module must have its parameters and buffers on device cuda:0

出现上述错误,是因为DP加载,默认会将模型和参数放在GPU设备上,如果想用cpu去推理DP加载的模型,必须对DP 模型进行下面一步转换才可以.

将torch.nn.DataParalle 放到cpu上

model = torch.nn.DataParallel(model, device_ids=[0,1,2,3])
model = model.module.to(torch.device('cpu'))