1.在加载了部分参数以后,直接另参数的requires_grad = False
self.encoder = BertModel.from_pretrained('bert-base-chinese')
for i, p in enumerate(self.parameters()):
p.requires_grad = False
2.优化器中去掉requires_grad = False 的参数
self.optimizer = torch.optim.Adam(filter(lambda p: p.requires_grad, self.model.parameters()), lr=2e-6)