61 lines
1.6 KiB
Python
61 lines
1.6 KiB
Python
_base_ = [
|
|
'../_base_/datasets/imagenet_bs32_byol.py',
|
|
'../_base_/schedules/imagenet_lars_coslr_200e.py',
|
|
'../_base_/default_runtime.py',
|
|
]
|
|
|
|
train_dataloader = dict(batch_size=256)
|
|
|
|
# model settings
|
|
model = dict(
|
|
type='BYOL',
|
|
base_momentum=0.01,
|
|
backbone=dict(
|
|
type='ResNet',
|
|
depth=50,
|
|
norm_cfg=dict(type='SyncBN'),
|
|
zero_init_residual=False),
|
|
neck=dict(
|
|
type='NonLinearNeck',
|
|
in_channels=2048,
|
|
hid_channels=4096,
|
|
out_channels=256,
|
|
num_layers=2,
|
|
with_bias=True,
|
|
with_last_bn=False,
|
|
with_avg_pool=True),
|
|
head=dict(
|
|
type='LatentPredictHead',
|
|
predictor=dict(
|
|
type='NonLinearNeck',
|
|
in_channels=256,
|
|
hid_channels=4096,
|
|
out_channels=256,
|
|
num_layers=2,
|
|
with_bias=True,
|
|
with_last_bn=False,
|
|
with_avg_pool=False),
|
|
loss=dict(type='CosineSimilarityLoss')),
|
|
)
|
|
|
|
# optimizer
|
|
optimizer = dict(type='LARS', lr=4.8, momentum=0.9, weight_decay=1e-6)
|
|
optim_wrapper = dict(
|
|
type='OptimWrapper',
|
|
optimizer=optimizer,
|
|
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),
|
|
}),
|
|
)
|
|
|
|
# runtime settings
|
|
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)
|