File size: 7,774 Bytes
814a594
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --------------------------------------------------------
# X-Decoder -- Generalized Decoding for Pixel, Image, and Language
# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Xueyan Zou ([email protected])
# --------------------------------------------------------

# Define Test/Trainer/Saving
PIPELINE: XDecoderPipeline
TRAINER: xdecoder
SAVE_DIR: './output'
base_path: "./"

# Resume Logistic
RESUME: false
WEIGHT: false
RESUME_FROM: ''
EVAL_AT_START: false
SAVE_CHECKPOINT: True

# Logging and Debug
WANDB: False
LOG_EVERY: 100
FIND_UNUSED_PARAMETERS: false

# Speed up training
FP16: false
PORT: '36873'

# misc
LOADER:
  JOINT: True
  KEY_DATASET: ""
  SAMPLE_PROB: "prop"    # sampling probability proportional to data size. Use "equal" for each bach from all datasets
  MIXING_LEVEL: 1    # num of different datasets for batch mixing on each GPU

RANDOM_SEED: 2024

STANDARD_TEXT_FOR_EVAL: False

##################
# Task settings
##################
VERBOSE: true
MODEL:
  DEVICE: "cuda"  # or "cpu" if no GPU available
  NAME: seem_model_v1
  HEAD: xdecoder_head
  MASK_ON: false
  KEYPOINT_ON: false
  LOAD_PROPOSALS: false
  DIM_PROJ: 512
  TEXT:
    ARCH: vlpencoder
    NAME: transformer
    TOKENIZER: clip
    CONTEXT_LENGTH: 77 #256 # 77
    WIDTH: 512 # 768  # 512
    HEADS: 8
    LAYERS: 12 # 6
    AUTOGRESSIVE: True
  BACKBONE:
    NAME: focal    # focal_dw    # focal
    PRETRAINED: ''
    LOAD_PRETRAINED: false
    FOCAL:
      PRETRAIN_IMG_SIZE: 224
      PATCH_SIZE: 4
      EMBED_DIM: 192    # 96    # 192
      DEPTHS: [2, 2, 18, 2]    # [2, 2, 6, 2]    # [2, 2, 18, 2]
      FOCAL_LEVELS: [4, 4, 4, 4]    # [3, 3, 3, 3]    # [4, 4, 4, 4]
      FOCAL_WINDOWS: [3, 3, 3, 3]
      DROP_PATH_RATE: 0.3
      MLP_RATIO: 4.0
      DROP_RATE: 0.0
      PATCH_NORM: True
      USE_CONV_EMBED: True
      SCALING_MODULATOR: True
      USE_CHECKPOINT: False
      USE_POSTLN: true
      USE_POSTLN_IN_MODULATION: false
      USE_LAYERSCALE: True
      OUT_FEATURES: ["res2", "res3", "res4", "res5"]
      OUT_INDICES: [0, 1, 2, 3]
  ENCODER:
    NAME: transformer_encoder_fpn
    IGNORE_VALUE: 255
    NUM_CLASSES: 16
    BINARY_CLASSES: False
    LOSS_WEIGHT: 1.0
    CONVS_DIM: 512
    MASK_DIM: 512
    NORM: "GN"
    IN_FEATURES: ["res2", "res3", "res4", "res5"]
    DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES: ["res3", "res4", "res5"]
    COMMON_STRIDE: 4
    TRANSFORMER_ENC_LAYERS: 6
  DECODER:
    NAME: seem_v1
    TRANSFORMER_IN_FEATURE: "multi_scale_pixel_decoder"
    MASK:
      ENABLED: True
    DETECTION: False
    SPATIAL:
      ENABLED: True
      MAX_ITER: 1
    GROUNDING:
      ENABLED: True
      MAX_LEN: 10
      TEXT_WEIGHT: 2.0
      CLASS_WEIGHT: 0.5
    RETRIEVAL:
      ENABLED: False
    LVIS:
      ENABLED: False
      THRES: 0.7
    OPENIMAGE:
      ENABLED: False
      NEGATIVE_SAMPLES: 5
      GROUNDING:
        ENABLED: False
        MAX_LEN: 5
    CAPTION:
      ENABLED: False
      PHRASE_PROB: 0.5
      SIM_THRES: 0.95
    DEEP_SUPERVISION: True
    NO_OBJECT_WEIGHT: 0.1
    GCLASS_WEIGHT: 0.4
    GMASK_WEIGHT: 1.0
    GDICE_WEIGHT: 1.0
    SCLASS_WEIGHT: 0.4
    SMASK_WEIGHT: 1.0
    SDICE_WEIGHT: 1.0
    OCLASS_WEIGHT: 0.4
    OMASK_WEIGHT: 1.0
    ODICE_WEIGHT: 1.0
    CLASS_WEIGHT: 2.0
    MASK_WEIGHT: 5.0
    DICE_WEIGHT: 5.0
    BBOX_WEIGHT: 5.0
    GIOU_WEIGHT: 2.0
    CAPTION_WEIGHT: 2.0
    COST_SPATIAL:
      CLASS_WEIGHT: 5.0
      MASK_WEIGHT: 2.0
      DICE_WEIGHT: 2.0
    HIDDEN_DIM: 512
    NUM_OBJECT_QUERIES: 101
    NHEADS: 8
    DROPOUT: 0.0
    DIM_FEEDFORWARD: 2048
    MAX_SPATIAL_LEN: [512, 512, 512, 512]
    # ENC_LAYERS: 0
    PRE_NORM: False
    ENFORCE_INPUT_PROJ: False
    SIZE_DIVISIBILITY: 32
    TRAIN_NUM_POINTS: 12544
    OVERSAMPLE_RATIO: 3.0
    IMPORTANCE_SAMPLE_RATIO: 0.75
    DEC_LAYERS: 10  # 9 decoder layers, add one for the loss on learnable query
    TOP_GROUNDING_LAYERS: 10
    TOP_CAPTION_LAYERS: 10
    TOP_SPATIAL_LAYERS: 10
    TOP_OPENIMAGE_LAYERS: 10
    TEST:
      SEMANTIC_ON: False
      INSTANCE_ON: False
      PANOPTIC_ON: False
      OVERLAP_THRESHOLD: 0.8
      OBJECT_MASK_THRESHOLD: 0.8
      SEM_SEG_POSTPROCESSING_BEFORE_INFERENCE: true

# Spatial sampler
STROKE_SAMPLER:
  MAX_CANDIDATE: 1
  CANDIDATE_PROBS: [0.25, 0.25, 0.25, 0.25] # for training only
  CANDIDATE_NAMES: ["Point", "Polygon", "Scribble", "Circle"]
  DILATION: 3
  CIRCLE:
    NUM_STROKES: 5
    STROKE_PRESET: ['object_like', 'object_like_middle', 'object_like_small']
    STROKE_PROB: [0.33, 0.33, 0.33]
  SCRIBBLE:
    NUM_STROKES: 5
    STROKE_PRESET: ['rand_curve', 'rand_curve_small']
    STROKE_PROB: [0.5, 0.5]
  POINT:
    NUM_POINTS: 20
  POLYGON:
    MAX_POINTS: 9
  EVAL:
    MODE: 'best' # best/random/best_random
    NEGATIVE: False
    MAX_ITER: 1
    IOU_ITER: 1
    GROUNDING: True

# Multi-modal Architecture, order matters
ATTENTION_ARCH:
  VARIABLE:
    queries: ['object', 'grounding', 'spatial']
    tokens: ['grounding', 'spatial']
    memories: ['spatial']
  SELF_ATTENTION:
    queries:
      object: ['queries_object']
      grounding: ['queries_grounding', 'tokens_grounding']
      spatial: ['queries_spatial', 'tokens_spatial', 'memories_spatial']
    tokens:
      grounding: ['queries_grounding', 'tokens_grounding']
      spatial: ['tokens_spatial']
    memories:
      spatial: ['memories_spatial']
  CROSS_ATTENTION:
    queries:
      object: True
      grounding: True
      spatial: True
    memories:
      spatial: True
    tokens:
      grounding: False
      spatial: False
  MASKING: ['tokens_spatial', 'tokens_grounding']
  DUPLICATION:
    queries:
      grounding: 'queries_object'
      spatial: 'queries_object'
  SPATIAL_MEMORIES: 32
  QUERY_NUMBER: 3

DATASETS:
  TRAIN: [
  'biomed_BiomedParseData-Demo_demo'    # Add your registered training datasets here
  ]



  TEST:  [
  'biomed_BiomedParseData-Demo_demo'   # Add your registered test datasets here
  ]
  
  CLASS_CONCAT: false
  SIZE_DIVISIBILITY: 32
  PROPOSAL_FILES_TRAIN: []

INPUT:
  PIXEL_MEAN: [123.675, 116.280, 103.530]
  PIXEL_STD: [58.395, 57.120, 57.375]

TRAIN:
  ASPECT_RATIO_GROUPING: true
  BATCH_SIZE_TOTAL: 4
  BATCH_SIZE_PER_GPU: 4
  SHUFFLE: true

TEST:
  DETECTIONS_PER_IMAGE: 100
  NAME: coco_eval
  IOU_TYPE: ['bbox', 'segm']
  USE_MULTISCALE: false
  BATCH_SIZE_TOTAL: 4
  MODEL_FILE: ''
  AUG:
    ENABLED: False

DATALOADER:
  FILTER_EMPTY_ANNOTATIONS: False
  NUM_WORKERS: 8
  LOAD_PROPOSALS: False
  SAMPLER_TRAIN: "TrainingSampler"
  ASPECT_RATIO_GROUPING: True


BioMed:
  INPUT:
    PIXEL_MEAN: [64.284, 59.293, 59.962]
    PIXEL_STD: [62.484, 60.865, 59.835]
    DATASET_MAPPER_NAME: "biomed_interactive"
    MIN_SIZE_TRAIN: 900
    MAX_SIZE_TRAIN: 1100
    MIN_SIZE_TRAIN_SAMPLING: 'choice'
    MIN_SIZE_TEST: 900
    MAX_SIZE_TEST: 1100
    IMAGE_SIZE: 1024
    MIN_SCALE: 0.9
    MAX_SCALE: 1.1
    IGNORE_VALUE: 255
    COLOR_AUG_SSD: False
    SIZE_DIVISIBILITY: 32
    RANDOM_FLIP: "none"
    RANDOM_ROTATE: False
    MASK_FORMAT: "polygon"
    MIN_AREA: 30
    FORMAT: "RGB"
    SPATIAL: True
    CROP:
      ENABLED: True
  DATASET: 
    DATASET: "biomed"


# Detectron2 training config for optimizer and lr scheduler
SOLVER:
  BASE_LR: 0.0001
  STEPS: [0.88889, 0.96296]
  MAX_ITER: 1
  GAMMA: 0.1
  WARMUP_FACTOR: 1.0
  WARMUP_ITERS: 10
  WARMUP_METHOD: "linear"
  WEIGHT_DECAY: 0.05
  OPTIMIZER: "ADAMW"
  LR_SCHEDULER_NAME: "WarmupMultiStepLR"
  LR_MULTIPLIER:
    backbone: 0.1
    lang_encoder: 0.1
  FIX_PARAM:
    backbone: True
    lang_encoder: True
    pixel_decoder: True
  WEIGHT_DECAY_NORM: 0.0
  WEIGHT_DECAY_EMBED: 0.0
  CLIP_GRADIENTS:
    ENABLED: True
    CLIP_TYPE: "full_model"
    CLIP_VALUE: 5.0 # 0.01
    NORM_TYPE: 2.0
  MAX_NUM_EPOCHS: 50