File size: 4,707 Bytes
32b542e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
_BASE_: "base_model_bert_l12_h192.yaml"

SHARED_TARGETS:

  - 
    NAME: 'ImageNet1k'
    SHARED_TARGETS_CFG:
      FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
      DISTRIBUTED: True
  
  # -
  #   NAME: 'Vocab_Word'
  #   SHARED_TARGETS_CFG:
  #     FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
  #     DISTRIBUTED: True



TASKS:

  - 
    NAME: imagenet
    DATASETS:
      TRAIN: 'ImageNetDataset'
      VAL: 'ImageNetDataset'
      TASK_TYPE: 'image_classification'
      DATASET_NAME: 'ImageNet1k'
      TARGET_SET: ['ImageNet1k']
      
    DATALOADER:
      TRAIN_BATCH_SIZE: 4
      TEST_BATCH_SIZE: 4
      NUM_WORKERS: 4
      FEATS_FOLDER: 'open_source_dataset/imagenet'
      S3_PATH: 'cluster2:s3://imagenet'
      ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
      SAMPLING_WEIGHT: 1.0
      CLASS_NAME_FILE: 'open_source_dataset/imagenet_class_name.pkl'
      MIXUP: 0.8
      CUTMIX: 1.0
      MIXUP_PROB: 1.0
      MIXUP_SWITCH_PROB: 0.5
      MIXUP_MODE: 'batch'
      MIXUP_LABEL_SMOOTHING: 0.1
    MODEL:
      MAX_SEQ_LEN: -1
      LABELS_NUM: 1000
      TEMP_NAME: logit_scale_img_cls
    LOSSES:
      NAMES: ['SoftTargetCrossEntropy', 'Accuracy']
      LOSS_WEIGHT: 1.0
      REDUCTION: 'mean'
      # LOSS_FP32: True
    INFERENCE:
      NAME: 'ImageNetEvaler'
      ID_KEY: 'image_id'
      VALUE: 'cls_logits'
      VAL_ANNFILE: 'open_source_dataset/imagenet/meta/val.txt'
      TEST_ANNFILE: ''
      GENERATION_MODE: False


ENGINE:
  NAME: 'UnifiedTrainer'
 
MODEL:
  META_ARCHITECTURE: 'MultiTaskTransformerEncoder'
  ENCODER: 'UnifiedBertEncoder'

  IN_TUNING: True # use IN1k instead of 22k
  SHARE_LAYERNORM: True
  BERT:
    NORMALIZE_DECISION: "BERTPre" 
    DROP_PATH_PROB: 0.0
    DROP_PATH_PROB_FIXED: True

    UNIFY_QKV: True
  

  OLD_CHECKPONT: True
  
  MODEL_EMA: False
  MODEL_EMA_DECAY: 0.9999

  MAEParamsInit: True
  POSEMBEDFIX: True


  IMG_INPUT_SIZE: 224
  PATCH_SIZE: 16
  # POSEMBED_SCALE: !!python/object/apply:eval ["160/224"]
  # CHECKPOINT_FILETER: False 

  LAYER_SCALE: True 
  LAYER_SCALE_INIT: 1e-3


DATALOADER:
  USE_WEIGHTED_SAMPLER: True
  UNIFIED_DATASET: True 
  NUM_WORKERS: 16

  PADDING_TO_MAX: False # True for debugging or token moe with distributed moe 


  
####################################### Optimizer #######################################
SOLVER:
  NAME: 'Adam'
  TORCH_OPTIMIZER: True
  PARAMS_SEPERATE: True
  # PARAMS_GROUP: True
  # EPOCH: 1
  MAX_ITER: 150000
  CHECKPOINT_PERIOD: 5000
  EVAL_PERIOD: 500000
  BASE_LR: 0.001
  BIAS_LR_FACTOR: 1.0
  WEIGHT_DECAY: 0.05
  WEIGHT_DECAY_NORM: 0.0
  WEIGHT_DECAY_BIAS: 0.0
  WEIGHT_DECAY_EMBEDDING: 0.0
  MOMENTUM: 0.9
  DAMPENING: 0.0
  NESTEROV: 0.0
  BETAS: [0.9, 0.95]
  EPS: 1e-6
  GRAD_CLIP: 0.1
  GRAD_CLIP_TYPE: 'norm'
  ACCUM_ITER: 0
  AMP_FP16: True
  APEX_FP16: False # dangerous

  WRITE_PERIOD: 50
  MIN_LOSS_SCLE: 2048.0
  # BF16: False # True
  # ZEROSTAGE: 2

  LOSS_SCALE_WINDOW: 200





  
####################################### lr scheduler ####################################### 
LR_SCHEDULER:
  NAME: 'WarmupCosine'
  WARMUP: 5000
  MIN_LR: 0.000001




####################################### evaluation ####################################### 
INFERENCE:

  VOCAB: 'CLIP'
  ITER_BASED: True


find_unused_parameters: true

# ENCODERS:
#   -
#     NAME: VisualEncoder
#     TYPE: VisualEncoder
#     DROP_PATH_PROB: 0.0
#     HIDDEN_SIZE: 192
#     HIDDEN_DROPOUT_PROB: 0.
#     HIDDEN_ACT: "gelu"
#     NUM_ATTENTION_HEADS: 3
#     INTERMEDIATE_SIZE: 768
#     INTERMEDIATE_DROP: 0.
#     FFN_DROPOUT_PROB: 0.
#     ATTENTION_PROBS_DROPOUT_PROB: 0.
#     NUM_HIDDEN_LAYERS: 6
#     NUM_GENERATION_LAYERS: 0
#     DROP_PATH_PROB_FIXED: True

#   -
#     NAME: TextEncoder
#     TYPE: TextEncoder
#     DROP_PATH_PROB: 0.0
#     HIDDEN_SIZE: 192
#     HIDDEN_DROPOUT_PROB: 0.
#     HIDDEN_ACT: "gelu"
#     NUM_ATTENTION_HEADS: 3
#     INTERMEDIATE_SIZE: 768
#     INTERMEDIATE_DROP: 0.
#     FFN_DROPOUT_PROB: 0.
#     ATTENTION_PROBS_DROPOUT_PROB: 0.
#     NUM_HIDDEN_LAYERS: 6
#     NUM_GENERATION_LAYERS: 0
#     DROP_PATH_PROB_FIXED: True 

MOE: 
  MOE: True 
  MOE_TYPE: 'attribute'
  TAG_Transform: True
  ATTRIBUTE_LENGTH: 8
  EP_WORLD_SIZE: 1 # tag moe only 
  NUM_EXPERTS: 8
  TOP_K: 2
  CAPACITY_FACTOR: 3.0 
  EVAL_MIN_CAPACITY: 4.0
  MIN_CAPACITY: 4
  NOISY_GATE_POLICY: 'vmoe'
  MOE_PARAM_GROUP: True 
  MOE_EXPERT_TYPE: 'FFN,SA'
  SA_LINEAR_OUT_MOE: True
  MOE_EXPERT_LOCATION: 'all' # 'odd'
  # MOE_LAYER_START_IDX: 3
  # MOE_LAYER_END_IDX: 21
  # MOE_LAYER_START_IDX: 18
  # MOE_LAYER_END_IDX: 12 
  BATCH_PRIO: True 
  USE_TUTEL: True
  FFN_SHARE_GATE_DECISION: True