【pytorch】freeze

2021-12-06 21:16:56 浏览数 (2)

freeze bn:

把所有相关的bn设置为 momentum=1.0 。

freeze 正常参数:

先比较两个state_dict,来freeze交集:

代码语言:javascript复制
def freeze_model(model, defined_dict, keep_step=None):

    for (name, param) in model.named_parameters():
        if name in defined_dict:
            param.requires_grad = False
        else:
            pass

    freezed_num, pass_num = 0, 0
    for (name, param) in model.named_parameters():
        if param.requires_grad == False:
            freezed_num  = 1
        else:
            pass_num  = 1

    return model, freezed_num, pass_num

之后再指定optimizer的时候要注意避开这部分参数,防止被freeze的参数重新被optimizer将requires_grad置为True:

代码语言:javascript复制
# 注意这里的 filter 是python3的写法,所以直接用就行,没必要加 list() 。
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=0.001, 
    betas=(0.9, 0.999), eps=1e-08, weight_decay=1e-5)

check freeze

代码语言:javascript复制
def check_state_dict_same(pre_dict, cur_dict):
    diff_lst = list()
    for key in cur_dict.keys():
        if key in pre_dict:
            if not torch.equal(cur_dict[key], pre_dict[key]):
                diff_lst.append(key)
    return diff_lst

diff_lst = check_state_dict_same(pre_dict=pre_state_dict, cur_dict=model.state_dict())
if diff_lst:
    print('nn Change by follow pars: n')
    print(diff_lst)
    print('nn')
    exit(0)
else:
    print('nn Model is successfully freezed . n')

bn在model.train()的模式下还是会自动更新参数的,就算放到 with torch.no_grad(): 里面或者把每个bn的参数都 p.requires_grad = False 也没用。

由于bn模块经常被写成可复用的形式。因此固定住bn的时候,记得另外写一套供不需被固定的分支所调用的bn模块。

0 人点赞