zeroshot_v06.yaml

MODEL:
  META_ARCHITECTURE: "GeneralizedRCNN"
  # Initial weight to load from image-caption training
  WEIGHT: "/home/alireza/workspace/ovo/runs-new/maskrcnn/130-rep/model_0120000.pth"
  # Trim the prefix of the checkpoint parameter names so they can be correctly loaded
  BACKBONE_PREFIX: "backbone.body."
  # Whether to load the vl_projection layer from the multimedia self-supervised learning head
  # If true, it loads it from the default mmss head defined by MODEL.MMSS_HEAD.DEFAULT_HEAD
  LOAD_EMB_PRED_FROM_MMSS_HEAD: True
  # Set true when resuming training. Otherwise should be False to prevent loading trainer
  # state from pretraining phase.
  LOAD_TRAINER_STATE: False
  # Always true for zero-shot settings, although it is false for regular Faster R-CNN
  # If false, it learns a bounding box regression for each (seen) class separately
  CLS_AGNOSTIC_BBOX_REG: True
  ROI_BOX_HEAD:
    # Note these are the number of classes for training only
    NUM_CLASSES: 49
    # Dimension of embeddings that will be loaded (300 for Glove, 768 for Bert)
    EMB_DIM: 768
    # Always true for zero-shot
    EMBEDDING_BASED: True
    # To balance background proposals vs. foreground. Especially important to tune for 
    # zero-shot settings, because a value too large would push unseen classes to background.
    LOSS_WEIGHT_BACKGROUND: 0.2
    # Whether or not to freeze the vl_projection layer. True is better. Only works if
    # MODEL.LOAD_EMB_PRED_FROM_MMSS_HEAD is true
    FREEZE_EMB_PRED: True
  ROI_HEADS:
    # At most how much of a batch should be filled with positive boxes. In zero-shot setting
    # having too many background hurts. Note 1.0 doesn't mean there won't be any background.
    # It is unlikely to have 512 positive boxes, and the rest is always filled with background.
    POSITIVE_FRACTION: 1.0
  BACKBONE:
    FREEZE_CONV_BODY_AT: 2
DATASETS:
  TRAIN: ("coco_zeroshot_train",)
  TEST: ("coco_not_zeroshot_val", "coco_zeroshot_val", "coco_generalized_zeroshot_val")
  DATASET_CLASS: "COCODataset"
  DATASET_ARGS: 
    # load embeddings from the annotation json. Note it doesn't work with original COCO json.
    # First run ipynb/003.ipynb or ipynb/004.ipynb to add embeddings to annotations
    LOAD_EMBEDDINGS: True
    # The key for embedding to load. We have BertEmb and GloVE for now.
    EMB_KEY: "BertEmb"
    # Dimension of embeddings (300 for Glove, 768 for Bert)
    EMB_DIM: 768
SOLVER:
  BASE_LR: 0.005
  WEIGHT_DECAY: 0.0001
  STEPS: (60000, 120000)
  MAX_ITER: 150000
  IMS_PER_BATCH: 8
  CHECKPOINT_PERIOD: 10000
  TEST_PERIOD: 10000
  LOG_PERIOD: 100
TEST:
  IMS_PER_BATCH: 8