mmselfsup/configs/benchmarks/classification/imagenet/swin-base_ft-8xb256-coslr-1...

75 lines
1.9 KiB
Python

_base_ = [
'../_base_/models/swin-base.py',
'../_base_/datasets/imagenet.py',
'../_base_/schedules/adamw_coslr-100e_in1k.py',
'../_base_/default_runtime.py',
]
# dataset
img_norm_cfg = dict(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
train_pipeline = [
dict(
type='RandomAug',
input_size=192,
color_jitter=0.4,
auto_augment='rand-m9-mstd0.5-inc1',
interpolation='bicubic',
re_prob=0.25,
re_mode='pixel',
re_count=1,
mean=(0.485, 0.456, 0.406),
std=(0.229, 0.224, 0.225))
]
test_pipeline = [
dict(type='Resize', size=219, interpolation=3),
dict(type='CenterCrop', size=192),
dict(type='ToTensor'),
dict(type='Normalize', **img_norm_cfg)
]
data = dict(
samples_per_gpu=256,
drop_last=False,
workers_per_gpu=32,
train=dict(pipeline=train_pipeline),
val=dict(pipeline=test_pipeline))
# model
model = dict(backbone=dict(init_cfg=dict()))
# optimizer
optimizer = dict(
lr=1.25e-3 * 2048 / 512,
paramwise_options={
'norm': dict(weight_decay=0.),
'bias': dict(weight_decay=0.),
'absolute_pos_embed': dict(weight_decay=0.),
'relative_position_bias_table': dict(weight_decay=0.)
},
constructor='TransformerFinetuneConstructor',
model_type='swin',
layer_decay=0.9)
# clip gradient
optimizer_config = dict(grad_clip=dict(max_norm=5.0))
# learning policy
lr_config = dict(
policy='CosineAnnealing',
min_lr=2.5e-7 * 2048 / 512,
warmup='linear',
warmup_iters=20,
warmup_ratio=2.5e-7 / 1.25e-3,
warmup_by_epoch=True,
by_epoch=False)
# mixed precision
fp16 = dict(loss_scale='dynamic')
# runtime
checkpoint_config = dict(interval=1, max_keep_ckpts=3, out_dir='')
persistent_workers = True
log_config = dict(
interval=100, hooks=[
dict(type='TextLoggerHook'),
])