forked from stefanklut/laypa
-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathextra_defaults.py
360 lines (296 loc) · 10.6 KB
/
extra_defaults.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
from detectron2.config import CfgNode as CN
_C = CN()
# Logging
_C.CONFIG_PATH = ""
_C.TRAINING_PATHS = []
_C.VALIDATION_PATHS = []
# Save name configuration while training, RUN_DIR will add a timestamp, NAME will add a custom name
_C.RUN_DIR = True
_C.NAME = ""
# Automatically filled do not expect values to remain
_C.LAYPA_UUID = ""
_C.LAYPA_GIT_HASH = ""
_C.SETUP_TIME = ""
# Model changes
_C.MODEL = CN()
_C.MODEL.RESUME = False
_C.MODEL.MODE = ""
_C.MODEL.SEM_SEG_HEAD = CN()
_C.MODEL.SEM_SEG_HEAD.WEIGHT = [1.0]
_C.MODEL.BINARY_SEG_HEAD = CN()
_C.MODEL.BINARY_SEG_HEAD.NAME = "BinarySegFPNHead"
_C.MODEL.BINARY_SEG_HEAD.IN_FEATURES = ["p2", "p3", "p4", "p5"]
# Label in the semantic segmentation ground truth that is ignored, i.e., no loss is calculated for
# the correposnding pixel.
_C.MODEL.BINARY_SEG_HEAD.IGNORE_VALUE = 255
# Number of classes in the semantic segmentation head
_C.MODEL.BINARY_SEG_HEAD.NUM_CLASSES = 54
# Number of channels in the 3x3 convs inside semantic-FPN heads.
_C.MODEL.BINARY_SEG_HEAD.CONVS_DIM = 128
# Outputs from semantic-FPN heads are up-scaled to the COMMON_STRIDE stride.
_C.MODEL.BINARY_SEG_HEAD.COMMON_STRIDE = 4
# Normalization method for the convolution layers. Options: "" (no norm), "GN".
_C.MODEL.BINARY_SEG_HEAD.NORM = "GN"
_C.MODEL.BINARY_SEG_HEAD.LOSS_WEIGHT = 1.0
# Automatic mixed precision settings
_C.MODEL.AMP_TRAIN = CN()
_C.MODEL.AMP_TRAIN.ENABLED = False
_C.MODEL.AMP_TRAIN.PRECISION = "bfloat16"
_C.MODEL.AMP_TEST = CN()
_C.MODEL.AMP_TEST.ENABLED = False
_C.MODEL.AMP_TEST.PRECISION = "bfloat16"
# Weights loaded during training
_C.TRAIN = CN()
_C.TRAIN.WEIGHTS = ""
# Weights loaded during testing/inference
_C.TEST = CN()
_C.TEST.WEIGHTS = ""
# Preprocessing
_C.PREPROCESS = CN()
_C.PREPROCESS.OUTPUT = [
("image", "png"),
("sem_seg", "png"),
("instances", "json"),
("pano", "png"),
]
# Preprocessing check if files specified in .txt file exist
_C.PREPROCESS.DISABLE_CHECK = False
# Overwrite existing files in temporary directory
_C.PREPROCESS.OVERWRITE = False
# PageXML region conversion
_C.PREPROCESS.REGION = CN()
_C.PREPROCESS.REGION.REGIONS = []
_C.PREPROCESS.REGION.MERGE_REGIONS = []
_C.PREPROCESS.REGION.REGION_TYPE = []
_C.PREPROCESS.REGION.RECTANGLE_REGIONS = []
_C.PREPROCESS.REGION.MIN_REGION_SIZE = 10
# PageXML baseline conversion
_C.PREPROCESS.BASELINE = CN()
_C.PREPROCESS.BASELINE.LINE_WIDTH = 5
_C.PREPROCESS.BASELINE.SQUARE_LINES = False
# Preprocessing resizing options
_C.PREPROCESS.RESIZE = CN()
_C.PREPROCESS.RESIZE.USE = False
_C.PREPROCESS.RESIZE.RESIZE_MODE = "shortest_edge"
_C.PREPROCESS.RESIZE.SCALING = 1.0
_C.PREPROCESS.RESIZE.RESIZE_SAMPLING = "choice"
_C.PREPROCESS.RESIZE.MIN_SIZE = [1024]
_C.PREPROCESS.RESIZE.MAX_SIZE = 2048
# DPI correction in resizing
_C.PREPROCESS.DPI = CN()
_C.PREPROCESS.DPI.AUTO_DETECT = False
_C.PREPROCESS.DPI.DEFAULT_DPI = 300
_C.PREPROCESS.DPI.TARGET_DPI = 300
_C.PREPROCESS.DPI.MANUAL_DPI = 300
_C.DATALOADER = CN()
_C.DATALOADER.PIN_MEMORY = True
_C.INPUT = CN()
# Run decoding and augmentation on GPU
_C.INPUT.ON_GPU = False
# Input augmentation
# Random flip
_C.INPUT.RANDOM_FLIP = "both"
# Resize options after loading image
_C.INPUT.RESIZE_MODE = "shortest_edge"
_C.INPUT.CROP_RESIZE = False
_C.INPUT.SCALING = 0.0
_C.INPUT.SCALING_TRAIN = 0.5
_C.INPUT.SCALING_TEST = 0.0
# DPI correction in resizing
_C.INPUT.DPI = CN()
_C.INPUT.DPI.AUTO_DETECT_TEST = False
_C.INPUT.DPI.DEFAULT_DPI_TEST = 300
_C.INPUT.DPI.TARGET_DPI_TEST = 300
_C.INPUT.DPI.MANUAL_DPI_TEST = 300
_C.INPUT.DPI.AUTO_DETECT_TRAIN = False
_C.INPUT.DPI.DEFAULT_DPI_TRAIN = 300
_C.INPUT.DPI.TARGET_DPI_TRAIN = 300
_C.INPUT.DPI.MANUAL_DPI_TRAIN = 300
# Convert entire image to grayscale
_C.INPUT.GRAYSCALE = CN()
_C.INPUT.GRAYSCALE.PROBABILITY = 0.0
# Adaptive thresholding to binarize image
_C.INPUT.ADAPTIVE_THRESHOLDING = CN()
_C.INPUT.ADAPTIVE_THRESHOLDING.PROBABILITY = 0.0
# Invert image
_C.INPUT.INVERT = CN()
_C.INPUT.INVERT.PROBABILITY = 0.0
_C.INPUT.INVERT.MAX_VALUE = 255
# Random JPEG compression to deal with compression artifacts
_C.INPUT.JPEG_COMPRESSION = CN()
_C.INPUT.JPEG_COMPRESSION.PROBABILITY = 0.0
_C.INPUT.JPEG_COMPRESSION.MIN_QUALITY = 0
_C.INPUT.JPEG_COMPRESSION.MAX_QUALITY = 100
# Add noise to image
_C.INPUT.NOISE = CN()
_C.INPUT.NOISE.PROBABILITY = 0.0
_C.INPUT.NOISE.MIN_STD = 0
_C.INPUT.NOISE.MAX_STD = 50
# Random brightness
_C.INPUT.BRIGHTNESS = CN()
_C.INPUT.BRIGHTNESS.PROBABILITY = 0.0
_C.INPUT.BRIGHTNESS.MIN_INTENSITY = 0.5
_C.INPUT.BRIGHTNESS.MAX_INTENSITY = 1.5
# Random contrast
_C.INPUT.CONTRAST = CN()
_C.INPUT.CONTRAST.PROBABILITY = 0.0
_C.INPUT.CONTRAST.MIN_INTENSITY = 0.5
_C.INPUT.CONTRAST.MAX_INTENSITY = 1.5
# Random saturation
_C.INPUT.SATURATION = CN()
_C.INPUT.SATURATION.PROBABILITY = 0.0
_C.INPUT.SATURATION.MIN_INTENSITY = 0.5
_C.INPUT.SATURATION.MAX_INTENSITY = 1.5
# Random hue
_C.INPUT.HUE = CN()
_C.INPUT.HUE.PROBABILITY = 0.0
_C.INPUT.HUE.MIN_DELTA = -0.5
_C.INPUT.HUE.MAX_DELTA = 0.5
# Random Gaussian filter (blur)
_C.INPUT.GAUSSIAN_FILTER = CN()
_C.INPUT.GAUSSIAN_FILTER.PROBABILITY = 0.0
_C.INPUT.GAUSSIAN_FILTER.MIN_SIGMA = 0.0
_C.INPUT.GAUSSIAN_FILTER.MAX_SIGMA = 3.0
# Horizontal flip
_C.INPUT.HORIZONTAL_FLIP = CN()
_C.INPUT.HORIZONTAL_FLIP.PROBABILITY = 0.0
# Vertical flip
_C.INPUT.VERTICAL_FLIP = CN()
_C.INPUT.VERTICAL_FLIP.PROBABILITY = 0.0
# Elastic deformation
_C.INPUT.ELASTIC_DEFORMATION = CN()
_C.INPUT.ELASTIC_DEFORMATION.PROBABILITY = 0.0
_C.INPUT.ELASTIC_DEFORMATION.ALPHA = 0.1
_C.INPUT.ELASTIC_DEFORMATION.SIGMA = 0.01
# Affine transformations
_C.INPUT.AFFINE = CN()
_C.INPUT.AFFINE.PROBABILITY = 0.0
_C.INPUT.AFFINE.TRANSLATION = CN()
_C.INPUT.AFFINE.TRANSLATION.PROBABILITY = 0.5
_C.INPUT.AFFINE.TRANSLATION.STANDARD_DEVIATION = 0.02
_C.INPUT.AFFINE.ROTATION = CN()
_C.INPUT.AFFINE.ROTATION.PROBABILITY = 0.5
_C.INPUT.AFFINE.ROTATION.KAPPA = 30.0
_C.INPUT.AFFINE.SHEAR = CN()
_C.INPUT.AFFINE.SHEAR.PROBABILITY = 0.5
_C.INPUT.AFFINE.SHEAR.KAPPA = 20.0
_C.INPUT.AFFINE.SCALE = CN()
_C.INPUT.AFFINE.SCALE.PROBABILITY = 0.5
_C.INPUT.AFFINE.SCALE.STANDARD_DEVIATION = 0.12
# Change orientation (rotate 0, 90, 180, 270 degrees)
_C.INPUT.ORIENTATION = CN()
_C.INPUT.ORIENTATION.PROBABILITY = 1.0
_C.INPUT.ORIENTATION.PERCENTAGES = [1.0, 0.0, 0.0, 0.0]
# Solver
_C.SOLVER = CN()
# weight decay on embedding
_C.SOLVER.WEIGHT_DECAY_EMBED = 0.0
# optimizer
_C.SOLVER.OPTIMIZER = "ADAMW"
_C.SOLVER.BACKBONE_MULTIPLIER = 0.1
_C.SOLVER.AMSGRAD = True
# Added for mask2former
_C.INPUT.DATASET_MAPPER_NAME = "mask_former_semantic"
# Color augmentation
_C.INPUT.COLOR_AUG_SSD = False
# We retry random cropping until no single category in semantic segmentation GT occupies more
# than `SINGLE_CATEGORY_MAX_AREA` part of the crop.
_C.INPUT.CROP = CN()
_C.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA = 1.0
# Pad image and segmentation GT in dataset mapper.
_C.INPUT.SIZE_DIVISIBILITY = -1
# mask_former model config
_C.MODEL.MASK_FORMER = CN()
# loss
_C.MODEL.MASK_FORMER.DEEP_SUPERVISION = True
_C.MODEL.MASK_FORMER.NO_OBJECT_WEIGHT = 0.1
_C.MODEL.MASK_FORMER.CLASS_WEIGHT = 1.0
_C.MODEL.MASK_FORMER.DICE_WEIGHT = 1.0
_C.MODEL.MASK_FORMER.MASK_WEIGHT = 20.0
# transformer config
_C.MODEL.MASK_FORMER.NHEADS = 8
_C.MODEL.MASK_FORMER.DROPOUT = 0.1
_C.MODEL.MASK_FORMER.DIM_FEEDFORWARD = 2048
_C.MODEL.MASK_FORMER.ENC_LAYERS = 0
_C.MODEL.MASK_FORMER.DEC_LAYERS = 6
_C.MODEL.MASK_FORMER.PRE_NORM = False
_C.MODEL.MASK_FORMER.HIDDEN_DIM = 256
_C.MODEL.MASK_FORMER.NUM_OBJECT_QUERIES = 100
_C.MODEL.MASK_FORMER.TRANSFORMER_IN_FEATURE = "res5"
_C.MODEL.MASK_FORMER.ENFORCE_INPUT_PROJ = False
# mask_former inference config
_C.MODEL.MASK_FORMER.TEST = CN()
_C.MODEL.MASK_FORMER.TEST.SEMANTIC_ON = True
_C.MODEL.MASK_FORMER.TEST.INSTANCE_ON = False
_C.MODEL.MASK_FORMER.TEST.PANOPTIC_ON = False
_C.MODEL.MASK_FORMER.TEST.OBJECT_MASK_THRESHOLD = 0.0
_C.MODEL.MASK_FORMER.TEST.OVERLAP_THRESHOLD = 0.0
_C.MODEL.MASK_FORMER.TEST.SEM_SEG_POSTPROCESSING_BEFORE_INFERENCE = False
# Sometimes `backbone.size_divisibility` is set to 0 for some backbone (e.g. ResNet)
# you can use this config to override
_C.MODEL.MASK_FORMER.SIZE_DIVISIBILITY = 32
# pixel decoder config
_C.MODEL.SEM_SEG_HEAD.MASK_DIM = 256
# adding transformer in pixel decoder
_C.MODEL.SEM_SEG_HEAD.TRANSFORMER_ENC_LAYERS = 0
# pixel decoder
_C.MODEL.SEM_SEG_HEAD.PIXEL_DECODER_NAME = "BasePixelDecoder"
# swin transformer backbone
_C.MODEL.SWIN = CN()
_C.MODEL.SWIN.PRETRAIN_IMG_SIZE = 224
_C.MODEL.SWIN.PATCH_SIZE = 4
_C.MODEL.SWIN.EMBED_DIM = 96
_C.MODEL.SWIN.DEPTHS = [2, 2, 6, 2]
_C.MODEL.SWIN.NUM_HEADS = [3, 6, 12, 24]
_C.MODEL.SWIN.WINDOW_SIZE = 7
_C.MODEL.SWIN.MLP_RATIO = 4.0
_C.MODEL.SWIN.QKV_BIAS = True
_C.MODEL.SWIN.QK_SCALE = None
_C.MODEL.SWIN.DROP_RATE = 0.0
_C.MODEL.SWIN.ATTN_DROP_RATE = 0.0
_C.MODEL.SWIN.DROP_PATH_RATE = 0.3
_C.MODEL.SWIN.APE = False
_C.MODEL.SWIN.PATCH_NORM = True
_C.MODEL.SWIN.OUT_FEATURES = ["res2", "res3", "res4", "res5"]
_C.MODEL.SWIN.USE_CHECKPOINT = False
# NOTE: maskformer2 extra configs
# transformer module
_C.MODEL.MASK_FORMER.TRANSFORMER_DECODER_NAME = "MultiScaleMaskedTransformerDecoder"
# LSJ aug
_C.INPUT.IMAGE_SIZE = 1024
_C.INPUT.MIN_SCALE = 0.1
_C.INPUT.MAX_SCALE = 2.0
# MSDeformAttn encoder configs
_C.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES = ["res3", "res4", "res5"]
_C.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_POINTS = 4
_C.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_N_HEADS = 8
# point loss configs
# Number of points sampled during training for a mask point head.
_C.MODEL.MASK_FORMER.TRAIN_NUM_POINTS = 112 * 112
# Oversampling parameter for PointRend point sampling during training. Parameter `k` in the
# original paper.
_C.MODEL.MASK_FORMER.OVERSAMPLE_RATIO = 3.0
# Importance sampling parameter for PointRend point sampling during training. Parametr `beta` in
# the original paper.
_C.MODEL.MASK_FORMER.IMPORTANCE_SAMPLE_RATIO = 0.75
# We retry random cropping until no single category in semantic segmentation GT occupies more
# than `SINGLE_CATEGORY_MAX_AREA` part of the crop.
_C.INPUT.CROP.SINGLE_CATEGORY_MAX_AREA = 1.0
# Used for `poly` learning rate schedule.
_C.SOLVER.POLY_LR_POWER = 0.9
_C.SOLVER.POLY_LR_CONSTANT_ENDING = 0.0
# Loss type, choose from `cross_entropy`, `hard_pixel_mining`.
_C.MODEL.SEM_SEG_HEAD.LOSS_TYPE = "hard_pixel_mining"
# DeepLab settings
_C.MODEL.SEM_SEG_HEAD.PROJECT_FEATURES = ["res2"]
_C.MODEL.SEM_SEG_HEAD.PROJECT_CHANNELS = [48]
_C.MODEL.SEM_SEG_HEAD.ASPP_CHANNELS = 256
_C.MODEL.SEM_SEG_HEAD.ASPP_DILATIONS = [6, 12, 18]
_C.MODEL.SEM_SEG_HEAD.ASPP_DROPOUT = 0.1
_C.MODEL.SEM_SEG_HEAD.USE_DEPTHWISE_SEPARABLE_CONV = False
# Backbone new configs
_C.MODEL.RESNETS = CN()
_C.MODEL.RESNETS.RES4_DILATION = 1
_C.MODEL.RESNETS.RES5_MULTI_GRID = [1, 2, 4]
# ResNet stem type from: `basic`, `deeplab`
_C.MODEL.RESNETS.STEM_TYPE = "deeplab"