_base_ = [
    '../_base_/datasets/imagenet_bs512_mocov3.py',
    '../_base_/default_runtime.py',
]

# model settings
temperature = 1.0
model = dict(
    type='MoCoV3',
    base_momentum=0.01,  # 0.01 for 100e and 300e, 0.004 for 1000e
    backbone=dict(type='ResNet', depth=50, norm_cfg=dict(type='SyncBN')),
    neck=dict(
        type='NonLinearNeck',
        in_channels=2048,
        hid_channels=4096,
        out_channels=256,
        num_layers=2,
        with_bias=False,
        with_last_bn=True,
        with_last_bn_affine=False,
        with_last_bias=False,
        with_avg_pool=True),
    head=dict(
        type='MoCoV3Head',
        predictor=dict(
            type='NonLinearNeck',
            in_channels=256,
            hid_channels=4096,
            out_channels=256,
            num_layers=2,
            with_bias=False,
            with_last_bn=False,
            with_last_bn_affine=False,
            with_last_bias=False,
            with_avg_pool=False),
        loss=dict(type='CrossEntropyLoss', loss_weight=2 * temperature),
        temperature=temperature))

# optimizer
optim_wrapper = dict(
    type='AmpOptimWrapper',
    loss_scale='dynamic',
    optimizer=dict(type='LARS', lr=9.6, weight_decay=1e-6, momentum=0.9),
    paramwise_cfg=dict(
        custom_keys={
            'bn': dict(decay_mult=0, lars_exclude=True),
            'bias': dict(decay_mult=0, lars_exclude=True),
            # bn layer in ResNet block downsample module
            'downsample.1': dict(decay_mult=0, lars_exclude=True),
        }),
)

# learning rate scheduler
param_scheduler = [
    dict(
        type='LinearLR',
        start_factor=1e-4,
        by_epoch=True,
        begin=0,
        end=10,
        convert_to_iter_based=True),
    dict(
        type='CosineAnnealingLR',
        T_max=90,
        by_epoch=True,
        begin=10,
        end=100,
        convert_to_iter_based=True)
]

# runtime settings
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=100)
# only keeps the latest 3 checkpoints
default_hooks = dict(checkpoint=dict(max_keep_ckpts=3))

# NOTE: `auto_scale_lr` is for automatically scaling LR
# based on the actual training batch size.
auto_scale_lr = dict(base_batch_size=4096)