sgd_tsm_100e.py

# optimizer
optimizer = dict(
    type='SGD',
    constructor='TSMOptimizerConstructor',
    paramwise_cfg=dict(fc_lr5=True),
    lr=0.01,  # this lr is used for 8 gpus
    momentum=0.9,
    weight_decay=0.0001)
optimizer_config = dict(grad_clip=dict(max_norm=20, norm_type=2))
# learning policy
lr_config = dict(policy='step', step=[40, 80])
total_epochs = 100