def build_opt_lr(model, cfg, args, epoch):
"""
获取优化方法和学习率
:param model:
:param cfg:
:param args:
:param epoch:
:return:
"""
# 获取要训练的网络
backbone_feature = model.features.param_groups(
cfg['lr']['start_lr'], cfg['lr']['feature_lr_mult'])
if len(backbone_feature) == 0:
# 获取要训练的rpn网络的参数
trainable_params = model.rpn_model.param_groups(
cfg['lr']['start_lr'], cfg['lr']['rpn_lr_mult'], 'mask')
else:
# 获取基础网络，rpn和mask网络的训练参数
trainable_params = backbone_feature + \
model.rpn_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['rpn_lr_mult']) + \
model.mask_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['mask_lr_mult'])
# 随机梯度下降算法优化
optimizer = torch.optim.SGD(trainable_params,
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# 获取学习率
lr_scheduler = build_lr_scheduler(optimizer, cfg['lr'], epochs=args.epochs)
# 更新学习率
lr_scheduler.step(epoch)
# 返回优化器和学习率
return optimizer, lr_scheduler
def build_opt_lr(model, cfg, args, epoch):
trainable_params = model.features.param_groups(cfg['lr']['start_lr'], cfg['lr']['feature_lr_mult']) + \
model.rpn_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['rpn_lr_mult'])
optimizer = torch.optim.SGD(trainable_params, args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
lr_scheduler = build_lr_scheduler(optimizer, cfg['lr'], epochs=args.epochs)
lr_scheduler.step(epoch)
return optimizer, lr_scheduler
def build_opt_lr(model, epoch):
trainable_params = []
for param in model.backbone.parameters():
param.requires_grad = False
for i in range(len(cfg.backbone.unfix_steps)):
if epoch >= cfg.backbone.unfix_steps[i]:
layer = getattr(model.backbone, cfg.backbone.unfix_layers[i])
for param in layer.parameters():
param.requires_grad = True
trainable_params += [{
'params': layer.parameters(),
'lr': cfg.backbone.unfix_lr[i]
}]
if cfg.adjust.adjust:
trainable_params += [{'params': model.neck.parameters(), 'lr': 1}]
trainable_params += [{'params': model.siamese.parameters(), 'lr': 1}]
if cfg.cornerdet.cornerdet:
if epoch >= cfg.train.pretrain_epoch:
if cfg.attention.attention:
for param in model.attention.parameters():
param.requires_grad = True
trainable_params += [{
'params': model.attention.parameters(),
'lr': 1
}]
for param in model.cornerdet.parameters():
param.requires_grad = True
trainable_params += [{
'params': model.cornerdet.parameters(),
'lr': 1
}]
else:
if cfg.attention.attention:
for param in model.attention.parameters():
param.requires_grad = False
for param in model.cornerdet.parameters():
param.requires_grad = False
optimizer = torch.optim.SGD(trainable_params,
momentum=cfg.train.momentum,
weight_decay=cfg.train.weight_decay)
lr_scheduler = build_lr_scheduler(optimizer, cfg.train.lr, cfg.train.epoch)
logger.info('change training parameters.')
logger.info("model\n{}".format(describe(model)))
return optimizer, lr_scheduler
def build_opt_lr(model, cfg, args, epoch):
backbone_feature = model.features.param_groups(
cfg['lr']['start_lr'], cfg['lr']['feature_lr_mult'])
if len(backbone_feature) == 0:
trainable_params = model.rpn_model.param_groups(
cfg['lr']['start_lr'], cfg['lr']['rpn_lr_mult'], 'mask')
else:
trainable_params = backbone_feature + \
model.rpn_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['rpn_lr_mult']) + \
model.kp_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['mask_lr_mult']) + \
model.pose_corr.param_groups(cfg['lr']['start_lr'], cfg['lr']['mask_lr_mult'])
optimizer = torch.optim.SGD(trainable_params,
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
lr_scheduler = build_lr_scheduler(optimizer, cfg['lr'], epochs=args.epochs)
lr_scheduler.step(epoch)
return optimizer, lr_scheduler
def build_opt_lr(model, cfg, args, epoch):
'''
对模型参数进行优化
:param model:
:param cfg:
:param args:
:param epoch:
:return:
'''
trainable_params = model.mask_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['mask_lr_mult']) + \
model.refine_model.param_groups(cfg['lr']['start_lr'], cfg['lr']['mask_lr_mult'])
# 随机梯度下降算法：
optimizer = torch.optim.SGD(trainable_params, args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# 获取学习率
lr_scheduler = build_lr_scheduler(optimizer, cfg['lr'], epochs=args.epochs)
# 更新学习率
lr_scheduler.step(epoch)
return optimizer, lr_scheduler