pytorch使用多块GPU训练代码

代码

# 模型，训练模式
net = Net(opt)
net.train()


# 损失函数，优化器
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr, (b1, b2))


# 设定gpu
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"

# 如果epoch不为0，就加载已经训练过的ckpt
if epoch != 0:
    ckpt = os.path.join(opt.out_path, 'net.pth')
    
    if torch.cuda.is_available():
        device = torch.device("cuda:0")
        net.to(device)
        criterion.to(device)

        # 多块gpu只转移模型
        if torch.cuda.device_count() > 1:
            net = nn.DataParallel(net, device_ids=[0,1,2,3])
            net.to(device)

        net.load_state_dict(torch.load(ckpt))

# 否则不加载
else:
    if torch.cuda.is_available():
        device = torch.device("cuda:0")
        net.to(device)
        criterion.to(device)
        
        if torch.cuda.device_count() > 1:
            net = nn.DataParallel(net, device_ids=[0,1,2,3])
            net.to(device)

模型转移

主要记载在GPU保存的模型读取到CPU，以及CPU保存的模型读取到GPU，主要的不同在于是否有module，无非就是加上和去掉的问题

CPU->GPU

# cpu保存的模型在gpu加载的方法
from collections import OrderedDict
new_state_dict = OrderedDict()
for key, value in torch.load(ckpt).items():
    name = 'module.' + key # 加上module
    new_state_dict[name] = value
    
net.load_state_dict(new_state_dict)

GPU->CPU

from collections import OrderedDict
state_dict_new = OrderedDict()
for key, value in torch.load(ckpt).items():
    name = key[7:] # 去掉module
    state_dict_new[name] = value
    
net.load_state_dict(state_dict_new)

或者还可以写得更简单：

net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(model_path, map_location=torch.device('cpu')).items()})

微调网络

官方方法

# 冻结除fc以外的层，再次写一遍fc层，默认求梯度
for param in net.parameters():
	param.requires_grad = False
	net.fc = nn.Linear(512, 100)

新方法

# 更加灵活决定那些层是不求梯度的
for param in list(net.parameters())[1:49]:
    param.requires_grad = False