mmpretrain/configs/_base_/schedules/imagenet_bs2048_AdamW.py

# optimizer
# In ClassyVision, the lr is set to 0.003 for bs4096.
# In this implementation(bs2048), lr = 0.003 / 4096 * (32bs * 64gpus) = 0.0015
optim_wrapper = dict(
    optimizer=dict(type='AdamW', lr=0.0015, weight_decay=0.3),
    # specific to vit pretrain
    paramwise_cfg=dict(custom_keys={
        '.cls_token': dict(decay_mult=0.0),
        '.pos_embed': dict(decay_mult=0.0)
    }),
)

# learning policy
warmup_epochs = 15  # about 10000 iterations for ImageNet-1k
param_scheduler = [
    # warm up learning rate scheduler
    dict(
        type='LinearLR',
        start_factor=1e-3,
        by_epoch=True,
        end=warmup_epochs,
        # update by iter
        convert_to_iter_based=True),
    # main learning rate scheduler
    dict(
        type='CosineAnnealingLR',
        eta_min=1e-5,
        by_epoch=True,
        begin=warmup_epochs)
]

# train, val, test setting
train_cfg = dict(by_epoch=True, max_epochs=300, val_interval=1)
val_cfg = dict()
test_cfg = dict()

# NOTE: `auto_scale_lr` is for automatically scaling LR,
# based on the actual training batch size.
auto_scale_lr = dict(base_batch_size=2048)