_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: False
  
  -
    NAME: 'Vocab_Word'
    SHARED_TARGETS_CFG:
      FILE_PATH: 'open_source_dataset/vocabulary_CLIP_with_endoftext.pkl'
      DISTRIBUTED: True

  # - 
  #   NAME: 'Kinetics400'
  #   SHARED_TARGETS_CFG:
  #     FILE_PATH: 'open_source_dataset/k400_class_name_CLIP_with_endoftext.pkl'
  #     DISTRIBUTED: False



TASKS:

  # - 
  #   NAME: imagenet
  #   DATASETS:
  #     TRAIN: 'ImageNetDataset'
  #     VAL: 'ImageNetDataset'
  #     TASK_TYPE: 'image_classification'
  #     DATASET_NAME: 'ImageNet1k'
  #     TARGET_SET: ['ImageNet1k']
      
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 720
  #     # TEST_BATCH_SIZE: 2
  #     NUM_WORKERS: 4
  #     FEATS_FOLDER: 'cluster2:s3://imagenet'
  #     ANNO_FOLDER:  'open_source_dataset/imagenet/meta'
  #     SAMPLING_WEIGHT: 2.5
  #     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: K400_retrieve
  #   DATASETS:
  #     TRAIN: 'VideoDataSet'
  #     VAL: 'VideoDataSet'
  #     TASK_TYPE: 'video_classification'
  #     DATASET_NAME: 'K400'
  #     TARGET_SET: ['Kinetics400']
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 12 # 256
  #     TEST_BATCH_SIZE: 4 # debug
  #     NUM_WORKERS: 4 # debug 4
  #     FEATS_FOLDER: 'open_source_dataset/K400_official'
  #     ANNO_FOLDER:  'open_source_dataset/K400_official'
  #     S3_PATH: 's3://K400/'
  #     FRAMES_PER_CLIP: 8
  #     STRIDE: 32
  #     FILE_EXTENSION: ''
  #     ANNO_FILE: 'annotation.json'
  #     TIMESFORMER_AUG: True
  #     SAMPLING_WEIGHT: 1.0
  #   MODEL:
  #     MAX_SEQ_LEN: -1
  #     TEMP_NAME: logit_scale_video_cls
  #   LOSSES:
  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
  #     LABELSMOOTHING: 0.1
  #     LOSS_WEIGHT: 1.0
  #   INFERENCE:
  #     NAME: 'MiTEvaler'
  #     ID_KEY: 'video_name'
  #     VALUE: 'label'
  #     VAL_ANNFILE: 'open_source_dataset/K400_official/annotation.json'
  #     TEST_ANNFILE: ''
  #     GENERATION_MODE: False
  #     NUM_VIEWS: 1
  
  # -
  #   NAME: bookswiki_pretrain
  #   DATASETS:
  #     TRAIN: 'GeneralCorpusDataset'
  #     TASK_TYPE: 'text_mlm'
  #     DATASET_NAME: 'BooksWiki'
  #     TARGET_SET: ['Vocab_Word']
  #     VERSION: 'v2'
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 512
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 2
  #     ANNO_FOLDER:  'open_source_dataset/text_corpus' # '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: 3.5
  #     RANDOM_MASK: True
  #   MODEL:
  #     MAX_SEQ_LEN: 128
  #     TEMP_NAME: logit_scale_text_mlm
  #   LOSSES:
  #     NAMES: ['CrossEntropy', 'Accuracy']
  #     LOSS_WEIGHT: 0.33333
  #     REDUCTION: 'mean'
  #   INFERENCE:
  #     VOCAB: 'CLIP'
  #     GENERATION_MODE: False
  # -
  #   NAME: mscoco_retrieve
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     TEST: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_retrieval'
  #     DATASET_NAME: 'MSCOCO'
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 100
  #     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/'
  #     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: 1.0
  #     TRANSFORM: 'clip_transforms'
  #   MODEL:
  #     MAX_SEQ_LEN: 50
  #     TEMP_NAME: logit_scale_retrieve
  #   LOSSES:
  #     NAMES: ['LabelSmoothingCrossEntropy', 'Accuracy']
  #     LABELSMOOTHING: 0.1
  #     LOSS_WEIGHT: 1.0
  #     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

 ########## Image Captioning ########### 


  # -
  #   NAME: cc12m_caption
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_caption'
  #     DATASET_NAME: 'CC12M'
  #     TARGET_SET: ['Vocab_Word']
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 300
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 2
  #     S3_ANNO_FOLDER:  's3://cc12m/'
      # ANNO_FOLDER:  'open_source_dataset/c12m/'
  #     ANNO_FILENAME: 'train_available.json'
  #     FEATS_FOLDER: 'open_source_dataset/c12m/'
  #     S3_PATH: 's3://cc12m/'
  #     SEQ_PER_SAMPLE:  1
  #     SAMPLER: NodeDistributed
  #     CACHE_MODE: True
  #     CIRCULAR_CACHE_MODE: False
  #     ZIP_MODE: False
  #     CACHE_ORIGIN_IMAGE: False
  #     RANDOM_CAPTION: False
  #     AS_NUMPY_AS_POSSIBLE: False
  #     SAMPLING_WEIGHT: 1.6889
  #     TRANSFORM: 'clip_transforms'
  #   MODEL:
  #     MAX_SEQ_LEN: 50
  #     TEMP_NAME: logit_scale_caption
  #   LOSSES:
  #     NAMES: ['CrossEntropy', 'Accuracy']
  #     LOSS_WEIGHT: 0.33333
  #     REDUCTION: 'mean'
  #   INFERENCE:
  #     VOCAB: 'CLIP'
  #     GENERATION_MODE: False

  # -
  #   NAME: cc3m_caption
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_caption'
  #     DATASET_NAME: 'CC3M'
  #     TARGET_SET: ['Vocab_Word']
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 300
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 2
  #     ANNO_FOLDER:  's3://cc3m/'
  #     ANNO_FILENAME: 'train_spacy.json'
  #     FEATS_FOLDER: 'open_source_dataset/cc3m/'
  #     S3_PATH: 's3://cc3m/'
  #     SEQ_PER_SAMPLE:  1
  #     SAMPLER: NodeDistributed
  #     CACHE_MODE: True
  #     CIRCULAR_CACHE_MODE: False
  #     ZIP_MODE: False
  #     CACHE_ORIGIN_IMAGE: False
  #     RANDOM_CAPTION: False
  #     AS_NUMPY_AS_POSSIBLE: False
  #     SAMPLING_WEIGHT: 0.8780
  #     TRANSFORM: 'clip_transforms'
  #   MODEL:
  #     MAX_SEQ_LEN: 50
  #     TEMP_NAME: logit_scale_caption
  #   LOSSES:
  #     NAMES: ['CrossEntropy', 'Accuracy']
  #     LOSS_WEIGHT: 0.33333
  #     REDUCTION: 'mean'
  #   INFERENCE:
  #     VOCAB: 'CLIP'
  #     GENERATION_MODE: False

  # -
  #   NAME: vg_caption
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_caption'
  #     DATASET_NAME: 'VG'
  #     TARGET_SET: ['Vocab_Word']
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 300
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 2
  #     FEATS_FOLDER: 'open_source_dataset/visual_genome/images'
  #     ANNO_FOLDER:  'open_source_dataset/visual_genome/annotations'
  #     S3_PATH: 's3://visual_genome/images'
  #     ANNO_FILENAME: 'vg_captions_128filter.json'
  #     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: 0.5895
  #     TRANSFORM: 'clip_transforms'
  #   MODEL:
  #     MAX_SEQ_LEN: 30
  #     TEMP_NAME: logit_scale_caption
  #   LOSSES:
  #     NAMES: ['CrossEntropy', 'Accuracy']
  #     LOSS_WEIGHT: 0.33333
  #     REDUCTION: 'mean'
  #   INFERENCE:
  #     VOCAB: 'CLIP'
  #     GENERATION_MODE: True
  
  -
    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: 32
      TEST_BATCH_SIZE: 2
      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: 0.3817
      TRANSFORM: 'clip_transforms'
      RANDOM_MASK: True
    MODEL:
      MAX_SEQ_LEN: 50
      EVAL_MAX_SEQ_LEN: 21
      TEMP_NAME: logit_scale_caption
    LOSSES:
      NAMES: ['CrossEntropy', 'Accuracy']
      LOSS_WEIGHT: 0.33333
      REDUCTION: 'mean'
    DECODE_STRATEGY:
      NAME: 'CaptionBeamSearcherV3'
      BEAM_SIZE: 2
      # LEN_PENALTY: 2.0
    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: sbu_caption
  #   DATASETS:
  #     TRAIN: 'ImageTextPairDataset'
  #     TASK_TYPE: 'image_caption'
  #     DATASET_NAME: 'SBU'
  #     TARGET_SET: ['Vocab_Word']
  #   DATALOADER:
  #     TRAIN_BATCH_SIZE: 300
  #     TEST_BATCH_SIZE: 32
  #     NUM_WORKERS: 1
  #     S3_ANNO_FOLDER: 's3://SBU/annotations'
      # ANNO_FOLDER: 'open_source_dataset/sbucaption/annotations'
  #     ANNO_FILENAME: 'subcaption.json'
  #     FEATS_FOLDER: 'open_source_dataset/sbucaption/'
  #     S3_PATH: 's3://SBU/images'
  #     SEQ_PER_SAMPLE:  1
  #     SAMPLER: NodeDistributed
  #     CACHE_MODE: True
  #     CIRCULAR_CACHE_MODE: False
  #     ZIP_MODE: False
  #     CACHE_ORIGIN_IMAGE: False
  #     RANDOM_CAPTION: False
  #     AS_NUMPY_AS_POSSIBLE: False
  #     SAMPLING_WEIGHT: 0.4618
  #     TRANSFORM: 'clip_transforms'
  #   MODEL:
  #     MAX_SEQ_LEN: 50
  #     TEMP_NAME: logit_scale_caption
  #   LOSSES:
  #     NAMES: ['CrossEntropy', 'Accuracy']
  #     LOSS_WEIGHT: 0.33333
  #     REDUCTION: 'mean'
  #   INFERENCE:
  #     VOCAB: 'CLIP'
  #     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

  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


  LAYER_SCALE_FP32: True
  GATE_FP32: False
  TAG_TRANSFORM_FP32: False


DATALOADER:
  USE_WEIGHTED_SAMPLER: True
  UNIFIED_DATASET: True 
  NUM_WORKERS: 32
  STRATEGY: 'turn'

  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



  FORCE_SOFTMAX_FP16: True
  FORCE_LN_FP16: True
  FORCE_NORM_FP16: True
  # FORCE_TEMP_FP16: True
  FORCE_EMBED_FP16: True





  
####################################### 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