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

[balakreshnan]

When i install and ran the sample i am getting this error - AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor'
python 3.8

[thk-cheng]

Just found out a quick fix for this

Assuming you are working inside a virtual environment with python3.8 installed

1. Head to ./lib/python3.8/site-packages/torch/nn/modules/upsampling.py
2. Amend the forward method of the Upsample class as follows:

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

3. Restart the ipykernel and the code should work again

Alternatively, downgrading Pytorch back to <= 1.10.0 would also do the trick. Check out #53 (comment) for more details.

[tobiaswuerth]

or see #53 (comment)

[mooooooood]

You need to select the appropriate pytorch version number according to your own graphics card driver and CUDA version,that's all

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人: "openai/DALL-E" ***@***.***&gt;; 发送时间:&nbsp;2022年4月8日(星期五) 晚上7:29 ***@***.***&gt;; ***@***.***&gt;; 主题:&nbsp;[openai/DALL-E] AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor' (Issue #54) When i install and ran the sample i am getting this error - AttributeError: 'Upsample' object has no attribute 'recompute_scale_factor' python 3.8 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***&gt;

[pzarzycki]

Google Colab default environment throws that error as of May 2022

[Lexaul]

Just found out a quick fix for this

Assuming you are working inside a virtual environment with python3.8 installed

1. Head to ./lib/python3.8/site-packages/torch/nn/modules/upsampling.py
2. Amend the forward method of the Upsample class as follows:

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

3. Restart the ipykernel and the code should work again

Alternatively, downgrading Pytorch back to <= 1.10.0 would also do the trick. Check out #53 (comment) for more details.

This helps, ty

[digitalShaman]

The error
'Upsample' object has no attribute 'recompute_scale_factor'

is related to a change in the torch Upscale class from 1.10 to 1.11.

It appears that 'old' Upscale objects are saved within the model after this line of code:
model = load_model("https://cdn.openai.com/dall-e/decoder.pkl", 'cuda')
i used the following code immediatly after the load_model call to patch this:

# Patch for torch 1.11 and higher: replace the old Upsample object by the new version
# that exposes recompute_scale_factor
_ = model.blocks.group_1.upsample
model.blocks.group_1.upsample = torch.nn.Upsample(scale_factor = _.scale_factor, mode= _.mode)
_ = model.blocks.group_2.upsample
model.blocks.group_2.upsample = torch.nn.Upsample(scale_factor = _.scale_factor, mode= _.mode)
_ = model.blocks.group_3.upsample
model.blocks.group_3.upsample = torch.nn.Upsample(scale_factor = _.scale_factor, mode= _.mode)

and it's running fine with torch 1.12.1!