configs/BERT_L12_H768_experiments/bw_mlm_training.yaml

_BASE_: "base_model_bert_l12_h768.yaml"

SHARED_TARGETS:

  # - 
  #   NAME: 'ImageNet1k'
  #   SHARED_TARGETS_CFG:
  #     FILE_PATH: 'open_source_dataset/imagenet_class_name_CLIP_with_endoftext.pkl'
  #     DISTRIBUTED: False

  -
    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: 256
  #     # TEST_BATCH_SIZE: 2
  #     NUM_WORKERS: 4
  #     FEATS_FOLDER: '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

  -
    NAME: bookswiki_pretrain
    DATASETS:
      TRAIN: 'GeneralCorpusDataset'
      TASK_TYPE: 'text_mlm'
      DATASET_NAME: 'BooksWiki'
      TARGET_SET: ['Vocab_Word']
    DATALOADER:
      TRAIN_BATCH_SIZE: 256
      TEST_BATCH_SIZE: 32
      NUM_WORKERS: 2
      ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
      # ANNO_FOLDER:  'open_source_dataset/bert_pretrain_data/bookswiki'
      SEQ_PER_SAMPLE:  1
      SAMPLER: NodeDistributed
      CACHE_MODE: True
      SEQ_PER_SAMPLE: 128
      MIN_SEQ_PER_SAMPLE: 128
      APPEND_EOS: True
      ONE_STREAM: False
      SAMPLING_WEIGHT: 1.0
      RANDOM_MASK: True
    MODEL:
      MAX_SEQ_LEN: 128
      TEMP_NAME: logit_scale_text_mlm
    LOSSES:
      NAMES: ['CrossEntropy', 'Accuracy']
      LOSS_WEIGHT: 0.5
      REDUCTION: 'mean'
    INFERENCE:
      VOCAB: 'CLIP'
      GENERATION_MODE: False

  # -
  #   NAME: mscoco_caption
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     # VAL: 'ImageTextPairDataset'
  #     # TEST: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_caption'
  #     DATASET_NAME: 'MSCOCO'
  #     TARGET_SET: ['Vocab_Word']
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 200
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 4
  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
  #     S3_PATH: 's3://coco/'
  #     SEQ_PER_SAMPLE:  1
      # CACHE_MODE: True
      # CIRCULAR_CACHE_MODE: False
      # ZIP_MODE: False
      # CACHE_ORIGIN_IMAGE: False
      # RANDOM_CAPTION: False
      # AS_NUMPY_AS_POSSIBLE: False
  #     SAMPLING_WEIGHT: 0.5
  #     TRANSFORM: 'clip_transforms'
  #     RANDOM_MASK: True
  #   MODEL:
  #     MAX_SEQ_LEN: 30
  #     EVAL_MAX_SEQ_LEN: 21
  #     TEMP_NAME: logit_scale_caption
  #   LOSSES:
  #     NAMES: ['CrossEntropy', 'Accuracy']
  #     LOSS_WEIGHT: 0.5
  #     REDUCTION: 'mean'
  #   DECODE_STRATEGY:
  #     NAME: 'CaptionBeamSearcherV3'
  #     BEAM_SIZE: 2
  #     # LEN_PENALTY: 1.0
  #   INFERENCE:
  #     NAME: 'COCOEvaler'
  #     VOCAB: 'CLIP'
  #     ID_KEY: 'image_id'
  #     VALUE: 'caption'
  #     VAL_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_val5k.json'
  #     TEST_ANNFILE: 'open_source_dataset/mscoco_dataset/new_annotations/captions_test5k.json'
  #     GENERATION_MODE: True

  # -
  #   NAME: mscoco_retrieve
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     # TEST: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_retrieval'
  #     DATASET_NAME: 'MSCOCO'
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 256
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 1
  #     FEATS_FOLDER: 'open_source_dataset/mscoco_dataset/coco_origin'
  #     ANNO_FOLDER:  'open_source_dataset/mscoco_dataset/new_annotations'
  #     S3_PATH: 's3://coco/'
      # CACHE_MODE: True
      # CIRCULAR_CACHE_MODE: False
      # ZIP_MODE: False
      # CACHE_ORIGIN_IMAGE: False
      # RANDOM_CAPTION: False
      # AS_NUMPY_AS_POSSIBLE: False
  #     SEQ_PER_SAMPLE:  1
  #     SAMPLING_WEIGHT: 0.5
  #     TRANSFORM: 'clip_transforms'
  #   MODEL:
  #     MAX_SEQ_LEN: 30
  #     TEMP_NAME: logit_scale_retrieve
  #   LOSSES:
  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
  #     LABELSMOOTHING: 0.1
  #     LOSS_WEIGHT: 0.5
  #     REDUCTION: 'mean'
  #   INFERENCE:
  #     VOCAB: 'CLIP'
  #     ID_KEY: 'image_id'
  #     VALUE: 'caption'
  #     NAME: 'RetrievalEvaler'
  #     VAL_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_val_set0_2014.jsonline'
  #     TEST_ANNFILE: 'open_source_dataset/flickr30k/all_data_final_test_set0_2014.jsonline'
  #     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: 0.1
    DROP_PATH_PROB_FIXED: True

    UNIFY_QKV: True
  
  MODEL_EMA: False
  MODEL_EMA_DECAY: 0.9999

  MAEParamsInit: True
  POSEMBEDFIX: True


  IMG_INPUT_SIZE: 224
  PATCH_SIZE: 16

  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: 450000
  CHECKPOINT_PERIOD: 50000
  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: 20000
  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