATA.TRAIN.OVERLAP = (0, 0)
941. # Padding to be done in (y,x)/(z,y,x) when reconstructing train data. Useful to avoid patch 'border effect'.
942. _C.DATA.TRAIN.PADDING = (0, 0)
943. # Train data resolution. It is not completely necessary but when configured it is taken into account when
944. # performing some augmentations, e.g. cutout. If defined it need to be (y,x)/(z,y,x) and needs to be to be a 2D
945. # tuple when using _C.PROBLEM.NDIM='2D' and 3D tuple when using _C.PROBLEM.NDIM='3D'
946. _C.DATA.TRAIN.RESOLUTION = (-1,)
947. # Order of the axes of the image when using Zarr/H5 images in train data.
948. _C.DATA.TRAIN.INPUT_IMG_AXES_ORDER = "TZCYX"
949. # Order of the axes of the mask when using Zarr/H5 images in train data.
950. _C.DATA.TRAIN.INPUT_MASK_AXES_ORDER = "TZCYX"
952. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
953. # 3.2.1 Training data filtering options
954. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
955. # DATA.TRAIN.FILTER_SAMPLES allows removing training images by the conditions based on their properties. When using Zarr each patch within the Zarr will be
956. # processed and will not depend on 'DATA.FILTER_BY_IMAGE' variable.
957. # Its three variables (PROPS, VALUES and SIGNS) define a set of conditions to remove the images from the training set. If an image satisfies any of the
958. # conditions, the image won't be used for training.
959. #
960. # In PROPS, we define the property to look at to establish the condition. The available properties are: ['foreground', 'mean', 'min', 'max',
961. # 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio'].
962. #
963. # * 'foreground' is defined as the percentage of pixels/voxels corresponding to the foreground mask. This option is only valid for
964. # SEMANTIC_SEG, INSTANCE_SEG and DETECTION.
965. # * 'mean' is defined as the mean intensity value of the raw image inputs.
966. # * 'min' is defined as the min intensity value of the raw image inputs.
967. # * 'max' is defined as the max intensity value of the raw image inputs.
968. # * 'diff' is defined as the difference between ground truth and raw images. Available for all workflows but SELF_SUPERVISED and DENOISING.
969. # * 'diff_by_min_max_ratio' is defined as the difference between ground truth and raw images multiplied by the ratio between raw image max and min. Available for all workflows but SELF_SUPERVISED and DENOISING.
970. # * 'target_mean' is defined as the mean intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
971. # * 'target_min' is defined as the min intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
972. # * 'target_max' is defined as the max intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
973. # * 'diff_by_target_min_max_ratio' is defined as the difference between ground truth and raw images multiplied by the ratio between ground truth image max and min. Available for all workflows but SELF_SUPERVISED and DENOISING.
974. #
975. # With VALUES and SIGNS, we define the specific values and the comparison operators of each property, respectively.
976. # The available operators are: ['gt', 'ge', 'lt', 'le'], that corresponds to "greather than" (or ">"), "greather equal" (or ">="), "less than" (or "<"),
977. # and "less equal" (or "<=").
978. #
979. # Here you have a full example of this filtering:
980. # If you want to remove those samples that have intensity values lower than 0.00001 and a mean average greater than 100 you should
981. # declare the above three variables as follows (notice you need to know the image data type in advance):
982. # _C.DATA.TRAIN.FILTER_SAMPLES.PROPS = [['foreground','mean']]
983. # _C.DATA.TRAIN.FILTER_SAMPLES.VALUES = [[0.00001, 100]]
984. # _C.DATA.TRAIN.FILTER_SAMPLES.SIGNS = [['lt', 'gt']]
985. # You can also concatenate more restrictions and they will be applied in order. For instance, if you want to filter those
986. # samples with a maximum intensity value greater than 1000, and do that before the condition described above, you can define the
987. # variables this way:
988. # _C.DATA.TRAIN.FILTER_SAMPLES.PROPS = [['max'], ['foreground','mean']]
989. # _C.DATA.TRAIN.FILTER_SAMPLES.VALUES = [[1000], [0.00001, 100]]
990. # _C.DATA.TRAIN.FILTER_SAMPLES.SIGNS = [['gt'], ['lt', 'gt']]
991. # This way, the images will be removed by 'max' and then by 'foreground' and 'mean'
992. _C.DATA.TRAIN.FILTER_SAMPLES = CN()
993. # Whether to enable or not the filtering by properties
994. _C.DATA.TRAIN.FILTER_SAMPLES.ENABLE = False
995. # List of lists of properties to apply a filter. Available properties are: ['foreground', 'mean', 'min', 'max',
996. # 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio']
997. _C.DATA.TRAIN.FILTER_SAMPLES.PROPS = []
998. # List of ints/float that represent the values of the properties listed in 'DATA.TRAIN.FILTER_SAMPLES.PROPS'
999. # that the images need to satisfy to not be dropped.
1000. _C.DATA.TRAIN.FILTER_SAMPLES.VALUES = []
1001. # List of list of signs to do the comparison. Options: ['gt', 'ge', 'lt', 'le'] that corresponds to "greather than", e.g. ">",
1002. # "greather equal", e.g. ">=", "less than", e.g. "<", and "less equal" e.g. "<=" comparisons.
1003. _C.DATA.TRAIN.FILTER_SAMPLES.SIGNS = []
1004. # Whether to normalize the samples before comparison
1005. _C.DATA.TRAIN.FILTER_SAMPLES.NORM_BEFORE = False
1008. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1009. # 3.3 Validation data options
1010. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1011. _C.DATA.VAL = CN()
1012. # Enabling distributed evaluation (recommended during training)
1013. _C.DATA.VAL.DIST_EVAL = True
1014. # Whether to create validation data from training set or read it from a directory
1015. _C.DATA.VAL.FROM_TRAIN = True
1016. # Use a cross validation strategy instead of just split the train data in two
1017. _C.DATA.VAL.CROSS_VAL = False
1018. # Number of folds. Used when _C.DATA.VAL.CROSS_VAL == True
1019. _C.DATA.VAL.CROSS_VAL_NFOLD = 5
1020. # Number of the fold to choose as validation. Used when _C.DATA.VAL.CROSS_VAL == True
1021. _C.DATA.VAL.CROSS_VAL_FOLD = 1
1022. # Percentage of the training data used as validation. Used when _C.DATA.VAL.FROM_TRAIN = True and _C.DATA.VAL.CROSS_VAL == False
1023. _C.DATA.VAL.SPLIT_TRAIN = 0.1
1024. # Create the validation data with random images of the training data. Used when _C.DATA.VAL.FROM_TRAIN = True
1025. _C.DATA.VAL.RANDOM = True
1026. # Used when _C.DATA.VAL.FROM_TRAIN = False, as DATA.VAL.FROM_TRAIN = True always implies DATA.VAL.IN_MEMORY = True
1027. _C.DATA.VAL.IN_MEMORY = True
1028. # Path to the validation data. Used when _C.DATA.VAL.FROM_TRAIN = False
1029. _C.DATA.VAL.PATH = os.path.join("user_data", "val", "x")
1030. # Path to the validation data mask. Used when _C.DATA.VAL.FROM_TRAIN = False
1031. _C.DATA.VAL.GT_PATH = os.path.join("user_data", "val", "y")
1032. # Whether your input Zarr contains the raw images and labels together or not. Use 'DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_RAW_PATH'
1033. # and 'DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_GT_PATH' to determine the tag to find within the Zarr
1034. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA = False
1035. # Paths to the raw and gt within the Zarr file. Only used when 'DATA.VAL.INPUT_ZARR_MULTIPLE_DATA' is True.
1036. # E.g. 'volumes.raw' for raw and 'volumes.labels.neuron_ids' for GT path.
1037. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_RAW_PATH = ""
1038. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_GT_PATH = ""
1039. # For synapse detection. The information must be stored as CREMI dataset (https://cremi.org/data/)
1040. # Path within the file where the ``ids`` are stored. Reference in CREMI: ``annotations/ids``
1041. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_ID_PATH = "annotations.ids"
1042. # Path within the file where the ``types`` are stored (not used). Reference in CREMI: ``annotations/types``
1043. # _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_TYPES_PATH = "annotations.types"
1044. # Path within the file where the ``partners`` are stored. Reference in CREMI: ``annotations/partners``
1045. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_PARTNERS_PATH = "annotations.presynaptic_site.partners"
1046. # Path within the file where the ``locations`` are stored. Reference in CREMI: ``annotations/locations``
1047. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_LOCATIONS_PATH = "annotations.locations"
1048. # Path within the file where the ``resolution`` is stored. Reference in CREMI: ``["volumes/raw"].attrs["offset"]``
1049. _C.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA_RESOLUTION_PATH = 'volumes.raw'
1050. # File to load/save data prepared with the appropiate channels in a instance segmentation problem.
1051. # E.g. _C.PROBLEM.TYPE ='INSTANCE_SEG' and _C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS != 'B'
1052. _C.DATA.VAL.INSTANCE_CHANNELS_MASK_DIR = os.path.join(
1053. "user_data", "val", "y_" + "".join(_C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS)
1054. )
1055. # Path to load/save detection masks prepared.
1056. _C.DATA.VAL.DETECTION_MASK_DIR = os.path.join("user_data", "val", "y_detection_masks")
1057. # Path to load/save SSL target prepared.
1058. _C.DATA.VAL.SSL_SOURCE_DIR = os.path.join("user_data", "val", "x_ssl_source")
1059. # Percentage of overlap in (y,x)/(z,y,x) when cropping validation. Set to 0 to calculate the minimun overlap.
1060. # The values must be floats between range [0, 1). It needs to be a 2D tuple when using _C.PROBLEM.NDIM='2D' and
1061. # 3D tuple when using _C.PROBLEM.NDIM='3D'. This is only used when the validation is loaded from disk, and thus,
1062. # not extracted from training.
1063. _C.DATA.VAL.OVERLAP = (0, 0)
1064. # Padding to be done in (y,x)/(z,y,x) when cropping validation data. Useful to avoid patch 'border effect'. This
1065. # is only used when the validation is loaded from disk, and thus, not extracted from training.
1066. _C.DATA.VAL.PADDING = (0, 0)
1067. # Not used yet.
1068. _C.DATA.VAL.RESOLUTION = (-1,)
1069. # Order of the axes of the image when using Zarr/H5 images in validation data.
1070. _C.DATA.VAL.INPUT_IMG_AXES_ORDER = "TZCYX"
1071. # Order of the axes of the mask when using Zarr/H5 images in validation data.
1072. _C.DATA.VAL.INPUT_MASK_AXES_ORDER = "TZCYX"
1073. # Remove validation images by the conditions based on their properties. When using Zarr each patch within the Zarr will be processed and will
1074. # not depend on 'DATA.FILTER_BY_IMAGE' variable
1075. # The three variables, DATA.VAL.FILTER_SAMPLES.PROPS, DATA.VAL.FILTER_SAMPLES.VALUES and DATA.VAL.FILTER_SAMPLES.SIGNS will compose a list of
1076. # conditions to remove the images. They are list of list of conditions. For instance, the conditions can be like this: [['A'], ['B','C']]. Then,
1077. # if the image satisfies the first list of conditions, only 'A' in this first case (from ['A'] list), or satisfy 'B' and 'C' (from ['B','C'] list)
1078. # it will be removed from the image. In each sublist all the conditions must be satisfied. Available properties are: ['foreground', 'mean', 'min', 'max',
1079. # 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio'].
1080. #
1081. # Each property descrition:
1082. # * 'foreground' is defined as the percentage of pixels/voxels corresponding to the foreground mask. This option is only valid for
1083. # SEMANTIC_SEG, INSTANCE_SEG and DETECTION.
1084. # * 'mean' is defined as the mean intensity value of the raw image inputs.
1085. # * 'min' is defined as the min intensity value of the raw image inputs.
1086. # * 'max' is defined as the max intensity value of the raw image inputs.
1087. # * 'diff' is defined as the difference between ground truth and raw images. Available for all workflows but SELF_SUPERVISED and DENOISING.
1088. # * 'diff_by_min_max_ratio' is defined as the difference between ground truth and raw images multiplied by the ratio between raw image max and min. Available for all workflows but SELF_SUPERVISED and DENOISING.
1089. # * 'target_mean' is defined as the mean intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
1090. # * 'target_min' is defined as the min intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
1091. # * 'target_max' is defined as the max intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
1092. # * 'diff_by_target_min_max_ratio' is defined as the difference between ground truth and raw images multiplied by the ratio between ground truth image max and min. Available for all workflows but SELF_SUPERVISED and DENOISING.
1093. #
1094. # A full example of this filtering:
1095. # If you want to remove those samples that have less than 0.00001 and a mean average more than 100 (you need to know image data type) you should
1096. # declare the above three variables as follows:
1097. # _C.DATA.VAL.FILTER_SAMPLES.PROPS = [['foreground','mean']]
1098. # _C.DATA.VAL.FILTER_SAMPLES.VALUES = [[0.00001, 100]]
1099. # _C.DATA.VAL.FILTER_SAMPLES.SIGNS = [['lt', 'gt']]
1100. # You can also concatenate more restrictions and they will be applied in order. For instance, if you want to filter those
1101. # samples with a max value more than 1000, and do that before the condition described above, you can define the
1102. # variables this way:
1103. # _C.DATA.VAL.FILTER_SAMPLES.PROPS = [['max'], ['foreground','mean']]
1104. # _C.DATA.VAL.FILTER_SAMPLES.VALUES = [[1000], [0.00001, 100]]
1105. # _C.DATA.VAL.FILTER_SAMPLES.SIGNS = [['gt'], ['lt', 'gt']]
1106. # This way, the images will be removed by 'max' and then by 'foreground' and 'mean'
1107. _C.DATA.VAL.FILTER_SAMPLES = CN()
1108. # Whether to enable or not the filtering by properties
1109. _C.DATA.VAL.FILTER_SAMPLES.ENABLE = False
1110. # List of lists of properties to apply a filter. Available properties are: ['foreground', 'mean', 'min', 'max',
1111. # 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio']
1112. _C.DATA.VAL.FILTER_SAMPLES.PROPS = []
1113. # List of ints/float that represent the values of the properties listed in 'DATA.VAL.FILTER_SAMPLES.PROPS'
1114. # that the images need to satisfy to not be dropped.
1115. _C.DATA.VAL.FILTER_SAMPLES.VALUES = []
1116. # List of list of signs to do the comparison. Options: ['gt', 'ge', 'lt', 'le'] that corresponds to "greather than", e.g. ">",
1117. # "greather equal", e.g. ">=", "less than", e.g. "<", and "less equal" e.g. "<=" comparisons.
1118. _C.DATA.VAL.FILTER_SAMPLES.SIGNS = []
1119. # Whether to normalize the samples before comparison
1120. _C.DATA.VAL.FILTER_SAMPLES.NORM_BEFORE = False
1122. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1123. # 3.4 Test data options
1124. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1125. _C.DATA.TEST = CN()
1126. # Whether to check if the data mask contains correct values, e.g. same classes as defined
1127. _C.DATA.TEST.CHECK_DATA = True
1128. _C.DATA.TEST.IN_MEMORY = False
1129. # Whether to load ground truth (GT)
1130. _C.DATA.TEST.LOAD_GT = False
1131. # Whether to use validation data as test instead of trying to load test from _C.DATA.TEST.PATH and
1132. # _C.DATA.TEST.GT_PATH. _C.DATA.VAL.CROSS_VAL needs to be True.
1133. _C.DATA.TEST.USE_VAL_AS_TEST = False
1134. # Path to load the test data from. Not used when _C.DATA.TEST.USE_VAL_AS_TEST == True
1135. _C.DATA.TEST.PATH = os.path.join("user_data", "test", "x")
1136. # Path to load the test data masks from. Not used when _C.DATA.TEST.USE_VAL_AS_TEST == True
1137. _C.DATA.TEST.GT_PATH = os.path.join("user_data", "test", "y")
1138. # File to load/save data prepared with the appropiate channels in a instance segmentation problem.
1139. # E.g. _C.PROBLEM.TYPE ='INSTANCE_SEG' and _C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS != 'B'
1140. _C.DATA.TEST.INSTANCE_CHANNELS_MASK_DIR = os.path.join(
1141. "user_data", "test", "y_" + "".join(_C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS)
1142. )
1143. # Path to load/save detection masks prepared.
1144. _C.DATA.TEST.DETECTION_MASK_DIR = os.path.join("user_data", "test", "y_detection_masks")
1145. # Path to load/save SSL target prepared.
1146. _C.DATA.TEST.SSL_SOURCE_DIR = os.path.join("user_data", "test", "x_ssl_source")
1147. # Percentage of overlap in (y,x)/(z,y,x) when cropping validation. Set to 0 to calculate the minimun overlap.
1148. # The values must be floats between range [0, 1). It needs to be a 2D tuple when using _C.PROBLEM.NDIM='2D' and
1149. # 3D tuple when using _C.PROBLEM.NDIM='3D'
1150. _C.DATA.TEST.OVERLAP = (0, 0)
1151. # Padding to be done in (y,x)/(z,y,xz) when reconstructing test data. Useful to avoid patch 'border effect'
1152. _C.DATA.TEST.PADDING = (0, 0)
1153. # Whether to use median values to fill padded pixels or zeros
1154. _C.DATA.TEST.MEDIAN_PADDING = False
1155. # Region of interest (ROI) mask restricting where the inference is done. Anything greater than 0 in the mask is
1156. # ROI. With 'TEST.BY_CHUNKS.ENABLE' the patches not overlapping the ROI are not predicted, and are left as
1157. # background in the output Zarr; otherwise the prediction is zeroed outside the ROI. The mask does not need to
1158. # have the same shape as the test image: it is mapped to it by scaling each axis, so a coarse mask (e.g. one
1159. # voxel per 64x64x64 block) is enough.
1160. _C.DATA.TEST.ROI_MASK = CN()
1161. # Whether to restrict the inference to the ROI mask
1162. _C.DATA.TEST.ROI_MASK.ENABLE = False
1163. # Path to the ROI mask. It can be a file (used for all the test samples) or a directory. In a directory, if only one
1164. # mask is found it is used for all the test samples, otherwise the mask named as the test sample is used for it.
1165. _C.DATA.TEST.ROI_MASK.PATH = ""
1166. # Order of the axes of the ROI mask. Leave it empty to use the axes order of the test image:
1167. # 'DATA.TEST.INPUT_IMG_AXES_ORDER' when reading Zarr/H5 by chunks and 'ZYX'/'YX' otherwise.
1168. _C.DATA.TEST.ROI_MASK.AXES_ORDER = ""
1170. # Test data resolution. Need to be provided in (z,y,x) order. Only applies when _C.PROBLEM.TYPE = 'DETECTION' now.
1171. _C.DATA.TEST.RESOLUTION = (-1,)
1172. # Whether to read the per-image test resolution from a 'resolution.json' file placed in the parent
1173. # folder of 'DATA.TEST.PATH' (a mapping of image basename -> (z, y, x) resolution) instead of using
1174. # 'DATA.TEST.RESOLUTION'. Images missing from the JSON fall back to 'DATA.TEST.RESOLUTION'.
1175. _C.DATA.TEST.RESOLUTION_FROM_JSON = False
1176. # Order of the axes of the image when using Zarr/H5 images in test data.
1177. _C.DATA.TEST.INPUT_IMG_AXES_ORDER = "TZCYX"
1178. # Order of the axes of the mask when using Zarr/H5 images in test data.
1179. _C.DATA.TEST.INPUT_MASK_AXES_ORDER = "TZCYX"
1180. # Whether your input Zarr contains the raw images and labels together or not. Use 'DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_RAW_PATH'
1181. # and 'DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_GT_PATH' to determine the tag to find within the Zarr
1182. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA = False
1183. # Paths to the raw and gt within the Zarr file. Only used when 'DATA.TEST.INPUT_ZARR_MULTIPLE_DATA' is True.
1184. # E.g. 'volumes.raw' for raw and 'volumes.labels.neuron_ids' for GT path.
1185. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_RAW_PATH = ""
1186. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_GT_PATH = ""
1187. # For synapse detection. The information must be stored as CREMI dataset (https://cremi.org/data/)
1188. # Path within the file where the ``ids`` are stored. Reference in CREMI: ``annotations/ids``
1189. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_ID_PATH = "annotations.ids"
1190. # Path within the file where the ``types`` are stored (not used). Reference in CREMI: ``annotations/types``
1191. # _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_TYPES_PATH = "annotations.types"
1192. # Path within the file where the ``partners`` are stored. Reference in CREMI: ``annotations/partners``
1193. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_PARTNERS_PATH = "annotations.presynaptic_site.partners"
1194. # Path within the file where the ``locations`` are stored. Reference in CREMI: ``annotations/locations``
1195. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_LOCATIONS_PATH = "annotations.locations"
1196. # Path within the file where the ``resolution`` is stored. Reference in CREMI: ``["volumes/raw"].attrs["offset"]``
1197. _C.DATA.TEST.INPUT_ZARR_MULTIPLE_DATA_RESOLUTION_PATH = 'volumes.raw'
1198. # Remove test images by the conditions based on their properties. When using Zarr each patch within the Zarr will be processed and will
1199. # not depend on 'DATA.FILTER_BY_IMAGE' variable
1200. # The three variables, DATA.TEST.FILTER_SAMPLES.PROPS, DATA.TEST.FILTER_SAMPLES.VALUES and DATA.TEST.FILTER_SAMPLES.SIGNS will compose a
1201. # list of conditions to remove the images. They are list of list of conditions. For instance, the conditions can be like this: [['A'], ['B','C']].
1202. # Then, if the image satisfies the first list of conditions, only 'A' in this first case (from ['A'] list), or satisfy 'B' and 'C' (from ['B','C'] list)
1203. # it will be removed from the image. In each sublist all the conditions must be satisfied. Available properties are: ['foreground', 'mean', 'min', 'max',
1204. # 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio'].
1205. #
1206. # Each property descrition:
1207. # * 'foreground' is defined as the percentage of pixels/voxels corresponding to the foreground mask. This option is only valid for
1208. # SEMANTIC_SEG, INSTANCE_SEG and DETECTION.
1209. # * 'mean' is defined as the mean intensity value of the raw image inputs.
1210. # * 'min' is defined as the min intensity value of the raw image inputs.
1211. # * 'max' is defined as the max intensity value of the raw image inputs.
1212. # * 'diff' is defined as the difference between ground truth and raw images. Available for all workflows but SELF_SUPERVISED and DENOISING.
1213. # * 'diff_by_min_max_ratio' is defined as the difference between ground truth and raw images multiplied by the ratio between raw image max and min. Available for all workflows but SELF_SUPERVISED and DENOISING.
1214. # * 'target_mean' is defined as the mean intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
1215. # * 'target_min' is defined as the min intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
1216. # * 'target_max' is defined as the max intensity value of the raw image targets. Available for all workflows but SELF_SUPERVISED and DENOISING.
1217. # * 'diff_by_target_min_max_ratio' is defined as the difference between ground truth and raw images multiplied by the ratio between ground truth image max and min. Available for all workflows but SELF_SUPERVISED and DENOISING.
1218. #
1219. #
1220. # A full example of this filtering:
1221. # If you want to remove those samples that have less than 0.00001 and a mean average more than 100 (you need to know image data type) you should
1222. # declare the above three variables as follows:
1223. # _C.DATA.TEST.FILTER_SAMPLES.PROPS = [['foreground','mean']]
1224. # _C.DATA.TEST.FILTER_SAMPLES.VALUES = [[0.00001, 100]]
1225. # _C.DATA.TEST.FILTER_SAMPLES.SIGNS = [['lt', 'gt']]
1226. # You can also concatenate more restrictions and they will be applied in order. For instance, if you want to filter those
1227. # samples with a max value more than 1000, and do that before the condition described above, you can define the
1228. # variables this way:
1229. # _C.DATA.TEST.FILTER_SAMPLES.PROPS = [['max'], ['foreground','mean']]
1230. # _C.DATA.TEST.FILTER_SAMPLES.VALUES = [[1000], [0.00001, 100]]
1231. # _C.DATA.TEST.FILTER_SAMPLES.SIGNS = [['gt'], ['lt', 'gt']]
1232. # This way, the images will be removed by 'max' and then by 'foreground' and 'mean'
1233. _C.DATA.TEST.FILTER_SAMPLES = CN()
1234. # Whether to enable or not the filtering by properties
1235. _C.DATA.TEST.FILTER_SAMPLES.ENABLE = False
1236. # List of lists of properties to apply a filter. Available properties are: ['foreground', 'mean', 'min', 'max',
1237. # 'target_mean', 'target_min', 'target_max', 'diff', 'diff_by_min_max_ratio', 'diff_by_target_min_max_ratio']
1238. _C.DATA.TEST.FILTER_SAMPLES.PROPS = []
1239. # List of ints/float that represent the values of the properties listed in 'DATA.TEST.FILTER_SAMPLES.PROPS'
1240. # that the images need to satisfy to not be dropped.
1241. _C.DATA.TEST.FILTER_SAMPLES.VALUES = []
1242. # List of list of signs to do the comparison. Options: ['gt', 'ge', 'lt', 'le'] that corresponds to "greather than", e.g. ">",
1243. # "greather equal", e.g. ">=", "less than", e.g. "<", and "less equal" e.g. "<=" comparisons.
1244. _C.DATA.TEST.FILTER_SAMPLES.SIGNS = []
1245. # Whether to normalize the samples before comparison
1246. _C.DATA.TEST.FILTER_SAMPLES.NORM_BEFORE = False
1248. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1249. # 3.5 Pre-processing data options
1250. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1251. # Same preprocessing will be applied to all selected datasets
1252. _C.DATA.PREPROCESS = CN()
1253. # Apply preprocessing to training dataset
1254. _C.DATA.PREPROCESS.TRAIN = False
1255. # Apply preprocessing to validation dataset
1256. _C.DATA.PREPROCESS.VAL = False
1257. # Apply preprocessing to testing dataset
1258. _C.DATA.PREPROCESS.TEST = False
1260. # Resize datasets
1261. _C.DATA.PREPROCESS.RESIZE = CN()
1262. _C.DATA.PREPROCESS.RESIZE.ENABLE = False
1263. # Desired resize size. when using 3D data, size must be also in 3D (ex. (512,512,512))
1264. _C.DATA.PREPROCESS.RESIZE.OUTPUT_SHAPE = (512, 512)
1265. # interpolation order: {0: Nearest-neighbor, 1: Bi-linear (default), 2: Bi-quadratic, 3: Bi-cubic, 4: Bi-quartic, 5: Bi-quintic}
1266. _C.DATA.PREPROCESS.RESIZE.ORDER = 1
1267. # Points outside the boundaries of the input are filled according to the given mode: {'constant', 'edge', 'symmetric', 'reflect', 'wrap'}
1268. _C.DATA.PREPROCESS.RESIZE.MODE = "reflect"
1269. # Used in conjunction with mode 'constant', the value outside the image boundaries.
1270. _C.DATA.PREPROCESS.RESIZE.CVAL = 0.0
1271. # Whether to clip the output to the range of values of the input image.
1272. _C.DATA.PREPROCESS.RESIZE.CLIP = True
1273. # Whether to keep the original range of values.
1274. _C.DATA.PREPROCESS.RESIZE.PRESERVE_RANGE = True
1275. # Whether to apply a Gaussian filter to smooth the image prior to downsampling.
1276. _C.DATA.PREPROCESS.RESIZE.ANTI_ALIASING = False
1278. # Zoom datasets.
1279. _C.DATA.PREPROCESS.ZOOM = CN()
1280. _C.DATA.PREPROCESS.ZOOM.ENABLE = False
1281. # WARNING: Only implemented for _C.TEST.BY_CHUNKS = True. It will change the zoom of each patch individually.
1282. # This is useful when the input image has a different resolution than the one used in the training. The value
1283. # is the zoom factor to be applied to each patch using scipy.ndimage.zoom.
1284. # "E.g. [1,2,1,3,3] that needs to match _C.DATA.TEST.INPUT_IMG_AXES_ORDER axes"
1285. _C.DATA.PREPROCESS.ZOOM.ZOOM_FACTOR = [1, 1, 1, 1, 1]
1287. # Gaussian blur
1288. _C.DATA.PREPROCESS.GAUSSIAN_BLUR = CN()
1289. _C.DATA.PREPROCESS.GAUSSIAN_BLUR.ENABLE = False
1290. # Standard deviation for Gaussian kernel.
1291. _C.DATA.PREPROCESS.GAUSSIAN_BLUR.SIGMA = 1
1292. # The mode parameter determines how the array borders are handled: {'reflect', 'constant', 'nearest', 'mirror', 'wrap'} 'constant' value = 0
1293. _C.DATA.PREPROCESS.GAUSSIAN_BLUR.MODE = "nearest"
1294. # If None, the image is assumed to be a grayscale (single channel) image.
1295. # Otherwise, this parameter indicates which axis of the array corresponds to channels.
1296. _C.DATA.PREPROCESS.GAUSSIAN_BLUR.CHANNEL_AXIS = None
1298. # Median blur
1299. _C.DATA.PREPROCESS.MEDIAN_BLUR = CN()
1300. _C.DATA.PREPROCESS.MEDIAN_BLUR.ENABLE = False
1301. # Desired kernel size (including channels). When using 3D data, size must be also in 3D (ex. (3,7,7,1) for (z,y,x,c))
1302. _C.DATA.PREPROCESS.MEDIAN_BLUR.KERNEL_SIZE = (3,3,1)
1304. # Histogram matching. More info at: https://en.wikipedia.org/wiki/Histogram_matching
1305. _C.DATA.PREPROCESS.MATCH_HISTOGRAM = CN()
1306. _C.DATA.PREPROCESS.MATCH_HISTOGRAM.ENABLE = False
1307. # the path of the reference images, from which the reference histogram will be extracted
1308. _C.DATA.PREPROCESS.MATCH_HISTOGRAM.REFERENCE_PATH = os.path.join("user_data", "test", "x")
1310. # Contrast Limited Adaptive Histogram Equalization. More info at: https://en.wikipedia.org/wiki/Adaptive_histogram_equalization#Contrast_Limited_AHE
1311. _C.DATA.PREPROCESS.CLAHE = CN()
1312. _C.DATA.PREPROCESS.CLAHE.ENABLE = False
1313. # Defines the shape of contextual regions used in the algorithm.
1314. # By default, kernel_size is 1/8 of image height by 1/8 of its width.
1315. _C.DATA.PREPROCESS.CLAHE.KERNEL_SIZE = None
1316. # Clipping limit, normalized between 0 and 1 (higher values give more contrast).
1317. _C.DATA.PREPROCESS.CLAHE.CLIP_LIMIT = 0.01
1319. # Canny or edge detection (only 2D - grayscale or RGB)
1320. _C.DATA.PREPROCESS.CANNY = CN()
1321. _C.DATA.PREPROCESS.CANNY.ENABLE = False
1322. # Lower bound for hysteresis thresholding (linking edges). If None, low_threshold is set to 10% of dtype's max.
1323. _C.DATA.PREPROCESS.CANNY.LOW_THRESHOLD = None
1324. # Upper bound for hysteresis thresholding (linking edges). If None, high_threshold is set to 20% of dtype's max.
1325. _C.DATA.PREPROCESS.CANNY.HIGH_THRESHOLD = None
1327. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1328. # 4. Data augmentation (DA)
1329. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1330. _C.AUGMENTOR = CN()
1331. # Flag to activate DA
1332. _C.AUGMENTOR.ENABLE = False
1333. # Per-augmentation probability of being applied (each in [0, 1]). Every enabled augmentation
1334. # is rolled independently against its own probability; there is no longer a single global
1335. # DA_PROB. The geometric augmentations (ZOOM, RANDOM_ROT, ROT90) are each rolled with their
1336. # own probability and then composed into a single resampling pass (see affine_transform).
1337. _C.AUGMENTOR.ZOOM_PROB = 0.5
1338. _C.AUGMENTOR.RANDOM_RESIZED_CROP_PROB = 0.5
1339. _C.AUGMENTOR.RANDOM_ROT_PROB = 0.5
1340. _C.AUGMENTOR.ROT90_PROB = 0.5
1341. _C.AUGMENTOR.SHEAR_PROB = 0.5
1342. _C.AUGMENTOR.SHIFT_PROB = 0.5
1343. _C.AUGMENTOR.VFLIP_PROB = 0.5
1344. _C.AUGMENTOR.HFLIP_PROB = 0.5
1345. _C.AUGMENTOR.ZFLIP_PROB = 0.5
1346. _C.AUGMENTOR.ELASTIC_PROB = 0.5
1347. _C.AUGMENTOR.G_BLUR_PROB = 0.5
1348. _C.AUGMENTOR.MEDIAN_BLUR_PROB = 0.5
1349. _C.AUGMENTOR.MOTION_BLUR_PROB = 0.5
1350. _C.AUGMENTOR.GAMMA_CONTRAST_PROB = 0.5
1351. _C.AUGMENTOR.BRIGHTNESS_PROB = 0.5
1352. _C.AUGMENTOR.CONTRAST_PROB = 0.5
1353. _C.AUGMENTOR.DROPOUT_PROB = 0.5
1354. _C.AUGMENTOR.CUTOUT_PROB = 0.5
1355. _C.AUGMENTOR.CUTBLUR_PROB = 0.5
1356. _C.AUGMENTOR.CUTMIX_PROB = 0.5
1357. _C.AUGMENTOR.CUTNOISE_PROB = 0.5
1358. _C.AUGMENTOR.MISALIGNMENT_PROB = 0.5
1359. _C.AUGMENTOR.MISSING_SECTIONS_PROB = 0.5
1360. _C.AUGMENTOR.GRAYSCALE_PROB = 0.5
1361. _C.AUGMENTOR.CHANNEL_SHUFFLE_PROB = 0.5
1362. _C.AUGMENTOR.GRIDMASK_PROB = 0.5
1363. _C.AUGMENTOR.GAUSSIAN_NOISE_PROB = 0.5
1364. _C.AUGMENTOR.POISSON_NOISE_PROB = 0.5
1365. _C.AUGMENTOR.SALT_PROB = 0.5
1366. _C.AUGMENTOR.PEPPER_PROB = 0.5
1367. _C.AUGMENTOR.SALT_AND_PEPPER_PROB = 0.5
1368. # Create samples of the DA made. Useful to check the output images made.
1369. _C.AUGMENTOR.AUG_SAMPLES = True
1370. # Draw a grid in the augenation samples generated. Used when _C.AUGMENTOR.AUG_SAMPLES=True
1371. _C.AUGMENTOR.DRAW_GRID = True
1372. # Number of samples to create
1373. _C.AUGMENTOR.AUG_NUM_SAMPLES = 10
1374. # Flag to shuffle the training data on every epoch
1375. _C.AUGMENTOR.SHUFFLE_TRAIN_DATA_EACH_EPOCH = True
1376. # Flag to shuffle the validation data on every epoch
1377. _C.AUGMENTOR.SHUFFLE_VAL_DATA_EACH_EPOCH = False
1378. # Rotation of 90º to the subvolumes
1379. _C.AUGMENTOR.ROT90 = False
1380. # Random rotation between a defined range
1381. _C.AUGMENTOR.RANDOM_ROT = False
1382. # Range of random rotations
1383. _C.AUGMENTOR.RANDOM_ROT_RANGE = (-180, 180)
1384. # Apply shear to images
1385. _C.AUGMENTOR.SHEAR = False
1386. # Shear range. Expected value range is around [-360, 360], with reasonable values being in the range of [-45, 45].
1387. _C.AUGMENTOR.SHEAR_RANGE = (-20, 20)
1388. # Apply zoom to images
1389. _C.AUGMENTOR.ZOOM = False
1390. # Zoom range. Scaling factor to use, where 1.0 denotes “no change” and 0.5 is zoomed out to 50 percent of the original size.
1391. _C.AUGMENTOR.ZOOM_RANGE = (0.5, 1.5)
1392. # Whether to apply or not zoom in Z axis (for 3D volumes).
1393. _C.AUGMENTOR.ZOOM_IN_Z = False
1394. # RandomResizedCrop-style augmentation (2D only): resize the whole image/mask so the usual
1395. # fixed-size crop covers a random area fraction of the original, instead of a fixed pixel
1396. # window. Rolled against AUGMENTOR.RANDOM_RESIZED_CROP_PROB.
1397. _C.AUGMENTOR.RANDOM_RESIZED_CROP = False
1398. # Area-fraction range of the original image the crop should cover, e.g. (0.7, 0.95).
1399. _C.AUGMENTOR.RANDOM_RESIZED_CROP_SCALE_RANGE = (0.7, 0.95)
1400. # Apply shift
1401. _C.AUGMENTOR.SHIFT = False
1402. # Shift range. Translation as a fraction of the image height/width (x-translation, y-translation), where 0 denotes
1403. # “no change” and 0.5 denotes “half of the axis size”.
1404. _C.AUGMENTOR.SHIFT_RANGE = (0.1, 0.2)
1405. # How to fill up the new values created with affine transformations (rotations, shear, shift and zoom).
1406. # Only keep modes common to skimage & scipy: 'constant', 'reflect', 'wrap' and 'symmetric
1407. # Dropped 'edge'/'nearest' for simplicity
1408. _C.AUGMENTOR.AFFINE_MODE = "reflect"
1409. # Make vertical flips
1410. _C.AUGMENTOR.VFLIP = False
1411. # Make horizontal flips
1412. _C.AUGMENTOR.HFLIP = False
1413. # Make z-axis flips
1414. _C.AUGMENTOR.ZFLIP = False
1415. # Elastic transformations
1416. _C.AUGMENTOR.ELASTIC = False
1417. # Strength of the distortion field. Higher values mean that pixels are moved further with respect to the distortion
1418. # field's direction. Set this to around 10 times the value of sigma for visible effects.
1419. _C.AUGMENTOR.E_ALPHA = (12, 16)
1420. # Standard deviation of the gaussian kernel used to smooth the distortion fields. Higher values (for 128x128 images
1421. # around 5.0) lead to more water-like effects, while lower values (for 128x128 images around 1.0 and lower) lead to
1422. # more noisy, pixelated images. Set this to around 1/10th of alpha for visible effects.
1423. _C.AUGMENTOR.E_SIGMA = 4
1424. # Parameter that defines the handling of newly created pixels with the elastic transformation
1425. _C.AUGMENTOR.E_MODE = "constant"
1426. # Gaussian blur
1427. _C.AUGMENTOR.G_BLUR = False
1428. # Standard deviation of the gaussian kernel. Values in the range 0.0 (no blur) to 3.0 (strong blur) are common.
1429. _C.AUGMENTOR.G_SIGMA = (1.0, 2.0)
1430. # To blur an image by computing median values over neighbourhoods
1431. _C.AUGMENTOR.MEDIAN_BLUR = False
1432. # Median blur kernel size
1433. _C.AUGMENTOR.MB_KERNEL = (3, 7)
1434. # Blur images in a way that fakes camera or object movements
1435. _C.AUGMENTOR.MOTION_BLUR = False
1436. # Kernel size to use in motion blur
1437. _C.AUGMENTOR.MOTB_K_RANGE = (8, 12)
1438. # Gamma contrast
1439. _C.AUGMENTOR.GAMMA_CONTRAST = False
1440. # Exponent for the contrast adjustment. Higher values darken the image
1441. _C.AUGMENTOR.GC_GAMMA = (1.25, 1.75)
1442. # To apply brightness changes to images
1443. _C.AUGMENTOR.BRIGHTNESS = False
1444. # Strength of the brightness range.
1445. _C.AUGMENTOR.BRIGHTNESS_FACTOR = (-0.1, 0.1)
1446. # To apply contrast changes to images
1447. _C.AUGMENTOR.CONTRAST = False
1448. # Strength of the contrast change range.
1449. _C.AUGMENTOR.CONTRAST_FACTOR = (-0.1, 0.1)
1450. # Set a certain fraction of pixels in images to zero (not get confused with the dropout concept of neural networks)
1451. _C.AUGMENTOR.DROPOUT = False
1452. # Range to take the probability to drop a pixel
1453. _C.AUGMENTOR.DROP_RANGE = (0, 0.2)
1454. # To fill one or more rectangular areas in an image using a fill mode
1455. _C.AUGMENTOR.CUTOUT = False
1456. # Range of number of areas to fill the image with. Reasonable values between range [0,4]
1457. _C.AUGMENTOR.COUT_NB_ITERATIONS = (1, 3)
1458. # Size of the areas in % of the corresponding image size
1459. _C.AUGMENTOR.COUT_SIZE = (0.05, 0.3)
1460. # Value to fill the area of cutout
1461. _C.AUGMENTOR.COUT_CVAL = 0.0
1462. # Apply cutout to the segmentation mask
1463. _C.AUGMENTOR.COUT_APPLY_TO_MASK = False
1464. # To apply cutblur operation
1465. _C.AUGMENTOR.CUTBLUR = False
1466. # Size of the region to apply cutblur
1467. _C.AUGMENTOR.CBLUR_SIZE = (0.2, 0.4)
1468. # Range of the downsampling to be made in cutblur
1469. _C.AUGMENTOR.CBLUR_DOWN_RANGE = (2, 8)
1470. # Whether to apply cut-and-paste just LR into HR image. If False, HR to LR will be applied also (see Figure 1
1471. # of the paper https://arxiv.org/pdf/2004.00448.pdf)
1472. _C.AUGMENTOR.CBLUR_INSIDE = True
1473. # Apply cutmix operation
1474. _C.AUGMENTOR.CUTMIX = False
1475. # Size of the region to apply cutmix
1476. _C.AUGMENTOR.CMIX_SIZE = (0.2, 0.4)
1477. # Apply noise to a region of the image
1478. _C.AUGMENTOR.CUTNOISE = False
1479. # Range to choose a value that will represent the % of the maximum value of the image that will be used as the std
1480. # of the Gaussian Noise distribution
1481. _C.AUGMENTOR.CNOISE_SCALE = (0.05, 0.1)
1482. # Number of areas to fill with noise
1483. _C.AUGMENTOR.CNOISE_NB_ITERATIONS = (1, 3)
1484. # Size of the regions
1485. _C.AUGMENTOR.CNOISE_SIZE = (0.2, 0.4)
1486. # Add miss-aligment augmentation
1487. _C.AUGMENTOR.MISALIGNMENT = False
1488. # Maximum pixel displacement in 'xy'-plane for misalignment
1489. _C.AUGMENTOR.MS_DISPLACEMENT = 16
1490. # Ratio of rotation-based mis-alignment
1491. _C.AUGMENTOR.MS_ROTATE_RATIO = 0.5
1492. # Augment the image by creating a black line in a random position
1493. _C.AUGMENTOR.MISSING_SECTIONS = False
1494. # Probability of adding a missing section in each channel
1495. _C.AUGMENTOR.MISSP_CHANNEL_PB = 0.5
1496. # Iterations to dilate the missing line with
1497. _C.AUGMENTOR.MISSP_ITERATIONS = (10, 30)
1498. # Convert images in grasycale gradually based on '_C.AUGMENTOR.GRAY_RANGE'
1499. _C.AUGMENTOR.GRAYSCALE = False
1500. # Shuffle channels of the images
1501. _C.AUGMENTOR.CHANNEL_SHUFFLE = False
1502. # Apply gridmask to the image. Original paper: https://arxiv.org/pdf/2001.04086v1.pdf
1503. _C.AUGMENTOR.GRIDMASK = False
1504. # Determines the keep ratio of an input image
1505. _C.AUGMENTOR.GRID_RATIO = 0.6
1506. # Range to choose a d value
1507. _C.AUGMENTOR.GRID_D_RANGE = (0.4, 1)
1508. # Rotation of the mask in GridMask. Needs to be between [0,1] where 1 is 360 degrees.
1509. _C.AUGMENTOR.GRID_ROTATE = 1.0
1510. # Whether to invert the mask
1511. _C.AUGMENTOR.GRID_INVERT = False
1512. # Add Gaussian noise
1513. _C.AUGMENTOR.GAUSSIAN_NOISE = False
1514. _C.AUGMENTOR.GAUSSIAN_NOISE_MEAN = 0.0
1515. _C.AUGMENTOR.GAUSSIAN_NOISE_VAR = 0.05
1516. _C.AUGMENTOR.GAUSSIAN_NOISE_USE_INPUT_IMG_MEAN_AND_VAR = False
1517. # Add Poisson noise
1518. _C.AUGMENTOR.POISSON_NOISE = False
1519. # Add salt (replaces random pixels with 1)
1520. _C.AUGMENTOR.SALT = False
1521. _C.AUGMENTOR.SALT_AMOUNT = 0.05
1522. # Add pepper (replaces random pixels with 0 (for unsigned images) or -1 (for signed images))
1523. _C.AUGMENTOR.PEPPER = False
1524. _C.AUGMENTOR.PEPPER_AMOUNT = 0.05
1525. # Whether to add Poisson noise
1526. _C.AUGMENTOR.SALT_AND_PEPPER = False
1527. _C.AUGMENTOR.SALT_AND_PEPPER_AMOUNT = 0.05
1528. _C.AUGMENTOR.SALT_AND_PEPPER_PROP = 0.5
1530. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1531. # 5. Model definition
1532. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1533. _C.MODEL = CN()
1534. # Whether to define manually the model ('biapy'), load a pretrained one from BioImage Model Zoo ('bmz') or use one
1535. # available in TorchVision ('torchvision').
1536. # Options: ["biapy", "bmz", "torchvision"]
1537. _C.MODEL.SOURCE = "biapy"
1539. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1540. # 5.1 BiaPy backend models options
1541. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1542. # Architecture of the network. Possible values are:
1543. # * Semantic segmentation: 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unetr', 'unext_v1', 'unext_v2',
1544. # 'hrnet' and 'stunet'
1545. # * Instance segmentation: 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unetr', 'unext_v1', 'unext_v2',
1546. # 'hrnet' and 'stunet'
1547. # * Detection: 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unetr', 'unext_v1', 'unext_v2', 'hrnet' and
1548. # 'stunet'
1549. # * Denoising: 'unet', 'resunet', 'resunet++', 'attention_unet', 'seunet', 'resunet_se', 'unext_v1', 'unext_v2', 'hrnet' and 'stunet'
1550. # * Super-resolution: 'edsr', 'rcan', 'dfcan', 'wdsr', 'unet', 'resunet', 'resunet++', 'seunet', 'resunet_se', 'attention_unet', 'multiresunet', 'unext_v1'
1551. # and 'unext_v2'
1552. # * Self-supervision: 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unetr', 'edsr', 'rcan', 'dfcan', 'wdsr', 'vit',
1553. # 'mae', 'unext_v1', 'unext_v2', 'hrnet' and 'stunet'
1554. # * Classification: 'simple_cnn', 'vit' and 'efficientnet_b[0-7]' (only 2D)
1555. # * Image to image: 'edsr', 'rcan', 'dfcan', 'wdsr', 'unet', 'resunet', 'resunet++', 'seunet', 'resunet_se', 'attention_unet', 'unetr', 'multiresunet', 'unext_v1',
1556. # 'unext_v2', 'hrnet' and 'stunet'
1557. _C.MODEL.ARCHITECTURE = "unet"
1559. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1560. # 5.1.1 U-Net-like architectures options
1561. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1562. # Number of feature maps on each level of the network. Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se',
1563. # 'unext_v1', 'unext_v2' architectures.
1564. _C.MODEL.FEATURE_MAPS = [16, 32, 64, 128, 256]
1565. # Values to make the dropout with. Set to 0 to prevent dropout. When using it with 'ViT' or 'unetr' a list with just one number must be provided. Works with
1566. # 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se' and 'unetr' architectures.
1567. _C.MODEL.DROPOUT_VALUES = [0.0, 0.0, 0.0, 0.0, 0.0]
1568. # Normalization layer (one of 'bn', 'sync_bn' 'in', 'ln', 'gn' or 'none'). Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet',
1569. # 'resunet_se', 'unetr', 'hrnet' architectures.
1570. _C.MODEL.NORMALIZATION = "in"
1571. # to set the kernel size for the convolutional layers. Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se'
1572. # and 'unetr' architectures.
1573. _C.MODEL.KERNEL_SIZE = 3
1574. # Upsampling layer to use in the model. Options: ["upsampling", "convtranspose"]. Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet',
1575. # 'seunet' and 'resunet_se' architectures.
1576. _C.MODEL.UPSAMPLE_LAYER = "convtranspose"
1577. # Activation function to use along the model (not in the final layer). Options: 'relu', 'tanh', 'leaky_relu', 'elu', 'gelu', 'silu', 'sigmoid', 'softmax', 'linear',
1578. # 'softplus' and 'none'. Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unetr' and 'hrnet' architectures.
1579. _C.MODEL.ACTIVATION = "elu"
1580. # Downsampling to be made in Z. This value will be the third integer of the MaxPooling operation. When facing anysotropic datasets set it to get better performance.
1581. # Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unext_v1' and 'unext_v2' architectures.
1582. _C.MODEL.Z_DOWN = [0, 0, 0, 0]
1583. # Downsampling to be made in XY. This value will be the first and second integer of the MaxPooling operation. When facing anysotropic datasets set it to get better
1584. # performance. Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unext_v1' and 'unext_v2' architectures.
1585. _C.MODEL.YX_DOWN = [0, 0, 0, 0]
1586. # For each level of the model (U-Net levels), set to true or false if the dimensions of the feature maps are isotropic. Works with 'unet', 'resunet', 'resunet++',
1587. # 'attention_unet', 'multiresunet', 'seunet', 'resunet_se', 'unext_v1' and 'unext_v2' architectures.
1588. _C.MODEL.ISOTROPY = [True, True, True, True, True]
1589. # Include extra convolutional layers with larger kernel at the beginning and end of the U-Net-like model. Works with 'unet', 'resunet', 'resunet++', 'attention_unet',
1590. # 'multiresunet', 'seunet' and 'resunet_se' architectures.
1591. _C.MODEL.LARGER_IO = False
1592. # Number of convolutional layers to stack at each level of the U-Net-like model, given as one value per level (i.e. per feature map). For 'unet', 'resunet', 'resunet++',
1593. # 'attention_unet', 'seunet' and 'resunet_se' architectures these are plain/residual convolutions ([2, 2, ...] reproduces the classic double-convolution U-Net); for
1594. # 'unext_v1' and 'unext_v2' these are the number of ConvNeXtBlocks in each level. If a single value is provided it is broadcast to all levels.
1595. _C.MODEL.CONV_LAYERS = [2, 2, 2, 2, 2] # CONV_LAYERS
1596. # Ordering of the convolution, normalization and activation layers inside each conv block.
1597. # Options: "conv_norm_act" (default, post-activation: Conv -> Norm -> Act, the historical
1598. # BiaPy block, keeps existing checkpoints loadable) and "norm_act_conv" (pre-activation:
1599. # Norm -> Act -> Conv, normalizes every convolution's input including the first;
1600. # more stable when training from scratch at large learning rates).
1601. # Works with 'unet', 'resunet', 'resunet++', 'attention_unet', 'seunet' and 'resunet_se'
1602. # architectures ('multiresunet' only supports "conv_norm_act").
1603. _C.MODEL.CONV_BLOCK_ORDER = "conv_norm_act"
1605. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1606. # 5.1.1.1 U-NeXT (v1 and v2) architectures options
1607. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1608. # "unext_v1" and "unext_v2" architectures variables. These architectures are based on the ConvNeXt architecture (https://arxiv.org/abs/2201.03545) and they
1609. # are adapted to the U-Net structure.
1610. # Maximum Stochastic Depth probability for the U-NeXt model.
1611. _C.MODEL.CONVNEXT_SD_PROB = 0.1
1612. # Layer Scale parameter for the U-NeXt model. Only valid for the "unext_v1" architecture.
1613. _C.MODEL.CONVNEXT_LAYER_SCALE = 1e-6
1614. # Size of the stem kernel in the U-NeXt model.
1615. _C.MODEL.CONVNEXT_STEM_K_SIZE = 2
1617. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1618. # 5.1.1.2 UNETR architecture options
1619. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1620. # Type of ViT model to use as UNETR's backbone. Options are "custom", "vit_base_patch16", "vit_large_patch16",
1621. # "vit_huge_patch14" and "sam3_vit". On "custom" setting the backbone is built with the 'MODEL.VIT_*' variables,
1622. # whereas with the rest of the options all of them ('MODEL.VIT_TOKEN_SIZE' included) are set automatically.
1623. # Notice that UNETR's decoder upsamples the ViT features by a factor of two on each of its levels, so the
1624. # resulting token size must be a power of two: "vit_huge_patch14" can not be used and "custom" must be selected
1625. # instead. "sam3_vit" builds the image encoder of SAM 3, which can be initialized with its pretrained weights
1626. # through 'MODEL.VIT_PRETRAINED_WEIGHTS'. As SAM 3's 14x14 tokens are not a power of two, 16x16 ones are used
1627. # here and its patch embedding is resized to them. It is 2D only, as SAM 3's pretrained weights are 2D.
1628. _C.MODEL.UNETR_VIT_MODEL = "custom"
1629. # Multiple of the transformer encoder layers from of which the skip connection signal is going to be extracted.
1630. # Leave it as -1 to decide it automatically based on the encoder selected, which spaces the skip connections
1631. # evenly along it taking one every 'MODEL.VIT_NUM_LAYERS // log2(token size)' blocks. That gives the 3 that
1632. # used to be the default with the classic 12-layer ViT and 16x16 tokens, and adapts it to deeper encoders,
1633. # e.g. it gives 8 with the 32 blocks of "sam3_vit" (so blocks 8, 16 and 24 feed the decoder).
1634. _C.MODEL.UNETR_VIT_HIDD_MULT = -1
1635. # Number of filters in the first UNETR's layer of the decoder. In each layer the previous number of filters is doubled.
1636. _C.MODEL.UNETR_VIT_NUM_FILTERS = 16
1638. # Specific for SR models based on U-Net architectures. Options are ["pre", "post"]
1639. _C.MODEL.UNET_SR_UPSAMPLE_POSITION = "pre"
1641. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1642. # 5.1.2 Transformer-based architectures options
1643. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1644. # Type of ViT model. Options are "custom", "vit_base_patch16", "vit_large_patch16", "vit_huge_patch14" and
1645. # "sam3_vit". On "custom" setting the rest of the ViT parameters can be modified as other options will set
1646. # them automatically. "sam3_vit" builds the image encoder of SAM 3 (Segment Anything Model 3), which can be
1647. # initialized with its pretrained weights through 'MODEL.VIT_PRETRAINED_WEIGHTS'. It uses SAM 3's 14x14
1648. # tokens, so 'DATA.PATCH_SIZE' must be a multiple of 14, and it is 2D only.
1649. _C.MODEL.VIT_MODEL = "custom"
1650. # Pretrained weights to initialize the ViT backbone with. Leave it empty to train from scratch. It can be a
1651. # Hugging Face repository, i.e. "facebook/sam3" or "facebook/sam3.1" (both share the same image encoder, so
1652. # "sam3_vit" builds the backbone for either of them), or the path to a local file with the weights. Only used when
1653. # the selected ViT is "sam3_vit" ('MODEL.VIT_MODEL' or 'MODEL.UNETR_VIT_MODEL'). Notice that SAM 3 is a gated
1654. # model, so its license needs to be accepted in https://huggingface.co/facebook/sam3 and this machine needs
1655. # to be authenticated (running "hf auth login" or exporting the HF_TOKEN environment variable) to download it.
1656. # The weights are not downloaded when 'MODEL.LOAD_CHECKPOINT' is enabled, as the checkpoint replaces them.
1657. _C.MODEL.VIT_PRETRAINED_WEIGHTS = ""
1658. # Size of the patches (tokens) that are extracted from the input image. Only used when the ViT model selected
1659. # is "custom", as the rest of them are built with the token size they were designed with (e.g. "sam3_vit"
1660. # uses SAM 3's 14x14 tokens, or the closest power of two under UNETR, as its decoder needs it).
1661. _C.MODEL.VIT_TOKEN_SIZE = 16
1662. # Dimension of the embedding space
1663. _C.MODEL.VIT_EMBED_DIM = 768
1664. # Number of transformer encoder layers
1665. _C.MODEL.VIT_NUM_LAYERS = 12
1666. # Number of heads in the multi-head attention layer.
1667. _C.MODEL.VIT_NUM_HEADS = 12
1668. # Size of the dense layers of the final classifier. This value will mutiply 'VIT_EMBED_DIM'
1669. _C.MODEL.VIT_MLP_RATIO = 4.0
1670. # Normalization layer epsion
1671. _C.MODEL.VIT_NORM_EPS = 1e-6
1673. # ViT architecture adapted for self-supervised learning with masked autoencoders (MAE). Original paper: https://arxiv.org/abs/2111.06377
1674. # Dimension of the embedding space for the MAE decoder
1675. _C.MODEL.MAE_DEC_HIDDEN_SIZE = 512
1676. # Number of transformer decoder layers
1677. _C.MODEL.MAE_DEC_NUM_LAYERS = 8
1678. # Number of heads in the multi-head attention layer.
1679. _C.MODEL.MAE_DEC_NUM_HEADS = 16
1680. # Size of the dense layers of the final classifier
1681. _C.MODEL.MAE_DEC_MLP_DIMS = 2048
1682. # Type of the masking strategy. Options: ["grid", "random"]
1683. _C.MODEL.MAE_MASK_TYPE = "grid"
1684. # Percentage of the input image to mask (applied only when MODEL.MAE_MASK_TYPE == "random"). Value between 0 and 1.
1685. _C.MODEL.MAE_MASK_RATIO = 0.5
1687. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1688. # 5.1.3 RCAN architecture options
1689. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1690. # Number of RG modules
1691. _C.MODEL.RCAN_RG_BLOCK_NUM = 10
1692. # Number of RCAB modules in each RG block
1693. _C.MODEL.RCAN_RCAB_BLOCK_NUM = 20
1694. # Filters in the convolutions
1695. _C.MODEL.RCAN_CONV_FILTERS = 16
1696. # Channel reduction ratio for channel attention
1697. _C.MODEL.RCAN_REDUCTION_RATIO = 16
1698. # Whether to maintain or not the upscaling layer.
1699. _C.MODEL.RCAN_UPSCALING_LAYER = True
1701. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1702. # 5.1.4 HRNet architecture options
1703. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1704. # These parameters can be used as a template for building custom HRNet versions
1705. _C.MODEL.HRNET = CN()
1706. # Whether to use a custom configuration for HRNet or use a predefined one. Options: "W18" (hrnet18), "W32" (hrnet32),
1707. # "W48" (hrnet48), "W64" (hrnet64) or "custom"
1708. _C.MODEL.HRNET.VARIANT = "W48"
1709. # Whether to downsample the input in Z or not
1710. _C.MODEL.HRNET.Z_DOWN = [0, 0, 0]
1711. # Downsampling to be made in XY. This value will be the first and second integer of the MaxPooling operation.
1712. # When facing anysotropic datasets set it to get better performance
1713. _C.MODEL.HRNET.YX_DOWN = [0, 0, 0]
1714. # Type of block to use in HRNet. Options: 'BASIC', 'BOTTLENECK', 'CONVNEXT_V1' and 'CONVNEXT_V2'
1715. _C.MODEL.HRNET.BLOCK_TYPE = 'BASIC'
1716. # Indicate whether to use a custom configuration for HRNet or use a predefined one. Options: "OCR", "ASPP", "PSP", "FCN"
1717. _C.MODEL.HRNET.HEAD_TYPE = "FCN"
1718. # Number of stages in the HRNet. This value will determine the length of the rest of the lists. Only used if MODEL.HRNET.VARIANT = "custom"
1719. _C.MODEL.HRNET.NUM_STAGES = 3
1720. # Number of modules in each stage. Only used if MODEL.HRNET.VARIANT = "custom". A module is a sequence of blocks (see MODEL.HRNET.BLOCK_TYPE)
1721. # that are not connected with the rest of branches. In each stage, after the modules, a fusion is made between all the branches. So, the
1722. # number of modules will determine how many times the fusion is made in each stage.
1723. _C.MODEL.HRNET.NUM_MODULES = [1, 4, 3]
1724. # Number of branches in each stage. Only used if MODEL.HRNET.VARIANT = "custom". The number of branches will determine how many parallel convolutions
1725. # are made in each stage and how many feature maps with different resolutions are generated.
1726. _C.MODEL.HRNET.NUM_BRANCHES = [2, 3, 4]
1727. # Number of blocks in each branch of each stage. Only used if MODEL.HRNET.VARIANT = "custom". A block is a convolutional operation
1728. # (see MODEL.HRNET.BLOCK_TYPE) that is repeated a certain number of times in each branch. The number of blocks will determine the
1729. # depth of the model.
1730. _C.MODEL.HRNET.NUM_BLOCKS = [[4, 4], [4, 4, 4], [4, 4, 4, 4]]
1731. # Number of channels in each block of each branch of each stage. Only used if MODEL.HRNET.VARIANT = "custom". The number of channels
1732. # will determine the width of the model.
1733. _C.MODEL.HRNET.NUM_CHANNELS = [[18, 36], [18, 36, 72], [18, 36, 72, 144]]
1735. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1736. # 5.1.5 STUNet architecture options
1737. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1738. _C.MODEL.STUNET = CN()
1739. # Variant of the STUNet model. Options are: 'small', 'base', 'large', 'custom'
1740. _C.MODEL.STUNET.VARIANT = 'base'
1741. # Whether to use a pretrained version of STUNet on ImageNet
1742. _C.MODEL.STUNET.PRETRAINED = False
1743. # Residual blocks per stage. Only used when VARIANT = "custom"; length must equal len(DIMS).
1744. _C.MODEL.STUNET.DEPTH = [1, 1, 1, 1, 1, 1]
1745. # Channels per stage. Only used when VARIANT = "custom".
1746. _C.MODEL.STUNET.DIMS = [32, 64, 128, 256, 512, 512]
1747. # Per-stage pooling/stride, (Z,Y,X), one entry per downsampling step (len(DIMS)-1). Only used
1748. # when VARIANT = "custom"; set anisotropic values (e.g. [1,2,2]) to avoid downsampling Z.
1749. _C.MODEL.STUNET.POOL_OP_KERNEL_SIZES = [[2, 2, 2], [2, 2, 2], [2, 2, 2], [2, 2, 2], [1, 1, 1]]
1750. # Per-stage conv kernel size, (Z,Y,X), one entry per stage (len(DIMS)). Only used when VARIANT
1751. # = "custom"; set anisotropic values (e.g. [1,3,3]) to avoid mixing information across Z.
1752. _C.MODEL.STUNET.CONV_KERNEL_SIZES = [[3, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3], [3, 3, 3]]
1753. # Indices (into DATA.PATCH_SIZE's channel axis) of an "exchangeable" group of input channels
1754. # whose channel-to-semantic mapping isn't stable (e.g. unsupervised per-slice GMM cluster
1755. # ids that don't track the same physical structure from slice to slice). When non-empty,
1756. # these channels are routed through a permutation-invariant encoder (shared per-channel 1x1
1757. # conv + pooling, see biapy.models.blocks.PermInvariantChannelSetEncoder) instead of the
1758. # network's regular first layer, so the result is exactly invariant to their order -- unlike
1759. # unioning them into a single channel, real per-channel information survives. The remaining
1760. # channels are passed through unchanged. Only used when VARIANT = "custom"; requires
1761. # CLASS_SET_OUT_CHANNELS > 0. Left empty (default), this is a no-op.
1762. _C.MODEL.STUNET.CLASS_SET_CHANNEL_IDXS = []
1763. # Output width of the CLASS_SET_CHANNEL_IDXS group's pooled encoding. Only used when
1764. # CLASS_SET_CHANNEL_IDXS is non-empty.
1765. _C.MODEL.STUNET.CLASS_SET_OUT_CHANNELS = 8
1766. # Pooling across the CLASS_SET_CHANNEL_IDXS group: 'max' or 'mean'. Only used when
1767. # CLASS_SET_CHANNEL_IDXS is non-empty.
1768. _C.MODEL.STUNET.CLASS_SET_POOLING = 'max'
1770. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1771. # 5.1.6 NafNet architecture options
1772. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1773. _C.MODEL.NAFNET = CN()
1774. # Number of NAFBlocks stacked at the bottleneck (deepest level).
1775. _C.MODEL.NAFNET.MIDDLE_BLK_NUM = 12
1776. # Number of NAFBlocks assigned to each downsampling level of the encoder.
1777. _C.MODEL.NAFNET.ENC_BLK_NUMS = [2, 2, 4, 8]
1778. # Number of NAFBlocks assigned to each upsampling level of the decoder.
1779. _C.MODEL.NAFNET.DEC_BLK_NUMS = [2, 2, 2, 2]
1780. # Channel expansion factor for the depthwise convolution within the gating unit.
1781. _C.MODEL.NAFNET.DW_EXPAND = 2
1782. # Expansion factor for the hidden layer within the feed-forward network.
1783. _C.MODEL.NAFNET.FFN_EXPAND = 2
1784. # Discriminator architecture
1785. _C.MODEL.NAFNET.ARCHITECTURE_D = "patchgan"
1786. # Discriminator PATCHGAN
1787. _C.MODEL.NAFNET.PATCHGAN = CN()
1788. # Number of initial convolutional filters in the first layer of the discriminator.
1789. _C.MODEL.NAFNET.PATCHGAN.BASE_FILTERS = 64
1790. # Generator backbone used inside the NAFNet GAN pipeline. 'nafnet' (default) uses NAFNet's own
1791. # NAFBlock encoder/decoder as the generator. 'stunet' swaps in STUNet (configured via
1792. # MODEL.STUNET) as the generator instead.
1793. _C.MODEL.NAFNET.GENERATOR_BACKBONE = "nafnet"
1795. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1796. # 5.1.7 RDBM (Residual Diffusion Bridge Model) architecture options
1797. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1798. _C.MODEL.RDBM = CN()
1799. # Base channel width of the conditional U-Net used as the bridge's denoising network.
1800. _C.MODEL.RDBM.BASE_DIM = 64
1801. # Channel multiplier per resolution level of the U-Net (encoder/decoder depth = len(DIM_MULTS)).
1802. _C.MODEL.RDBM.DIM_MULTS = [1, 2, 4, 8]
1803. # Number of discretization steps of the forward bridge process used during training.
1804. _C.MODEL.RDBM.TIMESTEPS = 100
1805. # Number of reverse (sampling) steps at validation/test time. Must be <= TIMESTEPS.
1806. _C.MODEL.RDBM.SAMPLING_TIMESTEPS = 10
1807. # Noise scale of the bridge's stochastic term.
1808. _C.MODEL.RDBM.LAMB = 1.0e-4
1810. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1811. # 5.1.8 Checkpoint options
1812. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1813. # To load a model (and more items if available) from a given checkpoint. Items that can be loaded are defined in 'MODEL.ITEMS_TO_LOAD_FROM_CHECKPOINT'.
1814. _C.MODEL.LOAD_CHECKPOINT = False
1815. # List of items to load from the checkpoint (if available). Options are:
1816. # * "weights": to load the model weights
1817. # * "norm": to load the normalization used. Unless you know that the checkpoint was trained with the same normalization as the one defined in the config,
1818. # it is recommended to load it from the checkpoint if "weights" is also specified, as the model will expect the input data to be normalized accordingly.
1819. # * "model_arch": to load the model architecture
1820. # * "optimizer": to load the optimizer state dict
1821. # * "epoch": to load the epoch number from which the training will be resumed
1822. #
1823. # Defining it with "weights" and "model_arch" will allow to load the model weights and architecture, but not the
1824. # optimizer state, which is useful when doing inference or fine-tunning. Defining it with all options
1825. # will allow to resume training from a checkpoint.
1826. _C.MODEL.ITEMS_TO_LOAD_FROM_CHECKPOINT = ["weights", "norm", "model_arch"]
1827. # Decide which checkpoint to load from job's dir if PATHS.CHECKPOINT_FILE is ''.
1828. # Options: 'best_on_val' or 'last_on_train'
1829. _C.MODEL.LOAD_CHECKPOINT_EPOCH = "best_on_val"
1830. # Format of the output checkpoint. Options are 'pth' (native PyTorch format) or 'safetensors' (https://github.com/huggingface/safetensors)
1831. _C.MODEL.OUT_CHECKPOINT_FORMAT = "pth"
1832. # To skip loading those layers that do not match in shape with the given checkpoint. If this is set to False a regular load function will be
1833. # done, which will fail if a layer mismatch is found. Only applicable if "weights" is in 'MODEL.ITEMS_TO_LOAD_FROM_CHECKPOINT'. It is useful
1834. # to set it to True when fine-tunning a model with a different head.
1835. _C.MODEL.SKIP_UNMATCHED_LAYERS = False
1836. # Epochs to save a checkpoint of the model apart from the best of the validation. Set it to -1 to not do it.
1837. _C.MODEL.SAVE_CKPT_FREQ = -1
1838. # List of regex patterns to match parameter names that should be frozen (i.e. excluded from gradient updates) during
1839. # training. Each entry is a Python 're' pattern tested against the full parameter name (e.g. "encoder\.layer1\..*").
1840. # Freezing is applied after loading the checkpoint so that the loaded weights are preserved. Parameters matched by
1841. # any pattern will have 'requires_grad' set to False and will be excluded from the optimizer. An empty list (default)
1842. # means no layers are frozen.
1843. # Examples: ["backbone\.layer1\..*", "backbone\.layer2\.conv.*"]
1844. _C.MODEL.FREEZE_LAYERS_MATCHING = []
1846. # Passed as 'find_unused_parameters' to DistributedDataParallel. Set to True if the model has
1847. # parameters that don't participate in every forward pass (e.g. some BMZ models), otherwise DDP
1848. # raises "Expected to have finished reduction...". Costs a small per-iteration overhead.
1849. _C.MODEL.FIND_UNUSED_PARAMETERS = False
1851. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1852. # 5.2 BioImage Model Zoo (BMZ) options
1853. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1854. # BMZ model export options
1855. _C.MODEL.BMZ = CN()
1856. # DOI or nickname of the model from BMZ to load. It can not be empty if MODEL.SOURCE = "bmz".
1857. _C.MODEL.BMZ.SOURCE_MODEL_ID = ""
1858. # Module to export the model to BMZ format. It will be activated when 'MODEL.BMZ.EXPORT.ENABLE' is set to True.
1859. _C.MODEL.BMZ.EXPORT = CN()
1860. # Whether to activate or not the exporation of the used model to the BMZ format after train and/or test
1861. _C.MODEL.BMZ.EXPORT.ENABLE = False
1862. # Name of the model to create. It should be something meaningful. Take other models in https://bioimage.io/#/ as reference.
1863. _C.MODEL.BMZ.EXPORT.MODEL_NAME = ""
1864. # Description of the model. It should be something meaningful. Take other models in https://bioimage.io/#/ as reference.
1865. # E.g. "Mitochondria segmentation for electron microscopy"
1866. _C.MODEL.BMZ.EXPORT.DESCRIPTION = ""
1867. # List of authors of the model. Each item must be a dict containing "name" and "githubuser".
1868. # E.g. [{"name": "Daniel", "github_user": "danifranco"}]
1869. _C.MODEL.BMZ.EXPORT.AUTHORS = []
1870. # License of the model.
1871. _C.MODEL.BMZ.EXPORT.LICENSE = "CC-BY-4.0"
1872. # Path to a .md extension file with the documentation of the model. If it is not set so the model documentation will be
1873. # automatically generated with the information provided. If you want to add more information you can use this option.
1874. # E.g. to provide more details of the training procedure, data, etc.
1875. _C.MODEL.BMZ.EXPORT.DOCUMENTATION = ""
1876. # List of tags. Here the type of dataset and the target object should be provided. BiaPy automatically sets the following tags:
1877. # * "biapy": to represent that the model was created with BiaPy.
1878. # * "pytorch": to represent that you are using Pytorch
1879. # * "2d" or "3d": depending on the image dimensions one or the other is selected.
1880. # * workflow tag: depending on the workflow the tag is set. E.g. "semantic-segmentation"
1881. #
1882. # So, what you can set for instance is: ["electron-microscopy", "mitochondria"]
1883. _C.MODEL.BMZ.EXPORT.TAGS = []
1884. # Citations. It must be a list of dictionaries with keys "text" and "doi". E.g.:
1885. # [{"text": "training library", "doi": "10.1038/s41592-025-02699-y"}, {"text": "architecture", "doi": "10.1109/LGRS.2018.2802944"},
1886. # {"text": "data", "doi": "10.48550/arXiv.1812.06024"}]
1887. _C.MODEL.BMZ.EXPORT.CITE = []
1888. # Information of the dataset. It must be a list of just one dict item with keys "name", "doi" and "image_modality". It can have
1889. # also "dataset_id" key but it is optional (must match a dataset in the BioImage Model Zoo). E.g:
1890. # [{
1891. # "name": "CartoCell",
1892. # "doi": "10.1016/j.crmeth.2023.100597",
1893. # "image_modality": "fluorescence microscopy",
1894. # "dataset_id": "biapy/cartocell_cyst_segmentation",
1895. # }]
1896. _C.MODEL.BMZ.EXPORT.DATASET_INFO = [{}]
1897. # Version of the model
1898. _C.MODEL.BMZ.EXPORT.MODEL_VERSION = "0.1.0"
1899. # If you are loading a BMZ model you can enable this option to avoid setting all above variables and instead reuse the same
1900. # information that was present in that model. You need still to set 'MODEL.BMZ.EXPORT.ENABLE' to 'True' and nothing else.
1901. _C.MODEL.BMZ.EXPORT.REUSE_BMZ_CONFIG = False
1903. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1904. # 5.2 TorchVision options (limited support)
1905. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1906. # BiaPy support using models of Torchvision. However, most of the models were trained in natural images (not biomedical) and most of them
1907. # are for classification. On top of that, some use bounding-box-annotations, which are not supported in BiaPy so only inference/prediction/test
1908. # can only be done. 'MODEL.TORCHVISION_MODEL_NAME' variable can not be empty if MODEL.SOURCE = "torchvision".
1909. # Models available here: https://pytorch.org/vision/stable/models.html
1910. # They can be listed with: "from torchvision.models import list_models; list_models()"
1911. #
1912. # Semantic segmentation (https://pytorch.org/vision/stable/models.html#semantic-segmentation):
1913. # 'deeplabv3_mobilenet_v3_large', 'deeplabv3_resnet101', 'deeplabv3_resnet50', 'fcn_resnet101', 'fcn_resnet50',
1914. # 'lraspp_mobilenet_v3_large'
1915. #
1916. # Object Detection (https://pytorch.org/vision/stable/models.html#object-detection-instance-segmentation-and-person-keypoint-detection)
1917. # 'fasterrcnn_mobilenet_v3_large_320_fpn', 'fasterrcnn_mobilenet_v3_large_fpn', 'fasterrcnn_resnet50_fpn',
1918. # 'fasterrcnn_resnet50_fpn_v2', 'fcos_resnet50_fpn', 'ssd300_vgg16', 'ssdlite320_mobilenet_v3_large',
1919. # 'retinanet_resnet50_fpn', 'retinanet_resnet50_fpn_v2',
1920. #
1921. # Instance Segmentation (https://pytorch.org/vision/stable/models.html#object-detection-instance-segmentation-and-person-keypoint-detection)
1922. # 'maskrcnn_resnet50_fpn', 'maskrcnn_resnet50_fpn_v2'
1923. #
1924. # Image classification (https://pytorch.org/vision/stable/models.html#classification):
1925. # 'alexnet', 'convnext_base', 'convnext_large', 'convnext_small', 'convnext_tiny', 'densenet121', 'densenet161',
1926. # 'densenet169', 'densenet201', 'efficientnet_b0', 'efficientnet_b1', 'efficientnet_b2', 'efficientnet_b3',
1927. # 'efficientnet_b4', 'efficientnet_b5', 'efficientnet_b6', 'efficientnet_b7', 'efficientnet_v2_l', 'efficientnet_v2_m',
1928. # 'efficientnet_v2_s', 'googlenet', 'inception_v3', 'maxvit_t', 'mnasnet0_5', 'mnasnet0_75', 'mnasnet1_0', 'mnasnet1_3',
1929. # 'mobilenet_v2', 'mobilenet_v3_large', 'mobilenet_v3_small', 'quantized_googlenet', 'quantized_inception_v3',
1930. # 'quantized_mobilenet_v2', 'quantized_mobilenet_v3_large', 'quantized_resnet18', 'quantized_resnet50',
1931. # 'quantized_resnext101_32x8d', 'quantized_resnext101_64x4d', 'quantized_shufflenet_v2_x0_5', 'quantized_shufflenet_v2_x1_0',
1932. # 'quantized_shufflenet_v2_x1_5', 'quantized_shufflenet_v2_x2_0', 'regnet_x_16gf', 'regnet_x_1_6gf', 'regnet_x_32gf',
1933. # 'regnet_x_3_2gf', 'regnet_x_400mf', 'regnet_x_800mf', 'regnet_x_8gf', 'regnet_y_128gf', 'regnet_y_16gf', 'regnet_y_1_6gf',
1934. # 'regnet_y_32gf', 'regnet_y_3_2gf', 'regnet_y_400mf', 'regnet_y_800mf', 'regnet_y_8gf', 'resnet101', 'resnet152',
1935. # 'resnet18', 'resnet34', 'resnet50', 'resnext101_32x8d', 'resnext101_64x4d', 'resnext50_32x4d', 'retinanet_resnet50_fpn',
1936. # 'shufflenet_v2_x0_5', 'shufflenet_v2_x1_0', 'shufflenet_v2_x1_5', 'shufflenet_v2_x2_0',
1937. # 'squeezenet1_0', 'squeezenet1_1', 'swin_b', 'swin_s', 'swin_t', 'swin_v2_b', 'swin_v2_s', 'swin_v2_t',
1938. # 'vgg11', 'vgg11_bn', 'vgg13', 'vgg13_bn', 'vgg16', 'vgg16_bn', 'vgg19', 'vgg19_bn', 'vit_b_16', 'vit_b_32',
1939. # 'vit_h_14', 'vit_l_16', 'vit_l_32', 'wide_resnet101_2', 'wide_resnet50_2'
1940. #
1941. # Listed but not supported:
1942. #
1943. # (NOT SUPPORTED) Video classification (https://pytorch.org/vision/stable/models.html#video-classification):
1944. # 'mc3_18', 'mvit_v1_b', 'mvit_v2_s', 'r2plus1d_18', 'r3d_18','swin3d_s', 'swin3d_t', 's3d', 'swin3d_b'
1945. #
1946. # (NOT SUPPORTED) Optical flow (https://pytorch.org/vision/stable/models.html#optical-flow):
1947. # 'raft_large', 'raft_small'
1948. #
1949. # (NOT SUPPORTED) Person Keypoint Detection (https://pytorch.org/vision/stable/models.html#object-detection-instance-segmentation-and-person-keypoint-detection)
1950. # 'keypointrcnn_resnet50_fpn'
1951. #
1952. _C.MODEL.TORCHVISION_MODEL_NAME = ""
1954. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1955. # 6. Loss definition options
1956. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1957. _C.LOSS = CN()
1958. # List of individually-weighted loss names to sum:
1959. # sum(LOSS.WEIGHTS[i] * loss_i(pred, target) for i, loss_i in enumerate(LOSS.TYPE))
1960. # 'LOSS.WEIGHTS' must be the same length. Empty -> per-workflow default (see each
1961. # workflow's 'define_metrics'). Old single-string combos (e.g. "W_MAE_SSIM") still work,
1962. # auto-converted in check_configuration.py. Valid names per workflow:
1963. # * Semantic segmentation: "CE" (default), "DICE", or ["DICE", "CE"] together (fused
1964. # CE+Dice, not decomposed).
1965. # * Instance segmentation: set via PROBLEM.INSTANCE_SEG.DATA_CHANNELS(_LOSSES).
1966. # * Detection: always CE. Classification: always "CE".
1967. # * Continuous-image (DENOISING, SUPER_RESOLUTION, SELF_SUPERVISED, IMAGE_TO_IMAGE):
1968. # "MAE" (default), "MSE", "PCC", "CHARBONNIER", "VGG", "LPIPS", "SSIM", "LAPLACIAN",
1969. # "FFT", "RFFT" -- see 'continuous_image_loss_registry' in metrics.py. DENOISING only
1970. # supports a single ["MSE"] or ["BCE"]/["HINGE"], never a multi-term mix.
1971. # * IMAGE_TO_IMAGE/DENOISING also accept "BCE" or "HINGE" (adversarial term, mutually
1972. # exclusive, needs a discriminator; knobs under LOSS.GAN below).
1973. # * IMAGE_TO_IMAGE with PROBLEM.IMAGE_TO_IMAGE.MEMBRANE_REPAIR.ENABLE: "BCE", "MALIS",
1974. # "CLDICE", "SVOX" (at least one positive weight).
1975. _C.LOSS.TYPE = []
1976. # Same length as LOSS.TYPE. Unlike the old "W_*" strings, weights don't need to sum to 1.
1977. _C.LOSS.WEIGHTS = []
1978. # To weight classes in an imbalanced dataset. Options available are:
1979. # * 'none': no class rebalancing is applied
1980. # * 'manual': the weights provided in LOSS.CLASS_WEIGHTS are used to weight each class. This is valid for semantic segmentation, instance segmentation (when instance+classes are predicted)
1981. # and detection workflows (when centroids + classes are predicted).
1982. _C.LOSS.CLASS_REBALANCE = "none"
1983. # If LOSS.CLASS_REBALANCE is set to 'manual', this list of weights will be used to weight each class in the loss calculation.
1984. # The length of the list must be equal to the number of classes.
1985. _C.LOSS.CLASS_WEIGHTS = []
1986. # Whether to ignore a value in the loss and metric calculation. This is only available when LOSS.TYPE == "CE". This value will not only
1987. # be ignored in the loss computation but in the metrics, e.g. IoU. In membrane repair it marks ignored voxels in the
1988. # membrane channel: they are excluded from its normalization, fed as 0 and the affinities between two of them ignored.
1989. _C.LOSS.IGNORE_INDEX = -1
1991. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1992. # 6.1 Contrastive learning definitions
1993. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1994. _C.LOSS.CONTRAST = CN()
1995. _C.LOSS.CONTRAST.ENABLE = False
1996. _C.LOSS.CONTRAST.MEMORY_SIZE = 5000
1997. _C.LOSS.CONTRAST.PROJ_DIM = 256
1998. _C.LOSS.CONTRAST.PIXEL_UPD_FREQ = 10
2000. # Adversarial-term knobs beyond its weight/formula (those live in LOSS.TYPE/LOSS.WEIGHTS,
2001. # picking "BCE" or "HINGE"). Used only when one of those is present.
2002. _C.LOSS.GAN = CN()
2003. # R1 gradient penalty coefficient (0.0 to disable).
2004. _C.LOSS.GAN.R1_GAMMA = 0.0
2005. # If True, scale the adversarial weight by the VQGAN adaptive weight (recon/adv
2006. # gradient-norm ratio) instead of using it as a fixed scalar.
2007. _C.LOSS.GAN.ADAPTIVE_GAN_WEIGHT = False
2009. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2010. # 7. Training phase options
2011. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2012. _C.TRAIN = CN()
2013. _C.TRAIN.ENABLE = False
2014. # Enable verbosity
2015. _C.TRAIN.VERBOSE = False
2016. # Optimizer(s) to use. Possible values: "SGD", "ADAM" or "ADAMW".
2017. _C.TRAIN.OPTIMIZER = ["SGD"]
2018. # Learning rate(s).
2019. _C.TRAIN.LR = [1.0e-4]
2020. # Weight decay
2021. _C.TRAIN.W_DECAY = 0.02
2022. # Coefficients used for computing running averages of gradient and its square. Used in ADAM and ADAMW optmizers
2023. _C.TRAIN.OPT_BETAS = [[0.9, 0.999]]
2024. # Batch size
2025. _C.TRAIN.BATCH_SIZE = 2
2026. # Number of epochs to train the model
2027. _C.TRAIN.EPOCHS = 360
2028. # Epochs to wait with no validation data improvement until the training is stopped
2029. _C.TRAIN.PATIENCE = -1
2030. # Metrics to apply during training. Depending on the workflow different ones can be applied. If empty, some
2031. # default metrics will be configured automatically:
2032. # * Semantic segmentation: 'iou' (called also Jaccard index)
2033. # * Instance segmentation: automatically set depending on the channels selected (PROBLEM.INSTANCE_SEG.DATA_CHANNELS).
2034. # * Detection: 'iou' (called also Jaccard index)
2035. # * Denoising: 'mae', 'mse'
2036. # * Super-resolution: "psnr", "mae", "mse", "ssim", "pcc"
2037. # * Self-supervision: "psnr", "mae", "mse", "ssim", "pcc"
2038. # * Classification: 'accuracy', 'top-5-accuracy'
2039. # * Image to image: "psnr", "mae", "mse", "ssim", "pcc"
2040. _C.TRAIN.METRICS = []
2042. # Gradient clipping max norm applied per optimizer. 0 = disabled.
2043. _C.TRAIN.GRADIENT_CLIP_NORM = 0.0
2045. # Callbacks
2046. # To determine which value monitor to consider which epoch consider the best to save. Currently not used.
2047. _C.TRAIN.CHECKPOINT_MONITOR = "val_loss"
2049. # Epochs to save a preview of the model's predictions on a fixed set of samples to PATHS.TRAIN_PRED_SAMPLES. Set it to -1 to not do it.
2050. _C.TRAIN.SAVE_TRAIN_PREDS_FREQ = -1
2051. # Number of fixed samples to track when 'TRAIN.SAVE_TRAIN_PREDS_FREQ' != -1.
2052. _C.TRAIN.SAVE_TRAIN_PREDS_NUM_SAMPLES = 1
2053. # Whether to also render the tracked samples inline (Jupyter/Colab only; ignored elsewhere).
2054. _C.TRAIN.SAVE_TRAIN_PREDS_SHOW = False
2055. # Add profiler callback to the training
2056. # _C.TRAIN.PROFILER = False
2057. # # Batch range to be analyzed
2058. # _C.TRAIN.PROFILER_BATCH_RANGE='10, 100'
2059. # _C.TRAIN.MAE_CALLBACK_EPOCHS = 5
2061. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2062. # 7.1 Learning rate (LE) scheduler options
2063. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2064. # LR Scheduler
2065. _C.TRAIN.LR_SCHEDULER = CN()
2066. _C.TRAIN.LR_SCHEDULER.NAME = "" # Possible options: 'warmupcosine', 'reduceonplateau', 'onecycle', 'warmupreduceonplateau'
2067. # Lower bound on the learning rate used in 'warmupcosine' and 'reduceonplateau'
2068. _C.TRAIN.LR_SCHEDULER.MIN_LR = [-1.0]
2070. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2071. # 7.1.1 Reduce on plateau options
2072. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2073. # The reduce on plateau scheduler reduces the learning rate when a metric has stopped improving.
2074. #
2075. # Factor by which the learning rate will be reduced
2076. _C.TRAIN.LR_SCHEDULER.REDUCEONPLATEAU_FACTOR = 0.5
2077. # Number of epochs with no improvement after which learning rate will be reduced. Need to be less than 'TRAIN.PATIENCE'
2078. # otherwise it makes no sense
2079. _C.TRAIN.LR_SCHEDULER.REDUCEONPLATEAU_PATIENCE = -1
2081. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2082. # 7.1.2 Cosine decay with warm up options
2083. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2084. # Cosine decay with a warm up consist in 2 phases: 1) a warm up phase which consists of increasing
2085. # the learning rate from TRAIN.LR_SCHEDULER.MIN_LR to TRAIN.LR value by a factor
2086. # during a certain number of epochs defined by 'TRAIN.LR_SCHEDULER.WARMUP_COSINE_DECAY_EPOCHS'
2087. # 2) after this will began the decay of the learning rate value using the cosine function.
2088. # Find a detailed explanation in: https://scorrea92.medium.com/cosine-learning-rate-decay-e8b50aa455b
2089. #
2090. # Epochs to do the warming up. -1/0 disables the ramp.
2091. _C.TRAIN.LR_SCHEDULER.WARMUP_COSINE_DECAY_EPOCHS = -1
2092. # Delays when decay starts: after warm up (if any), TRAIN.LR is held constant until
2093. # (1 - COSINE_DECAY_FRACTION) fraction of TRAIN.EPOCHS have passed, then cosine-decayed
2094. # over the final fraction ("delayed cosine decay"). Combinable with the warm up above
2095. # (ramp, then hold, then decay). Set to a value in (0, 1] to enable; -1 (default) disables
2096. # it, so decay starts right after warm up as usual.
2097. _C.TRAIN.LR_SCHEDULER.COSINE_DECAY_FRACTION = -1.0
2099. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2100. # 8. Test/inference phase options
2101. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2102. _C.TEST = CN()
2103. _C.TEST.ENABLE = False
2104. # Tries to reduce the memory footprint by separating crop/merge operations and by changing dtype of the predictions.
2105. # It is slower and not as precise as the "normal" inference process but saves memory
2106. _C.TEST.REDUCE_MEMORY = False
2107. # Whether to compute the metrics in CPU instead of GPU to reduce GPU memory consumption.
2108. _C.TEST.METRICS_IN_CPU = False
2109. # Whether to save the raw output of the model (before any post-processing) alongside the final prediction. It is placed normally
2110. # in a folder called 'per_image'
2111. _C.TEST.SAVE_MODEL_RAW_OUTPUT = True
2112. # Enable verbosity
2113. _C.TEST.VERBOSE = True
2114. # Make test-time augmentation. Infer over 8 possible rotations for 2D img and 16 when 3D
2115. _C.TEST.AUGMENTATION = False
2116. # Select test-time augmentation mode. Options: "mean" (default), "min", "max".
2117. # "min"/"max" are only applied to the channels they are meaningful for: a component-wise
2118. # minimum over several predicted flow/offset fields is not a flow field, so signed vector
2119. # channels (Cellpose/Omnipose 'Gv'/'Gh'/'Gz', HoVerNet 'V'/'H'/'Z', EmbedSeg 'E_offset')
2120. # are always averaged no matter what is set here.
2121. _C.TEST.AUGMENTATION_MODE = "mean"
2122. # Largest set of test-time augmentation orientations to use. Options:
2123. # - "auto" (default)/"full": 90 degree rotations + flips (8 in 2D, 16 in 3D).
2124. # - "flips": axis flips only (4 in 2D, 8 in 3D). What Cellpose does upstream, and the safe
2125. # choice for representations whose rotations cannot be undone exactly.
2126. # - "none": no augmentation (identity only).
2127. # Instance-segmentation representations that cannot represent an orientation exactly shrink
2128. # this set automatically (e.g. 3D StarDist rays, or EmbedSeg on anisotropic data); the
2129. # orientations actually used are printed once at the start of inference.
2130. _C.TEST.AUGMENTATION_GROUP = "auto"
2131. # Stack 2D images into a 3D image and then process it entirely instead of going image per image
2132. _C.TEST.ANALIZE_2D_IMGS_AS_3D_STACK = False
2133. # Whether to reuse the existing ones (from file) or calculate predictions using the model
2134. _C.TEST.REUSE_PREDICTIONS = False
2135. # If PROBLEM.NDIM = '2D' this can be activated to process each image entirely instead of patch by patch. Only can be done
2136. # if the neural network is fully convolutional. Implemented in semantic-segmentation, instance-segmentation and detection workflows.
2137. _C.TEST.FULL_IMG = False
2138. # Metrics to apply during test/inference. Depending on the workflow different ones can be applied. If empty, some
2139. # default metrics will be configured automatically:
2140. # * Semantic segmentation: 'iou' (called also Jaccard index)
2141. # * Instance segmentation: automatically set depending on the channels selected (PROBLEM.INSTANCE_SEG.DATA_CHANNELS).
2142. # Instance metrics will be always calculated.
2143. # * Detection: 'iou' (called also Jaccard index)
2144. # * Denoising: 'mae', 'mse'
2145. # * Super-resolution: "psnr", "mae", "mse", "ssim", "pcc". Additionally, if only if PROBLEM.NDIM == '2D', these
2146. # can also be selected: "fid", "is", "lpips"
2147. # * Self-supervision: "psnr", "mae", "mse", "ssim", "pcc". Additionally, if only if PROBLEM.NDIM == '2D', these
2148. # can also be selected: "fid", "is", "lpips"
2149. # * Classification: 'accuracy'. Always calculated: Confusion matrix
2150. # * Image to image: "psnr", "mae", "mse", "ssim", "pcc". Additionally, if only if PROBLEM.NDIM == '2D', these
2151. # can also be selected: "fid", "is", "lpips"
2152. _C.TEST.METRICS = []
2154. # Number of pixels/voxels to exclude from each border when computing test/inference metrics,
2155. # regardless of workflow. Order is: [z, y, x] (3D) and [y, x] (2D). For example, with an image
2156. # of 10x100x200 to ignore the first/last Z slices and a 15-pixel border on Y/X, use [1, 15, 15].
2157. # Predictions near the border are usually less reliable (the model has no context beyond the
2158. # edge, and patch-merging seams concentrate there too), which can inflate errors that don't
2159. # reflect the model's real performance. This never changes the predictions themselves (nor
2160. # anything saved to disk) -- only the region considered when scoring each metric, so only the
2161. # center of the image is evaluated:
2162. # * Semantic segmentation / voxel-level instance segmentation & membrane-repair metrics /
2163. # denoising / super-resolution / image-to-image / self-supervised: the border region of the
2164. # prediction and GT is excluded before computing pixel-wise metrics (IoU, MAE, MSE, SSIM,
2165. # PSNR, etc).
2166. # * Object-level instance segmentation & membrane-repair matching: a copy of the predicted/GT
2167. # instance label images has this border region blacked out (kept at the original shape)
2168. # before matching, so the TP/FP/FN color map stays full-size too.
2169. # * Detection: points (predicted or GT) whose coordinates fall in the border region are
2170. # excluded from the precision/recall/F1 computation.
2171. # * Classification: not applicable (no spatial dimension to crop); must be left empty.
2172. _C.TEST.EVAL_BORDER_CROP = []
2174. ### Instance segmentation
2175. # Whether to calculate matching statistics (average overlap, accuracy, recall, precision, etc.)
2176. _C.TEST.MATCHING_STATS = True
2177. # Theshold of overlap to consider a TP when calculating the metrics. If more than one value is provided
2178. # the process is repeated with each of the threshold values
2179. _C.TEST.MATCHING_STATS_THS = [0.3, 0.5, 0.75]
2180. # Decide in which thresholds to create a colored image of the TPs, FNs and FPs
2181. _C.TEST.MATCHING_STATS_THS_COLORED_IMG = []
2182. # How to aggregate the matching statistics across the dataset. If ``True`` the metrics (accuracy/AP_dsb,
2183. # precision, recall, etc.) are computed per image and then averaged over images (macro-average), which matches
2184. # the "AP_dsb" definition reported by EmbedSeg/StarDist. If ``False`` the TP/FP/FN counts are pooled across all
2185. # images and the metrics are computed once from those totals (micro-average).
2186. _C.TEST.MATCHING_STATS_BY_IMAGE = True
2188. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2189. # 8.1 Test/inference by chunks options
2190. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2191. # In the processing of 3D images, the primary image is segmented into smaller patches. These patches are subsequently
2192. # passed through a computational network. The outcome is a new image, typically saved as a TIF file, that retains the
2193. # dimensions of the original input. Notably, if the input image is sizable, this process can be memory-intensive. This
2194. # is because the quantity of patches is contingent on both the dimensions of the input and the selected padding/overlap
2195. # parameters (defined as 'DATA.TEST.PADDING' and 'DATA.TEST.OVERLAP').
2196. # To alleviate potential memory constraints, we offer an alternative: producing an H5/Zarr file with the predicted patches.
2197. # This method ensures efficient memory usage, as patches are individually incorporated into the H5/Zarr file in their respective
2198. # positions. This negates the need to store all patches simultaneously for image reconstruction. Importantly, in this
2199. # approach, only the 'DATA.TEST.PADDING' parameter is considered, excluding 'DATA.TEST.OVERLAP', which sufficiently
2200. # addresses border effect issues. If the source image is also an H5/Zarr file, it will be processed by chunks, further
2201. # optimizing memory usage. This process can only be applied in semantic segmentation, instance segmentation and detection
2202. # workflows.
2203. #
2204. _C.TEST.BY_CHUNKS = CN()
2205. _C.TEST.BY_CHUNKS.ENABLE = False
2206. # In the process of 'TEST.BY_CHUNKS' you can enable this variable to save the reconstructed prediction as a TIF too.
2207. # Be aware of this option and be sure that the prediction can fit in you memory entirely, as it is needed for saving as TIF.
2208. _C.TEST.BY_CHUNKS.SAVE_OUT_TIF = False
2209. # In how many iterations the H5 writer needs to flush the data. No need to do so with Zarr files.
2210. _C.TEST.BY_CHUNKS.FLUSH_EACH = 100
2211. # Z slice index (inclusive) at which chunk processing starts. Use -1 (default) to start from the beginning.
2212. # Useful for distributing large-volume prediction across multiple cluster jobs, each handling a Z sub-range.
2213. _C.TEST.BY_CHUNKS.Z_START = -1
2214. # Z slice index (exclusive) at which chunk processing ends. Use -1 (default) to process until the end.
2215. # The output Zarr is always the full data shape so multiple jobs can write concurrently.
2216. _C.TEST.BY_CHUNKS.Z_END = -1
2217. # Whether to write the raw model predictions into a Zarr file. They are not needed to create the final output, as
2218. # each tile is post-processed as soon as it is predicted, and they are heavy to store (e.g. 12 float32 channels
2219. # per voxel for affinities), so they are only written on demand. Forced when the workflow process is disabled or
2220. # set to 'entire_pred', when 'TEST.REUSE_PREDICTIONS' is enabled or when the raw prediction is the final output
2221. # of the workflow (e.g. denoising).
2222. _C.TEST.BY_CHUNKS.WRITE_RAW_PREDICTIONS = False
2223. # Phases to execute in this job. Allows splitting the full pipeline across multiple cluster jobs.
2224. # Available phases:
2225. # * 'prediction' : run the model on each patch. Post-processing runs on the fly unless this is the only
2226. # phase of the job, in which case the raw predictions are written to a Zarr file.
2227. # * 'instance_creation' : run per-chunk watershed (instance segmentation, 'chunk_by_chunk' only). Runs within
2228. # 'prediction' when both phases are selected, otherwise it reads the raw predictions Zarr.
2229. # * 'instance_merging' : resolve cross-chunk instance IDs (Passes B–E, instance segmentation only).
2230. # Example multi-job setup for a large volume:
2231. # Job 1 — Z_START=0, Z_END=500, PHASES=["prediction", "instance_creation"]
2232. # Job 2 — Z_START=500, Z_END=1000, PHASES=["prediction", "instance_creation"]
2233. # Job 3 — (no Z range) PHASES=["instance_merging"]
2234. _C.TEST.BY_CHUNKS.PHASES = ["prediction", "instance_creation", "instance_merging"]
2236. # After passing all the patches through the model we obtain the model's raw predictions, which may be subsequently processed
2237. # to generate the final prediction. Each workflow has its own steps for this process:
2238. #
2239. # * Semantic segmentation: all raw predictions, which are probabilities of each class, are merged together to create the final
2240. # predicted image by argmax operation. This process is not memory intensive as the raw predictions are merged patch by
2241. # patch and not all at once.
2242. # * Instance segmentation: the raw predictions are merged together to create the final predicted image by a watershed process.
2243. # This process is memory intensive as the entire predicted image needs to be loaded in memory to do it.
2244. # * Detection: All the points of interest are detected from the raw predictions. This process is not memory intensive as the
2245. # points are detected patch by patch and not all at once.
2246. _C.TEST.BY_CHUNKS.WORKFLOW_PROCESS = CN()
2247. _C.TEST.BY_CHUNKS.WORKFLOW_PROCESS.ENABLE = True
2248. # How the workflow process is going to be done. There are two options:
2249. # * 'chunk_by_chunk' : each chunk will be considered as an individual file. Select this operation if you have not enough
2250. # memory to process the entire prediction image with 'entire_pred'.
2251. # * 'entire_pred': the predicted image will be loaded in memory and processed entirely (be aware of your memory budget)
2252. _C.TEST.BY_CHUNKS.WORKFLOW_PROCESS.TYPE = "chunk_by_chunk"
2253. # Patches, in (z,y,x) order, grouped into each tile of the workflow process ('chunk_by_chunk' mode only). A tile
2254. # is post-processed at once, using the 'DATA.TEST.PADDING' of its border patches as context, so it covers
2255. # PATCHES_PER_TILE * ('DATA.PATCH_SIZE' - 2 * 'DATA.TEST.PADDING') voxels per axis. Tiles divide the patches
2256. # among them, so this does not change the number of patches passed through the model nor their overlap: larger
2257. # tiles only mean fewer seams between them, and therefore fewer instances split across tiles that instance
2258. # segmentation has to merge back. Raising it improves the result, and the limit is the RAM available.
2259. #
2260. # Each tile being filled holds the prediction of its region plus the padding around it, i.e.
2261. # (tile + 2 * padding voxels per axis) * model output channels * 4 bytes
2262. # and post-processing one takes about two more tiles of working memory. Every one of the 'SYSTEM.NUM_GPUS'
2263. # processes fills as many tiles at the same time as workers it has ('SYSTEM.NUM_WORKERS'), so a job needs about
2264. # NUM_GPUS * ((NUM_WORKERS + 2) * tile size + 3 GiB)
2265. # where those 3 GiB are what BiaPy itself takes per GPU (model, CUDA context and the patches in flight).
2266. #
2267. # Measured on a (1024,1024,128) volume with 4 GPUs and 5 workers each, a patch of (128,128,128), a padding of
2268. # (10,10,10) and 12 output channels (peak RAM of the whole job, and the instances split across tiles that the
2269. # merging has to join back):
2270. # PATCHES_PER_TILE tile tile size RAM instances found
2271. # (1,1,1) (108,108,108) 96 MiB 13.7 GiB 953
2272. # (2,2,2) (216,216,216) 327 MiB 19.7 GiB 836
2273. # (3,3,3) (324,324,324) 693 MiB 22.9 GiB 828
2274. #
2275. # Values to ask for in the same setup, but on a volume large enough for the tiles not to be cut by its borders:
2276. # RAM of the job PATCHES_PER_TILE tile needs
2277. # 100 GiB (3,3,3) (324,324,324) 63 GiB
2278. # 300 GiB (5,5,5) (540,540,540) 232 GiB
2279. # 500 GiB (6,6,6) (648,648,648) 385 GiB
2280. # 700 GiB (7,7,7) (756,756,756) 597 GiB
2281. # 1000 GiB (8,8,8) (864,864,864) 877 GiB
2282. # Halve the values above if the model outputs twice the channels, and note that the tile size grows with the
2283. # cube of this value when the three axes are raised together: raise only the axes that need it if the memory
2284. # gets tight (e.g. (1,2,2) on anisotropic data, or when an axis is already covered by a single tile).
2285. _C.TEST.BY_CHUNKS.WORKFLOW_PROCESS.PATCHES_PER_TILE = (1, 1, 1)
2286. # Minimum normalised IoU required to merge two instances across a chunk boundary
2287. # (instance segmentation, 'chunk_by_chunk' mode only). Must be in (0, 1].
2288. _C.TEST.BY_CHUNKS.WORKFLOW_PROCESS.INSTANCE_SEG_MERGE_IOU_TH = 0.3
2291. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2292. # 8.2 Detection test/inference options
2293. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2294. # To decide which function is going to be used to create point from probabilities. Options: ['peak_local_max', 'blob_log']
2295. # 'peak_local_max': https://scikit-image.org/docs/stable/api/skimage.feature.html#skimage.feature.peak_local_max
2296. # 'blob_log': https://scikit-image.org/docs/stable/api/skimage.feature.html#skimage.feature.blob_log
2297. _C.TEST.DET_POINT_CREATION_FUNCTION = "peak_local_max"
2298. # The minimal allowed distance separating peaks. To find the maximum number of peaks, use min_distance=1.
2299. _C.TEST.DET_PEAK_LOCAL_MAX_MIN_DISTANCE = 1
2300. # Whether the threshold are going to be set as automaticaly (with Otsu thresholding) or manually.
2301. # Options available: 'auto' or 'manual'. If this last is used TEST.DET_MIN_TH_TO_BE_PEAK needs to be set.
2302. _C.TEST.DET_TH_TYPE = "manual"
2303. # Minimun value to consider a point as a peak. Corresponds to 'threshold_abs' argument of the function
2304. # 'peak_local_max' of skimage.feature
2305. _C.TEST.DET_MIN_TH_TO_BE_PEAK = 0.2
2306. # Corresponds to 'exclude_border' argument of 'peak_local_max' or 'blob_log' function of skimage. If True it will exclude
2307. # peaks from the border of the image to avoid partial detection.
2308. _C.TEST.DET_EXCLUDE_BORDER = False
2309. # Corresponds to 'min_sigma' argument of 'blob_log' function. It is the minimum standard deviation for Gaussian kernel.
2310. # Keep this low to detect smaller blobs. The standard deviations of the Gaussian filter are given for each axis as a
2311. # sequence, or as a single number, in which case it is equal for all axes.
2312. _C.TEST.DET_BLOB_LOG_MIN_SIGMA = 5
2313. # Corresponds to 'max_sigma' argument of 'blob_log' function. It is the maximum standard deviation for Gaussian kernel.
2314. # Keep this high to detect larger blobs. The standard deviations of the Gaussian filter are given for each axis as a
2315. # sequence, or as a single number, in which case it is equal for all axes.
2316. _C.TEST.DET_BLOB_LOG_MAX_SIGMA = 10
2317. # Corresponds to 'num_sigma' argument of 'blob_log' function. The number of intermediate values of standard deviations
2318. # to consider between min_sigma and max_sigma.
2319. _C.TEST.DET_BLOB_LOG_NUM_SIGMA = 2
2320. # Maximum distance far away from a GT point to consider a point as a true positive
2321. _C.TEST.DET_TOLERANCE = 10
2323. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2324. # 8.3 Post-processing options
2325. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2326. _C.TEST.POST_PROCESSING = CN()
2327. # To apply median filtering to the data
2328. _C.TEST.POST_PROCESSING.MEDIAN_FILTER = False
2329. # List of median filters to apply. They are going to be applied in the list order. This can only be used in
2330. # 'SEMANTIC_SEG', 'INSTANCE_SEG' and 'DETECTION' workflows. There are multiple options to compose the list:
2331. # * 'xy' or 'yx': to apply the filter in x and y axes together
2332. # * 'zy' or 'yz': to apply the filter in y and z axes together
2333. # * 'zx' or 'xz': to apply the filter in x and z axes together
2334. # * 'z': to apply the filter only in z axis
2335. # Those filter that imply 'z' axis are going to be applied only in 3D or in 2D if TEST.ANALIZE_2D_IMGS_AS_3D_STACK is selected
2336. _C.TEST.POST_PROCESSING.MEDIAN_FILTER_AXIS = []
2337. _C.TEST.POST_PROCESSING.MEDIAN_FILTER_SIZE = []
2339. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2340. # 8.3.1 Instance segmentation post-processing options
2341. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2342. # Instance refinement:
2343. # Whether to refine the instances with morphological and filtering operations after being created and before any other post-processing
2344. # such as Voronoi. This instance refinement is applied on each instance individually and sequentially. The two variables,
2345. # TEST.POST_PROCESSING.INSTANCE_REFINEMENT.PROPS and TEST.POST_PROCESSING.INSTANCE_REFINEMENT.VALUES of refinements to be applied.
2346. # For instance, the conditions can be like this: ['A', 'B', 'C'] and they will be applied sequentially.
2347. # A full example of this refinement:
2348. # _C.DATA.VAL.FILTER_SAMPLES.PROPS = ['fill_holes', 'remove_small_objects', 'remove_large_objects', 'clear_border']
2349. # _C.DATA.VAL.FILTER_SAMPLES.VALUES = ['none', 10, 2000, 'none']
2350. _C.TEST.POST_PROCESSING.INSTANCE_REFINEMENT = CN()
2351. # Enable/disable instance refinement
2352. _C.TEST.POST_PROCESSING.INSTANCE_REFINEMENT.ENABLE = False
2353. # List of morphological operations to apply. They are going to be applied in the list order. Available operations are:
2354. # * 'dilation': to dilate instances
2355. # * 'erosion': to erode instances
2356. # * 'fill_holes': to fill holes inside instances
2357. # * 'clear_border': to remove instances touching the image border
2358. # * 'remove_small_objects': to remove small objects
2359. # * 'remove_large_objects': to remove large objects
2360. _C.TEST.POST_PROCESSING.INSTANCE_REFINEMENT.OPERATIONS = []
2361. # Values associated to each operation. For 'dilation' and 'erosion' it corresponds to the size of the structuring element (it can also be a list).
2362. # For 'remove_small_objects' and 'remove_large_objects' it corresponds to the size threshold in pixels.
2363. # For 'fill_holes' and 'clear_border' no value is needed so put None in those cases.
2364. _C.TEST.POST_PROCESSING.INSTANCE_REFINEMENT.VALUES = []
2366. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2367. # 8.3.1.1 Instance property measurement and filtering options
2368. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2369. # Whether to measure morphological features on each instances, i.e. 'circularity' (2D), 'elongation' (2D), 'npixels', 'area', 'diameter',
2370. # 'perimeter', 'sphericity' (3D)
2371. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES = CN()
2372. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.ENABLE = False
2373. # List of properties to measure on each instance. The following properties will be always calculated: label, npixels, areas, centers, elongation (2D),
2374. # sphericities (2D)/circularities (3D), diameters, perimeter (2D)/surface_area (3D). Apart from them, you can select more properties to be calculated
2375. # based on scikit-image regionprops function. Check the following link for a detailed list of the extra available properties you can request:
2376. # https://scikit-image.org/docs/stable/api/skimage.measure.html#skimage.measure.regionprops
2377. # Default property description is as follows:
2378. # * 'circularity' is defined as the ratio of area over the square of the perimeter, normalized such that the value for a disk equals
2379. # one: (4 * PI * area) / (perimeter^2). Only measurable for 2D images (use sphericity for 3D images). While values of circularity
2380. # range theoretically within the interval [0;1], the measurements errors of the perimeter may produce circularity values above 1
2381. # (Lehmann et al., 201211 ; https://doi.org/10.1093/bioinformatics/btw413).
2382. #
2383. # * 'elongation' is the inverse of the circularity. The values of elongation range from 1 for round particles and increase for
2384. # elongated particles. Calculated as: (perimeter^2)/(4 * PI * area) . It is only measurable for 2D images.
2385. #
2386. # * 'npixels' corresponds to the sum of pixels that compose an instance.
2387. #
2388. # * 'area' corresponds to the number of pixels taking into account the image resolution (we call it 'area' also even in a 3D
2389. # image for simplicity, but that will be the volume in that case). In the resulting statistics 'volume' will appear in that
2390. # case too.
2391. #
2392. # * 'diameter' calculated with the bounding box and by taking the maximum value of the box in x and y axes. In 3D, z axis
2393. # is also taken into account. Does not take into account the image resolution.
2394. #
2395. # * 'perimeter', in 2D, approximates the contour as a line through the centers of border pixels using a 4-connectivity. In 3D,
2396. # it corresponds to the surface area.
2397. #
2398. # * 'sphericity', in 3D, it is the ratio of the squared volume over the cube of the surface area, normalized such that the value
2399. # for a ball equals one: (36 * PI)*((volume^2)/(perimeter^3)). Only measurable for 3D images (use circularity for 2D images).
2400. #
2401. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.EXTRA_PROPS = []
2402. # Remove instances by the conditions based in each instance properties. The three variables, TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.PROPS,
2403. # TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.VALUES and TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.SIGNS will compose a list
2404. # of conditions to remove the instances. They are list of list of conditions. For instance, the conditions can be like this: [['A'], ['B','C']]. Then, if the instance satisfies
2405. # the first list of conditions, only 'A' in this first case (from ['A'] list), or satisfy 'B' and 'C' (from ['B','C'] list) it will be
2406. # removed from the image. In each sublist all the conditions must be satisfied. Available properties are: ['circularity', 'elongation',
2407. # 'npixels', 'area', 'diameter', 'perimeter', 'sphericity']. When this post-processing step is selected two .csv files
2408. # will be created, one with the properties of each instance from the original image (will be placed in PATHS.RESULT_DIR.PER_IMAGE_INSTANCES
2409. # path), and another with only instances that remain once this post-processing has been applied (will be placed in
2410. # PATHS.RESULT_DIR.PER_IMAGE_POST_PROCESSING path). In those csv files two more information columns will appear: a list of conditions
2411. # that each instance has satisfy or not ('Satisfied', 'No satisfied' respectively), and a comment with two possible values, 'Removed'
2412. # and 'Correct', telling you if the instance has been removed or not, respectively. Some of the properties follow the formulas used in
2413. # MorphoLibJ library for Fiji https://doi.org/10.1093/bioinformatics/btw413
2414. #
2415. # A full example of this post-processing:
2416. # If you want to remove those instances that have less than 100 pixels and circularity less equal to 0.7 you should
2417. # declare the above three variables as follows:
2418. # _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.PROPS = [['npixels', 'circularity']]
2419. # _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.VALUES = [[100, 0.7]]
2420. # _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.SIGNS = [['lt', 'le']]
2421. # You can also concatenate more restrictions and they will be applied in order. For instance, if you want to remove those
2422. # instances that are bigger than an specific area, and do that before the condition described above, you can define the
2423. # variables this way:
2424. # _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.PROPS = [['area'], ['npixels', 'circularity']]
2425. # _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.VALUES = [[500], [100, 0.7]]
2426. # _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.SIGNS = [['gt'], ['lt', 'le']]
2427. # This way, the instances will be removed by 'area' and then by 'npixels' and 'circularity'
2428. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES = CN()
2429. # Whether to enable or not the filtering by properties
2430. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.ENABLE = False
2431. # List of lists of properties to apply a filter. Available properties are: ['circularity', 'elongation', 'npixels', 'area', 'diameter',
2432. # 'perimeter', 'sphericity']
2433. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.PROPS = []
2434. # List of ints/float that represent the values of the properties listed in TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES
2435. # that the instances need to satisfy to not be dropped.
2436. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.VALUES = []
2437. # List of list of signs to do the comparison. Options: ['gt', 'ge', 'lt', 'le'] that corresponds to "greather than", e.g. ">",
2438. # "greather equal", e.g. ">=", "less than", e.g. "<", and "less equal" e.g. "<=" comparisons.
2439. _C.TEST.POST_PROCESSING.MEASURE_PROPERTIES.REMOVE_BY_PROPERTIES.SIGNS = []
2441. # Whether to apply Voronoi using 'BC' or 'M' channels need to be present
2442. _C.TEST.POST_PROCESSING.VORONOI_ON_MASK = False
2443. # Threshold to be applied to the 'M' channel when expanding the instances with Voronoi. Need to be in [0,1] range.
2444. # Leave it to 0 to adjust the threhold with Otsu
2445. _C.TEST.POST_PROCESSING.VORONOI_TH = 0.0
2446. # Set it to try to repare large instances by merging their neighbors with them and removing possible central holes.
2447. # Its value determines which instances are going to be repared by size (number of pixels that compose the instance)
2448. # This option is useful when PROBLEM.INSTANCE_SEG.DATA_CHANNELS is 'BP', as multiple central seeds may appear in big
2449. # instances. Only works in Instance segmentation workflow.
2450. _C.TEST.POST_PROCESSING.REPARE_LARGE_BLOBS_SIZE = -1
2452. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2453. # 8.3.2 Detection post-processing options
2454. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2455. # To remove close points to each other. This can also be set when using 'BP' channels for instance segmentation.
2456. _C.TEST.POST_PROCESSING.REMOVE_CLOSE_POINTS = False
2457. # Distance between points to be considered the same. Only applies when TEST.POST_PROCESSING.REMOVE_CLOSE_POINTS = True
2458. # This can also be set when using 'BP' channels for instance segmentation.
2459. _C.TEST.POST_PROCESSING.REMOVE_CLOSE_POINTS_RADIUS = 0
2460. # Whether to apply a watershed to grow the points detected
2461. _C.TEST.POST_PROCESSING.DET_WATERSHED = False
2462. # Structure to dilate the initial seeds before watershed. For instance in a 3D problem: [10,10,4]
2463. _C.TEST.POST_PROCESSING.DET_WATERSHED_FIRST_DILATION = [-1, -1]
2464. # List of classes to be consider as 'donuts'. For those class points, the 'donuts' type cell means that their nucleus is
2465. # to big and that the seeds need to be dilated more so the watershed can grow the instances properly.
2466. _C.TEST.POST_PROCESSING.DET_WATERSHED_DONUTS_CLASSES = [-1]
2467. # Patch shape to extract all donuts type cells. It needs to be a bit greater than bigest donuts type cell so all of them can
2468. # be contained in this patch. This is used to analize that area for each point of class `DET_WATERSHED_DONUTS_CLASSES`.
2469. _C.TEST.POST_PROCESSING.DET_WATERSHED_DONUTS_PATCH = [13, 120, 120]
2470. # Diameter (in pixels) that a cell need to have to be considered as donuts type
2471. _C.TEST.POST_PROCESSING.DET_WATERSHED_DONUTS_NUCLEUS_DIAMETER = 30
2473. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2474. # 9. Auxiliary paths
2475. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2476. _C.PATHS = CN()
2478. # Directories to store the results
2479. _C.PATHS.RESULT_DIR = CN()
2480. _C.PATHS.RESULT_DIR.PATH = os.path.join(job_dir, "results", job_identifier)
2481. _C.PATHS.RESULT_DIR.PER_IMAGE = os.path.join(_C.PATHS.RESULT_DIR.PATH, "per_image")
2482. _C.PATHS.RESULT_DIR.PER_IMAGE_BIN = os.path.join(_C.PATHS.RESULT_DIR.PATH, "per_image_binarized")
2483. _C.PATHS.RESULT_DIR.PER_IMAGE_INSTANCES = os.path.join(_C.PATHS.RESULT_DIR.PATH, "per_image_instances")
2484. _C.PATHS.RESULT_DIR.PER_IMAGE_POST_PROCESSING = os.path.join(
2485. _C.PATHS.RESULT_DIR.PATH, "per_image_post_processing"
2486. )
2487. _C.PATHS.RESULT_DIR.FULL_IMAGE = os.path.join(_C.PATHS.RESULT_DIR.PATH, "full_image")
2488. _C.PATHS.RESULT_DIR.FULL_IMAGE_BIN = os.path.join(_C.PATHS.RESULT_DIR.PATH, "full_image_binarized")
2489. _C.PATHS.RESULT_DIR.FULL_IMAGE_INSTANCES = os.path.join(_C.PATHS.RESULT_DIR.PATH, "full_image_instances")
2490. _C.PATHS.RESULT_DIR.FULL_IMAGE_POST_PROCESSING = os.path.join(
2491. _C.PATHS.RESULT_DIR.PATH, "full_image_post_processing"
2492. )
2493. _C.PATHS.RESULT_DIR.AS_3D_STACK = os.path.join(_C.PATHS.RESULT_DIR.PATH, "as_3d_stack")
2494. _C.PATHS.RESULT_DIR.AS_3D_STACK_BIN = os.path.join(_C.PATHS.RESULT_DIR.PATH, "as_3d_stack_binarized")
2495. _C.PATHS.RESULT_DIR.AS_3D_STACK_POST_PROCESSING = os.path.join(
2496. _C.PATHS.RESULT_DIR.PATH, "as_3d_stack_post_processing"
2497. )
2498. _C.PATHS.RESULT_DIR.DET_LOCAL_MAX_COORDS_CHECK = os.path.join(
2499. _C.PATHS.RESULT_DIR.PATH, "per_image_local_max_check"
2500. )
2501. _C.PATHS.RESULT_DIR.DET_LOCAL_MAX_COORDS_CHECK_POST_PROCESSING = os.path.join(
2502. _C.PATHS.RESULT_DIR.PATH, "per_image_local_max_check_post_processing"
2503. )
2504. _C.PATHS.RESULT_DIR.DET_ASSOC_POINTS = os.path.join(_C.PATHS.RESULT_DIR.PATH, "point_associations")
2505. _C.PATHS.RESULT_DIR.INST_ASSOC_POINTS = os.path.join(_C.PATHS.RESULT_DIR.PATH, "instance_associations")
2506. # Path to store the BMZ model created
2507. _C.PATHS.BMZ_EXPORT_PATH = os.path.join(_C.PATHS.RESULT_DIR.PATH, "BMZ_files")
2509. # Path to store profiler files
2510. _C.PATHS.PROFILER = os.path.join(_C.PATHS.RESULT_DIR.PATH, "profiler")
2512. # Name of the folder where the charts of the loss and metrics values while training the network are stored.
2513. _C.PATHS.CHARTS = os.path.join(_C.PATHS.RESULT_DIR.PATH, "charts")
2514. # Folder where TRAIN.SAVE_TRAIN_PREDS_FREQ previews are stored.
2515. _C.PATHS.TRAIN_PRED_SAMPLES = os.path.join(_C.PATHS.RESULT_DIR.PATH, "train_pred_samples")
2516. # Folder where samples of DA will be stored
2517. _C.PATHS.DA_SAMPLES = os.path.join(_C.PATHS.RESULT_DIR.PATH, "aug")
2518. # Folder where generator samples (X) will be stored
2519. _C.PATHS.GEN_CHECKS = os.path.join(_C.PATHS.RESULT_DIR.PATH, "gen_check")
2520. # Folder where generator samples (Y) will be stored
2521. _C.PATHS.GEN_MASK_CHECKS = os.path.join(_C.PATHS.RESULT_DIR.PATH, "gen_mask_check")
2522. # Paths where a few samples of instance channels created will be stored just to check id there is any problem
2523. _C.PATHS.TRAIN_INSTANCE_CHANNELS_CHECK = os.path.join(
2524. _C.PATHS.RESULT_DIR.PATH,
2525. "train_" + "".join(_C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS) + "_instance_channels",
2526. )
2527. _C.PATHS.VAL_INSTANCE_CHANNELS_CHECK = os.path.join(
2528. _C.PATHS.RESULT_DIR.PATH,
2529. "val_" + "".join(_C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS) + "_instance_channels",
2530. )
2531. _C.PATHS.TEST_INSTANCE_CHANNELS_CHECK = os.path.join(
2532. _C.PATHS.RESULT_DIR.PATH,
2533. "test_" + "".join(_C.PROBLEM.INSTANCE_SEG.DATA_CHANNELS) + "_instance_channels",
2534. )
2535. # Name of the folder where weights files will be stored/loaded from.
2536. _C.PATHS.CHECKPOINT = os.path.join(job_dir, "checkpoints")
2537. # Checkpoint file to load/store the model weights
2538. _C.PATHS.CHECKPOINT_FILE = ""
2539. # Name of the folder to store the probability map to avoid recalculating it on every run.
2540. # Overridden at runtime to "<DATA.TRAIN.GT_PATH>_prob_maps" (see create_train_val_augmentors),
2541. # mirroring the DATA.TRAIN.INSTANCE_CHANNELS_MASK_DIR/DETECTION_MASK_DIR naming convention.
2542. _C.PATHS.PROB_MAP_DIR = os.path.join(job_dir, "prob_map")
2543. _C.PATHS.PROB_MAP_FILENAME = "prob_map.npy"
2544. # Watershed debugging folder
2545. _C.PATHS.WATERSHED_DIR = os.path.join(_C.PATHS.RESULT_DIR.PATH, "watershed")
2546. # Path where the images used in MAE will be saved suring inference
2547. _C.PATHS.MAE_OUT_DIR = os.path.join(_C.PATHS.RESULT_DIR.PATH, "MAE_checks")
2548. # Directory to save filtered images.
2549. _C.PATHS.FIL_SAMPLES_DIR = os.path.join(_C.PATHS.RESULT_DIR.PATH, "filtering_information")
2551. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2552. # 10. Logging
2553. # ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2554. _C.LOG = CN()
2555. _C.LOG.LOG_DIR = os.path.join(job_dir, "logs")
2556. _C.LOG.TENSORBOARD_LOG_DIR = os.path.join(_C.PATHS.RESULT_DIR.PATH, "tensorboard")
2557. _C.LOG.LOG_FILE_PREFIX = job_identifier
2558. _C.LOG.CHART_CREATION_FREQ = 5
2560. self._C = _C
2562. def get_cfg_defaults(self) -> CN:
2563. """
2564. Get a clone of the default configuration.
2566. Returns
2567. -------
2568. CN
2569. A cloned YACS CfgNode object with default values.
2570. """
2571. # Return a clone so that the defaults will not be altered
2572. # This is for the "local variable" use pattern
2573. return self._C.clone()
2575. def to_dict(self):
2576. """
2577. Convert the configuration to a Python dictionary.
2579. Returns
2580. -------
2581. dict
2582. Dictionary representation of the configuration.
2583. """
2584. return dict(self._C)
2586. def copy(self):
2587. """
2588. Create a deep copy of the Config object.
2590. Returns
2591. -------
2592. Config
2593. A deep copy of the current Config instance.
2594. """
2595. return copy.deepcopy(self)
2597. def __str__(self):
2598. """
2599. Return a string representation of the Config object.
2601. Returns
2602. -------
2603. str
2604. String representation of the configuration.
2605. """
2606. return str(self.__dict__)
2608. def __repr__(self):
2609. """
2610. Return a string representation of the Config object.
2612. Returns
2613. -------
2614. str
2615. String representation of the configuration.
2616. """
2617. return str(self.__dict__)
2619. def update_dependencies(cfg) -> None:
2620. """
2621. Update dependent configuration variables after merging user config.
2623. This function should be called after merging a user-provided .cfg file
2624. to ensure that all dependent paths and variables are updated accordingly.
2625. That is, this function should be called after YACS's merge_from_file().
2627. Parameters
2628. ----------
2629. cfg : Config or CN
2630. The configuration object to update.
2631. """
2632. call = getattr(cfg, "_C") if bool(getattr(cfg, "_C", False)) else cfg
2633. # Remove possible / characters at the end of the paths
2634. call.DATA.TRAIN.PATH = (
2635. call.DATA.TRAIN.PATH if call.DATA.TRAIN.PATH[-1] != "/" else call.DATA.TRAIN.PATH[:-1]
2636. )
2637. call.DATA.TRAIN.GT_PATH = (
2638. call.DATA.TRAIN.GT_PATH if call.DATA.TRAIN.GT_PATH[-1] != "/" else call.DATA.TRAIN.GT_PATH[:-1]
2639. )
2640. call.DATA.VAL.PATH = (
2641. call.DATA.VAL.PATH if call.DATA.VAL.PATH[-1] != "/" else call.DATA.VAL.PATH[:-1]
2642. )
2643. call.DATA.VAL.GT_PATH = (
2644. call.DATA.VAL.GT_PATH if call.DATA.VAL.GT_PATH[-1] != "/" else call.DATA.VAL.GT_PATH[:-1]
2645. )
2646. call.DATA.TEST.PATH = (
2647. call.DATA.TEST.PATH if call.DATA.TEST.PATH[-1] != "/" else call.DATA.TEST.PATH[:-1]
2648. )
2649. call.DATA.TEST.GT_PATH = (
2650. call.DATA.TEST.GT_PATH if call.DATA.TEST.GT_PATH[-1] != "/" else call.DATA.TEST.GT_PATH[:-1]
2651. )
2652. tdata = call.DATA.TRAIN.GT_PATH if not call.DATA.TRAIN.INPUT_ZARR_MULTIPLE_DATA else call.DATA.TRAIN.PATH
2653. call.DATA.TRAIN.INSTANCE_CHANNELS_MASK_DIR = tdata
2654. cpd = '_'.join([str(x) for x in call.PROBLEM.DETECTION.CENTRAL_POINT_DILATION])
2655. call.DATA.TRAIN.DETECTION_MASK_DIR = call.DATA.TRAIN.GT_PATH + "_detection_masks_" + str(cpd)
2656. call.DATA.TRAIN.SSL_SOURCE_DIR = call.DATA.TRAIN.PATH + "_ssl_source"
2657. vdata = call.DATA.VAL.GT_PATH if not call.DATA.VAL.INPUT_ZARR_MULTIPLE_DATA else call.DATA.VAL.PATH
2658. call.DATA.VAL.INSTANCE_CHANNELS_MASK_DIR = vdata
2659. # If value is not the default
2660. call.DATA.VAL.DETECTION_MASK_DIR = call.DATA.VAL.GT_PATH + "_detection_masks_" + str(cpd)
2661. call.DATA.VAL.SSL_SOURCE_DIR = call.DATA.VAL.PATH + "_ssl_source"
2662. tdata = call.DATA.TEST.GT_PATH if not call.DATA.TEST.INPUT_ZA

config.py at commit c560d40, under MIT · at the source