AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor'

Question

I get error on line x_stats = dec(z).float().

import torch.nn.functional as F

z_logits = enc(x)
z = torch.argmax(z_logits, axis=1)
z = F.one_hot(z, num_classes=enc.vocab_size).permute(0, 3, 1, 2).float()

x_stats = dec(z).float()
x_rec = unmap_pixels(torch.sigmoid(x_stats[:, :3]))
x_rec = T.ToPILImage(mode='RGB')(x_rec[0])

display_markdown('Reconstructed image:')
display(x_rec)

I tried to downgrade and reinstall the torch package but that didn't help the issue. My package version is torch==1.11.0

Full traceback:

AttributeError                            Traceback (most recent call last)
/Users/hanpham/github/DALL-E/notebooks/usage.ipynb Cell 4' in <cell line: 7>()
      4 z = torch.argmax(z_logits, axis=1)
      5 z = F.one_hot(z, num_classes=enc.vocab_size).permute(0, 3, 1, 2).float()
----> 7 x_stats = dec(z).float()
      8 x_rec = unmap_pixels(torch.sigmoid(x_stats[:, :3]))
      9 x_rec = T.ToPILImage(mode='RGB')(x_rec[0])

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/module.py:1110, in Module._call_impl(self, *input, **kwargs)
   1106 # If we don't have any hooks, we want to skip the rest of the logic in
   1107 # this function, and just call forward.
   1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1109         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110     return forward_call(*input, **kwargs)
   1111 # Do not call functions when jit is used
   1112 full_backward_hooks, non_full_backward_hooks = [], []

File /opt/homebrew/lib/python3.9/site-packages/dall_e/decoder.py:94, in Decoder.forward(self, x)
     91 if x.dtype != torch.float32:
     92     raise ValueError('input must have dtype torch.float32')
---> 94 return self.blocks(x)

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/module.py:1110, in Module._call_impl(self, *input, **kwargs)
   1106 # If we don't have any hooks, we want to skip the rest of the logic in
   1107 # this function, and just call forward.
   1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1109         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110     return forward_call(*input, **kwargs)
   1111 # Do not call functions when jit is used
   1112 full_backward_hooks, non_full_backward_hooks = [], []

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/container.py:141, in Sequential.forward(self, input)
    139 def forward(self, input):
    140     for module in self:
--> 141         input = module(input)
    142     return input

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/module.py:1110, in Module._call_impl(self, *input, **kwargs)
   1106 # If we don't have any hooks, we want to skip the rest of the logic in
   1107 # this function, and just call forward.
   1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1109         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110     return forward_call(*input, **kwargs)
   1111 # Do not call functions when jit is used
   1112 full_backward_hooks, non_full_backward_hooks = [], []

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/container.py:141, in Sequential.forward(self, input)
    139 def forward(self, input):
    140     for module in self:
--> 141         input = module(input)
    142     return input

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/module.py:1110, in Module._call_impl(self, *input, **kwargs)
   1106 # If we don't have any hooks, we want to skip the rest of the logic in
   1107 # this function, and just call forward.
   1108 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1109         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1110     return forward_call(*input, **kwargs)
   1111 # Do not call functions when jit is used
   1112 full_backward_hooks, non_full_backward_hooks = [], []

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/upsampling.py:154, in Upsample.forward(self, input)
    152 def forward(self, input: Tensor) -> Tensor:
    153     return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners,
--> 154                          recompute_scale_factor=self.recompute_scale_factor)

File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/module.py:1185, in Module.__getattr__(self, name)
   1183     if name in modules:
   1184         return modules[name]
-> 1185 raise AttributeError("'{}' object has no attribute '{}'".format(
   1186     type(self).__name__, name))

AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor'

Answer 1

Comment out the recompute_scale_factor=self.recompute_scale_factor part in the source code. File /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/upsampling.py:154, in Upsample.forward(self, input)

152 def forward(self, input: Tensor) -> Tensor:
153     return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
154                          
#recompute_scale_factor=self.recompute_scale_factor)

Answer 2

Install Torch version, this will solve the issue

---

pip install torchvision==0.10.1
pip install torch==1.9.1

Answer 3

Also getting this error with torch 1.11.0 Would love to hear how people solve it

Looks like it's an issue with 1.11.0: https://github.com/openai/DALL-E/issues/54

Edit: Following these instructions solved it for me: https://github.com/openai/DALL-E/issues/54#issuecomment-1092826376

Answer 4

I think your issue might be along the lines of https://github.com/ultralytics/yolov5/issues/6948.

I'm not familiar with pytorch; but suggestions were:

1. wait for the next version (not really that great a suggestion, sorry)

2. comment out the code as pointed in https://github.com/ultralytics/yolov5/issues/6948#issuecomment-1075528897, that is:

In /opt/homebrew/lib/python3.9/site-packages/torch/nn/modules/upsampling.py in line 153-154:

Change:

  return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners,
recompute_scale_factor=self.recompute_scale_factor)

To:

  return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
# recompute_scale_factor=self.recompute_scale_factor)

or

  return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners,
# recompute_scale_factor=self.recompute_scale_factor
)

In my opinion, as a 'workaround', you could do the comment out option until a new version comes out at which, you can undo the comment out, and upgrade.

While I agree this is an 'answer', it isn't the perfect answer. My apologies.