File size: 6,981 Bytes
edcf5ee | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 | # --------------------------------------------------------
# SimMIM
# Copyright (c) 2021 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Written by Ze Liu
# Modified by Zhenda Xie
# --------------------------------------------------------
import json
from functools import partial
from torch import optim as optim
def build_optimizer(config, model, logger, is_pretrain):
if is_pretrain:
return build_pretrain_optimizer(config, model, logger)
else:
return build_finetune_optimizer(config, model, logger)
def build_pretrain_optimizer(config, model, logger):
logger.info('>>>>>>>>>> Build Optimizer for Pre-training Stage')
skip = {}
skip_keywords = {}
if hasattr(model, 'no_weight_decay'):
skip = model.no_weight_decay()
logger.info(f'No weight decay: {skip}')
if hasattr(model, 'no_weight_decay_keywords'):
skip_keywords = model.no_weight_decay_keywords()
logger.info(f'No weight decay keywords: {skip_keywords}')
parameters = get_pretrain_param_groups(model, logger, skip, skip_keywords)
opt_lower = config.TRAIN.OPTIMIZER.NAME.lower()
optimizer = None
if opt_lower == 'sgd':
optimizer = optim.SGD(parameters, momentum=config.TRAIN.OPTIMIZER.MOMENTUM, nesterov=True,
lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)
elif opt_lower == 'adamw':
optimizer = optim.AdamW(parameters, eps=config.TRAIN.OPTIMIZER.EPS, betas=config.TRAIN.OPTIMIZER.BETAS,
lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)
logger.info(optimizer)
return optimizer
def get_pretrain_param_groups(model, logger, skip_list=(), skip_keywords=()):
has_decay = []
no_decay = []
has_decay_name = []
no_decay_name = []
for name, param in model.named_parameters():
if not param.requires_grad:
continue
if len(param.shape) == 1 or name.endswith(".bias") or (name in skip_list) or \
check_keywords_in_name(name, skip_keywords):
no_decay.append(param)
no_decay_name.append(name)
else:
has_decay.append(param)
has_decay_name.append(name)
logger.info(f'No decay params: {no_decay_name}')
logger.info(f'Has decay params: {has_decay_name}')
return [{'params': has_decay},
{'params': no_decay, 'weight_decay': 0.}]
def build_finetune_optimizer(config, model, logger):
logger.info('>>>>>>>>>> Build Optimizer for Fine-tuning Stage')
if config.MODEL.TYPE == 'swin':
depths = config.MODEL.SWIN.DEPTHS
num_layers = sum(depths)
get_layer_func = partial(get_swin_layer, num_layers=num_layers + 2, depths=depths)
elif config.MODEL.TYPE == 'vit':
num_layers = config.MODEL.VIT.DEPTH
get_layer_func = partial(get_vit_layer, num_layers=num_layers + 2)
else:
raise NotImplementedError
scales = list(config.TRAIN.LAYER_DECAY ** i for i in reversed(range(num_layers + 2)))
skip = {}
skip_keywords = {}
if hasattr(model, 'no_weight_decay'):
skip = model.no_weight_decay()
logger.info(f'No weight decay: {skip}')
if hasattr(model, 'no_weight_decay_keywords'):
skip_keywords = model.no_weight_decay_keywords()
logger.info(f'No weight decay keywords: {skip_keywords}')
parameters = get_finetune_param_groups(
model, logger, config.TRAIN.BASE_LR, config.TRAIN.WEIGHT_DECAY,
get_layer_func, scales, skip, skip_keywords)
opt_lower = config.TRAIN.OPTIMIZER.NAME.lower()
optimizer = None
if opt_lower == 'sgd':
optimizer = optim.SGD(parameters, momentum=config.TRAIN.OPTIMIZER.MOMENTUM, nesterov=True,
lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)
elif opt_lower == 'adamw':
optimizer = optim.AdamW(parameters, eps=config.TRAIN.OPTIMIZER.EPS, betas=config.TRAIN.OPTIMIZER.BETAS,
lr=config.TRAIN.BASE_LR, weight_decay=config.TRAIN.WEIGHT_DECAY)
logger.info(optimizer)
return optimizer
def get_vit_layer(name, num_layers):
if name in ("cls_token", "mask_token", "pos_embed"):
return 0
elif name.startswith("patch_embed"):
return 0
elif name.startswith("rel_pos_bias"):
return num_layers - 1
elif name.startswith("blocks"):
layer_id = int(name.split('.')[1])
return layer_id + 1
else:
return num_layers - 1
def get_swin_layer(name, num_layers, depths):
if name in ("mask_token"):
return 0
elif name.startswith("patch_embed"):
return 0
elif name.startswith("layers"):
layer_id = int(name.split('.')[1])
block_id = name.split('.')[3]
if block_id == 'reduction' or block_id == 'norm':
return sum(depths[:layer_id + 1])
layer_id = sum(depths[:layer_id]) + int(block_id)
return layer_id + 1
else:
return num_layers - 1
def get_finetune_param_groups(model, logger, lr, weight_decay, get_layer_func, scales, skip_list=(), skip_keywords=()):
parameter_group_names = {}
parameter_group_vars = {}
for name, param in model.named_parameters():
if not param.requires_grad:
continue
if len(param.shape) == 1 or name.endswith(".bias") or (name in skip_list) or \
check_keywords_in_name(name, skip_keywords):
group_name = "no_decay"
this_weight_decay = 0.
else:
group_name = "decay"
this_weight_decay = weight_decay
if get_layer_func is not None:
layer_id = get_layer_func(name)
group_name = "layer_%d_%s" % (layer_id, group_name)
else:
layer_id = None
if group_name not in parameter_group_names:
if scales is not None:
scale = scales[layer_id]
else:
scale = 1.
parameter_group_names[group_name] = {
"group_name": group_name,
"weight_decay": this_weight_decay,
"params": [],
"lr": lr * scale,
"lr_scale": scale,
}
parameter_group_vars[group_name] = {
"group_name": group_name,
"weight_decay": this_weight_decay,
"params": [],
"lr": lr * scale,
"lr_scale": scale
}
parameter_group_vars[group_name]["params"].append(param)
parameter_group_names[group_name]["params"].append(name)
logger.info("Param groups = %s" % json.dumps(parameter_group_names, indent=2))
return list(parameter_group_vars.values())
def check_keywords_in_name(name, keywords=()):
isin = False
for keyword in keywords:
if keyword in name:
isin = True
return isin |