Bug report: Unable to Interpolate Models

[infinitemonkeytheory]

Information:

Error

An error occurred in a PyTorch Interpolate Models node:

Input and output sizes should be greater than 0, but got input (H: 1, W: 1) output (H: 0, W: 0)

ChaiNNer: chaiNNer 0.15.2 OS: macOS 10.18

Description When attempting to interpolate models the above error is thrown.

Logs

[2022-11-24 14:56:57.864] [info]  Backend: [67906] [ERROR] Traceback (most recent call last):
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 344, in __process
    raw_output, execution_time = await self.loop.run_in_executor(
  File "/Users/imt/Library/Application Support/chaiNNer/python/python/lib/python3.9/concurrent/futures/thread.py", line 58, in run
    result = self.fn(*self.args, **self.kwargs)
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 222, in wrapper
    result = supplier()
  File "/Applications/chaiNNer.app/Contents/Resources/src/nodes/nodes/pytorch/interpolate_models.py", line 101, in run
    if not self.check_can_interp(state_a, state_b):
  File "/Applications/chaiNNer.app/Contents/Resources/src/nodes/nodes/pytorch/interpolate_models.py", line 81, in check_can_interp
    t_out = model(img_tensor)
  File "/Users/imt/Library/Application Support/chaiNNer/python/python/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/Applications/chaiNNer.app/Contents/Resources/src/nodes/utils/architecture/GFPGAN/gfpganv1_clean_arch.py", line 335, in forward
    feat = self.conv_body_down[i](feat)
  File "/Users/imt/Library/Application Support/chaiNNer/python/python/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/Applications/chaiNNer.app/Contents/Resources/src/nodes/utils/architecture/GFPGAN/gfpganv1_clean_arch.py", line 162, in forward
    out = F.interpolate(
  File "/Users/imt/Library/Application Support/chaiNNer/python/python/lib/python3.9/site-packages/torch/nn/functional.py", line 3731, in interpolate
    return torch._C._nn.upsample_bilinear2d(input, output_size, align_corners, scale_factors)
RuntimeError: Input and output sizes should be greater than 0, but got input (H: 1, W: 1) output (H: 0, W: 0)

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
  File "/Applications/chaiNNer.app/Contents/Resources/src/run.py", line 243, in run
    await executor.run()
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 487, in run
    await self.__process_nodes(self.__get_output_nodes())
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 475, in __process_nodes
    await self.process(output_node)
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 282, in process
    return await self.__process(node)
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 309, in __process
    processed_input = await self.process(node_input.id)
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 282, in process
    return await self.__process(node)
  File "/Applications/chaiNNer.app/Contents/Resources/src/process.py", line 365, in __process
    raise NodeExecutionError(node, str(e), input_dict) from e
process.NodeExecutionError: Input and output sizes should be greater than 0, but got input (H: 1, W: 1) output (H: 0, W: 0)

[joeyballentine]

I assume this was with SwinIR, Swin2SR, or hat?

[infinitemonkeytheory]

I assume this was with SwinIR, Swin2SR, or hat?

I should've included more data. I started out seeing about merging various GFPGAN models (out of curiosity), which is where I hit above error. And as predicted with, e.g., SwinIR or HAT I get essentially what I assume (perhaps totally incorrectly) is a similar error to above:

Argument #4: Padding size should be less than the corresponding input dimension, but got: padding (0, 5) at dimension 3 of input 4

and

Argument #4: Padding size should be less than the corresponding input dimension, but got: padding (0, 13) at dimension 3 of input 4

Real-ESRGAN Compact works as expected (e.g. tested merging realesr-general-x4v3.pth with realesr-general-wdn-x4v3.pth and encountered no issues plus was able to successfully use merged model).

I'm afraid that's all I've tested so far.

[joeyballentine]

Yeah then i know why this happens. Since interpolation between models that don't share a pretrained model produces broken models i do a small 3x3 upscale test to make sure it doesn't spit out a garbage result. However gfpgan requires a 512x512 image no matter what, and a similar thing is needed for those other models where they have size restrictions. I just need to update the interpolation check to either not happen for those models or simply use the correct sizes.