Description
🐛 Inplace RELU issues with backward hooks
Hello there 👋
I encountered a bug I cannot manage to handle with in place ReLUs. In a project of mine, I use backward hooks to get intermediate feature maps and gradients. It used to be working fine up until a few versions ago where densenet was throwing warnings caused by the in place relu over here: https://github.com/pytorch/vision/blob/master/torchvision/models/densenet.py#L217
I updated my torch/torchvision versions recently and now it has turned into an error. I tried on my end and was able to reproduce the bug with a simple sequential that includes an in place ReLU. I have no clue how to get around this, anyone got an idea? The strange thing is that the gradient flows back properly, it's just the backward hook that throws the error.
I had initially opened pytorch/vision#4164 thinking the issue was with densenet, but seeing how I was able to reproduce this with a simple sequential, I think the root problem is with PyTorch actually!
Happy to help solving this, thanks in advance!
To Reproduce
import torch
from torch import nn
# Get the model
mod = nn.Sequential(nn.Conv2d(3, 8, 3, padding=1), nn.ReLU(), nn.Conv2d(8, 8, 3, padding=1), nn.ReLU(inplace=True))
# Hook a module before the in place ReLU
def hook_g(m, input, output): print(input.shape)
mod[2].register_full_backward_hook(hook_g)
# Inference
out = mod(torch.rand((1, 3, 224, 224)))yields:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-6-d15e02b5edd3> in <module>
----> 1 out = mod(torch.rand((1, 3, 224, 224)))
~/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1049 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1050 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051 return forward_call(*input, **kwargs)
1052 # Do not call functions when jit is used
1053 full_backward_hooks, non_full_backward_hooks = [], []
~/miniconda3/lib/python3.8/site-packages/torch/nn/modules/container.py in forward(self, input)
137 def forward(self, input):
138 for module in self:
--> 139 input = module(input)
140 return input
141
~/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
1049 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
1050 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1051 return forward_call(*input, **kwargs)
1052 # Do not call functions when jit is used
1053 full_backward_hooks, non_full_backward_hooks = [], []
~/miniconda3/lib/python3.8/site-packages/torch/nn/modules/activation.py in forward(self, input)
100
101 def forward(self, input: Tensor) -> Tensor:
--> 102 return F.relu(input, inplace=self.inplace)
103
104 def extra_repr(self) -> str:
~/miniconda3/lib/python3.8/site-packages/torch/nn/functional.py in relu(input, inplace)
1294 return handle_torch_function(relu, (input,), input, inplace=inplace)
1295 if inplace:
-> 1296 result = torch.relu_(input)
1297 else:
1298 result = torch.relu(input)
RuntimeError: Output 0 of BackwardHookFunctionBackward is a view and is being modified inplace. This view was created inside a custom Function (or because an input was returned as-is) and the autograd logic to handle view+inplace would override the custom backward associated with the custom Function, leading to incorrect gradients. This behavior is forbidden. You can remove this warning by cloning the output of the custom Function.
Expected behavior
It used to be throwing a warning but this is now causing an error, and I'm unsure how to get around this.
Environment
- PyTorch / torchvision Version: 1.9.0 / 0.10.0
- OS: Ubuntu 20.04.2 LTS (x86_64)
- How you installed PyTorch / torchvision: conda
- Python version: 3.8 (64-bit runtime)
- CUDA/cuDNN version: 11.2.152 (cuDNN 8.2.0)
cc @ezyang @albanD @zou3519 @gqchen @pearu @nikitaved @soulitzer @lezcano @Varal7 @mruberry @jbschlosser