RuntimeError: CUDA out of memory

今天在训练模型的时候突然报了显存不够的问题，然后分析了一下，找到了解决的办法，这里记录一下，方便以后查阅。

注：以下的解决方案是在模型测试而不是模型训练时出现这个报错的！

完整的报错信息：

Traceback (most recent call last):File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/main.py", line 420, in <module>main()File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/main.py", line 414, in maintrain_with_cross_validate(training_epochs, kfolds, train_indices, eval_indices, X_train, Y_train, model, losser, optimizer)File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/main.py", line 77, in train_with_cross_validateval_probs = model(inputs)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_implreturn forward_call(*input, **kwargs)File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/model/S_CAMLP_Net.py", line 235, in forwardx = self.camlp_mixer(x) # (batch_size, F, C, L)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_implreturn forward_call(*input, **kwargs)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/container.py", line 139, in forwardinput = module(input)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_implreturn forward_call(*input, **kwargs)File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/model/S_CAMLP_Net.py", line 202, in forwardx = self.time_mixing_unit(x)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_implreturn forward_call(*input, **kwargs)File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/model/S_CAMLP_Net.py", line 186, in forwardx = self.mixing_unit(x)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_implreturn forward_call(*input, **kwargs)File "/home/pytorch/LiangXiaohan/MI_Same_limb/Joint_Motion_Decoding/SelfAten_Mixer/model/S_CAMLP_Net.py", line 147, in forwardx = self.activate(x)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1130, in _call_implreturn forward_call(*input, **kwargs)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/modules/activation.py", line 772, in forwardreturn F.leaky_relu(input, self.negative_slope, self.inplace)File "/home/pytorch/anaconda3/envs/pytorch_env/lib/python3.7/site-packages/torch/nn/functional.py", line 1633, in leaky_reluresult = torch._C._nn.leaky_relu(input, negative_slope)
RuntimeError: CUDA out of memory. Tried to allocate 2.49 GiB (GPU 0; 23.70 GiB total capacity; 21.49 GiB already allocated; 550.81 MiB free; 21.53 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

因为自己写的程序训练完成一轮会有输出，所以这些信息是在模型预测过程中发生的。

关键的报错信息：

RuntimeError: CUDA out of memory. Tried to allocate 2.49 GiB (GPU 0; 23.70 GiB total capacity; 21.49 GiB already allocated; 550.81 MiB free; 21.53 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

大体意思就是显存不够了。

通过下面的代码查看程序运行过程中显卡的状态：

nvidia-smi -l 1

模型加载完成后，此时的显卡状态：

模型训练过程中显卡的状态：

模型训练完成，开始模型预测阶段，并且是数据输入模型之后，紧接着出现如下的显卡状态，并且这个状态持续时间很短，在显示过程中，只有一次输出结果是这样的：

紧接着程序报错，显卡内存被释放，显卡的任务栏中，运行的程序也没有了：

然后，就感觉很奇怪，觉得是梯度的问题，因为在训练的时候很正常，然后模型预测就出现问题了，然后模型训练需要梯度信息，模型预测不需要梯度信息，就尝试着解决梯度的问题：

就是在模型训练代码的前面加入下面这句话：

with torch.no_grad():

更改后的代码如下所示：

with torch.no_grad():# validationmodel.eval()inputs = x_eval.to(device)val_probs = model(inputs)val_acc = (val_probs.argmax(dim=1) == y_eval.to(device)).float().mean()# print(f"Eval : Epoch : {iter} - kfold : {kfold+1} - acc: {val_acc:.4f}\n")epoch_val_acc += val_acc

更改之后模型预测阶段显卡的状态如下所示：

然后开始新一轮的训练过程，显卡的显存占用情况也没有再发生变化。

这样就不再报错了！！！