eval_resnet_8gpu_transfer_openimages_linear.yaml

# @package _global_
config:
  VERBOSE: True
  LOG_FREQUENCY: 10
  TEST_ONLY: False
  TEST_EVERY_NUM_EPOCH: 1
  TEST_MODEL: True
  SEED_VALUE: 1
  MULTI_PROCESSING_METHOD: forkserver
  HOOKS:
    PERF_STATS:
      MONITOR_PERF_STATS: True
  DATA:
    NUM_DATALOADER_WORKERS: 2
    TRAIN:
      DATA_SOURCES: [everstore]
      LABEL_SOURCES: [disk_filelist]
      DATASET_NAMES: [openimages_v6]
      BATCHSIZE_PER_REPLICA: 16
      TRANSFORMS:
        - name: RandomResizedCrop
          size: 224
        - name: RandomHorizontalFlip
        - name: ToTensor
        - name: Normalize
          mean: [0.485, 0.456, 0.406]
          std: [0.229, 0.224, 0.225]
      MMAP_MODE: True
      COPY_TO_LOCAL_DISK: False
      COPY_DESTINATION_DIR: /tmp/openimages_v6/
      COLLATE_FUNCTION: targets_one_hot_default_collator
      COLLATE_FUNCTION_PARAMS: {"num_classes": 9605}
    TEST:
      DATA_SOURCES: [everstore]
      LABEL_SOURCES: [disk_filelist]
      DATASET_NAMES: [openimages_v6]
      BATCHSIZE_PER_REPLICA: 16
      TRANSFORMS:
        - name: Resize
          size: 256
        - name: CenterCrop
          size: 224
        - name: ToTensor
        - name: Normalize
          mean: [0.485, 0.456, 0.406]
          std: [0.229, 0.224, 0.225]
      MMAP_MODE: True
      COPY_TO_LOCAL_DISK: False
      COPY_DESTINATION_DIR: /tmp/openimages_v6/
      COLLATE_FUNCTION: targets_one_hot_default_collator
      COLLATE_FUNCTION_PARAMS: {"num_classes": 9605}
  METERS:
    name: mean_ap_list_meter
    # training data size is 5.6M , we just disable the training meter
    enable_training_meter: False
    mean_ap_list_meter:
      num_classes: 9605
      num_meters: 1
  TRAINER:
    TRAIN_STEP_NAME: standard_train_step
  MODEL:
    FEATURE_EVAL_SETTINGS:
      EVAL_MODE_ON: True
      FREEZE_TRUNK_ONLY: True
      SHOULD_FLATTEN_FEATS: False
      LINEAR_EVAL_FEAT_POOL_OPS_MAP: [
        # ["conv1", ["AvgPool2d", [[10, 10], 10, 4]]],
        # ["res2", ["AvgPool2d", [[16, 16], 8, 0]]],
        # ["res3", ["AvgPool2d", [[13, 13], 5, 0]]],
        # ["res4", ["AvgPool2d", [[8, 8], 3, 0]]],
        # ["res5", ["AvgPool2d", [[6, 6], 1, 0]]],
        ["res5avg", ["Identity", []]],
      ]
    TRUNK:
      NAME: resnet
      TRUNK_PARAMS:
        RESNETS:
          DEPTH: 50
    HEAD:
      PARAMS: [
        # ["eval_mlp", {"in_channels": 64, "dims": [9216, 9605]}],
        # ["eval_mlp", {"in_channels": 256, "dims": [9216, 9605]}],
        # ["eval_mlp", {"in_channels": 512, "dims": [8192, 9605]}],
        # ["eval_mlp", {"in_channels": 1024, "dims": [9216, 9605]}],
        # ["eval_mlp", {"in_channels": 2048, "dims": [8192, 9605]}],
        ["mlp", {"dims": [2048, 9605]}],
      ]
    WEIGHTS_INIT:
      PARAMS_FILE: "specify the model weights"
      STATE_DICT_KEY_NAME: classy_state_dict
      # STATE_DICT_KEY_NAME: model_state_dict
    SYNC_BN_CONFIG:
      CONVERT_BN_TO_SYNC_BN: True
      SYNC_BN_TYPE: apex
      GROUP_SIZE: 8
  LOSS:
    name: bce_logits_multiple_output_single_target
    bce_logits_multiple_output_single_target:
      reduction: "none"
      world_size: 64
  OPTIMIZER:
      name: sgd
      # In the OSS Caffe2 benchmark, RN50 models use 1e-4 and AlexNet models 5e-4
      weight_decay: 0.0001
      momentum: 0.9
      num_epochs: 16
      nesterov: True
      regularize_bn: True
      regularize_bias: True
      param_schedulers:
        lr:
          auto_lr_scaling:
            auto_scale: false
            base_value: 0.003
            base_lr_batch_size: 32
          name: multistep
          values: [0.003, 0.0003, 0.00003, 0.000003]
          milestones: [2, 4, 6]
          update_interval: epoch
  DISTRIBUTED:
    BACKEND: nccl
    NUM_NODES: 8
    NUM_PROC_PER_NODE: 8
    INIT_METHOD: tcp
    RUN_ID: auto
  MACHINE:
    DEVICE: gpu
  CHECKPOINT:
    DIR: "."
    AUTO_RESUME: True
    CHECKPOINT_FREQUENCY: 1