configs/BERT_L12_H192_experiments/in1k_training_moe.yaml

_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