resnet18_linear_eval_imagenet.yaml

SEED: 100
MODEL:
    ARCH: us_resnet18
    INPUTSHAPE: [224, 224]
    #PRETRAINED: /data/train_log_OFA/imagenet/r18/simclr_momenrtum_max_with_momentum_r18_imagenet_sandwich3T_asymmertric_mse_distill_feat_2_3_4_distill_mse_groupreg_A/checkpoint_width_1.0.pth.tar
    #CHECKPOINT: /data/train_log_OFA/resnet18_linear_eval_imagenet_0.75_ours/checkpoint.pth.tar
    #CHECKPOINT: /data/train_log_OFA/resnet18_linear_eval_imagenet_0.25_baseline/checkpoint.pth.tar
    #PRETRAINED: /data/train_log_OFA/imagenet/r18/simsiam_momentum_us_r18_imagenet_sandwich3T_asymmertric_infoncev2_distill_new_head_200ep/checkpoint_width_0.25.pth.tar
    #PRETRAINED: /data/train_log_OFA/imagenet/r18/byol_us_r18_imagenet_200ep_single_0.25_rerun/checkpoint.pth.tar
    NUM_CLASSES: 1000
    OFA:
        CALIBRATE_BN: False
        NUM_SAMPLE_TRAINING: 1
        WIDTH_MULT: 0.25
        SANDWICH: False
TRAIN:
    EPOCHS: 60
    USE_DDP: True
    LINEAR_EVAL: True
    DATASET: imagenet
    BATCH_SIZE: 256 # per-gpu
    OPTIMIZER: 
        NAME: lars
        MOMENTUM: 0.9
        WEIGHT_DECAY: 0.000 # 1e-5
    LR_SCHEDULER:
        WARMUP_EPOCHS: 0
        WARMUP_LR: 0.0002 # 1e-4
        BASE_LR: 0.8 # 1e-2
        MIN_LR: 0.
        TYPE: cosine
        DECAY_RATE: 0.1
        DECAY_MILESTONES : [30, 40]
    LOSS: 
        CRITERION:
            NAME: CrossEntropy
        #REGULARIZER:
        #    NAME: PACT
        LAMBDA: 0.0001
    METER:
        NAME: ACC
        ACC:
            TOPK: [1, 5]
    RUNNER:
        NAME: default