1. 模型保存以及加载方法
torch.save(model, '/path/to/save')
model = torch.load('/path/to/load')
torch.save(model.state_dict(), '/path/to/save')
model = NET()
mode.load_state_dict(torch.load('/path/to/load'))
state = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch
}
torch.save(state, '/path/to/save')
checkpoint = torch.load('/path/to/load')
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
epoch = checkpoint['epoch']
2. 单 GPU 和 单 CPU 参数-模型相互加载
model.load_state_dict(torch.load('/path/to/load'))
state_dict = torch.load('/path/to/load', map_location=lambda storage, loc: storage)
state_dict = torch.load('/path/to/load', map_location='cpu')
model.load_state_dict(states_dict)
state_dict = torch.load('/path/to/load', map_location=lambda storage, loc: storage.cuda)
state_dict = torch.load('/path/to/load', map_location='cuda:0')
state_dict = torch.load('/path/to/load', map_location=lambda storage, loc: storage.cuda(1))
state_dict = torch.load('/path/to/load', map_location='cuda:1')
3. 多 GPU 模型-参数
torch.save(model.module, '/path/to/save')
torch.save(model.module.state_dict(), '/path/to/save')
4. 单 GPU or CPU 模型加载多 GPU 参数
device = torch.device('cpu')
model = NET().to(device)
state_dict = torch.load('/path/to/load', map_location=device)
state_dict_new = {}
for k, v in state_dict.items():
new_k = k[7:]
state_dict_new[new_k] = v
model.load_state_dict(state_dict_new)
5. 单 GPU or CPU 加载 多GPU模型+参数
model_cpu = NET().to('cpu')
model_gpu = NET().to('cuda:0')
pretrained_model = torch.load('/path/to/load')
pretrained_dict = pretrained_model.module.state_dict()
model_cpu.load_state_dict(pretrained_dict)
model_gpu.load_state_dicr(pretrained_dicr)
6. 多 GPU 加载 多GPU参数
model = NET().to('cuda:0')
model = torch.nn.DataParallel(model, device_ids=[0, 1])
state_dict = torch.load('/path/to/load')
model.load_state_dict(state_dict)