原因,model需要先放到cuda上,根据你的设置,所以是load_state_dict错了。
原来:
model = se_resnet20_v2()
model.load_state_dict(torch.load('/home/zj/senetial/save_models/SENet/SEnet_model_2.pth')
if use_gpu and len(device_ids)>1:#多gpu训练
model = model.cuda(device_ids[0])
model = nn.DataParallel(model, device_ids=device_ids)
if use_gpu and len(device_ids)==1:#单gpu训练
model = model.cuda()
print model
#model=TSnet(17)
model = se_resnet20_v2()
改:
if use_gpu and len(device_ids)>1:#多gpu训练
model = model.cuda(device_ids[0])
model = nn.DataParallel(model, device_ids=device_ids)
if use_gpu and len(device_ids)==1:#单gpu训练
model = model.cuda()
print model
model.load_state_dict(torch.load('/home/zj/senetial/save_models/SENet/SEnet_model_2.pth'))
单GPU:
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
多GPU:
device_ids = [0,1,2,3]
model = model.cuda(device_ids[0])
model = nn.DataParallel(model, device_ids=device_ids)
optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9, weight_decay=0.001)
optimizer = nn.DataParallel(optimizer, device_ids=device_ids)
optimizer.module.step()
for param_lr in optimizer.module.param_groups: # 同样是要加module
# param_lr['lr'] = param_lr['lr'] * 0.999
加载多GPU预训练模型
model = ft_net()
pretained_model = torch.load('./model/all/8_model.pkl')
pretained_dict = pretained_model.module.state_dict()
model = ft_net()
model.load_state_dict(pretained_dict)