Instructions for torch use

1.torch.load_state_dict() Function Usage Summary
stay Pytorch After building a model in , Generally, pre training is needed to load the weight .torch.load_state_dict() The function is used to load the pre trained parameter weights into the new model , The operation mode is as follows ：

sd_net = torchvision.models.resnte50(pretrained=False)
sd_net.load_state_dict(torch.load('*.pth'), strict=True)

The focus of this blog post is attribute strict; When strict=True, It is required that the key value of the pre training weight level should be completely consistent with the name of the weight level in the newly constructed model ; If the newly built model is partially fine tuned in the number of layers , The above code will report an error ： say key No correspondence .

here , If we adopt strict=False Can perfectly solve this problem . That is to say , The key value of the matching layer in the training weight and the newly constructed network is used , If there is no, it will be initialized by default .
2.class torch.optim.lr_scheduler.CosineAnnealingLR

torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,T_max,eta_min=0,last_epoch=-1)

T_max： Because the learning rate changes periodically , This parameter represents the 1/2, in other words , The initial learning rate is a, after 2*T_max After time , After a cycle of changes, the learning rate is still a.
eta_min： Represents the minimum variable learning rate , The default is 0