easytorch-training - ERROR RuntimeError: Expected 2D (unbatched) or 3D (batched) input to conv1d, but got input of size: [8, 32, 207, 13] about step HOT 7 OPEN

graphwavenet.py中的3个nn.conv1d改成2d 就可以了

from step.

这个是torchc版本的问题，推荐使用1.10.0或者1.9.1，这俩我都测试过。更高版本的torch似乎不适配某些算子，就会导致出现这个问题。

from step.

行了运行了一下程序：
(STEP) PS D:\python-venv\STEP> conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c conda-forge

pip install -r requirements.txt

from step.

)
2022-12-07 15:07:14,588 - easytorch-training - INFO - Set lr_scheduler: <torch.optim.lr_scheduler.MultiStepLR object at 0x00000298A3100970>
2022-12-07 15:07:14,594 - easytorch-training - INFO - Initializing validation.
2022-12-07 15:07:14,594 - easytorch-training - INFO - Building val data loader.
val len: 3425
test len: 6850
2022-12-07 15:07:14,770 - easytorch-training - INFO - Epoch 1 / 100
31%|████████████████████████████████████████████████████████████████▌ | 935/2997 [06:52<14:51, 2.31it/s]

from step.

似乎已经顺利运行。

from step.

After torch 1.10, the kernel size of torch.nn.conv1d is restricted to 1d, which used to could take tuple. But intuitively, when the kernel is (1,1) 2D, it should be conv2d not conv1d, and pytorch changed that after 1.10.

The fix will be changing
nn.Conv1d(in_channels=dilation_channels, out_channels=skip_channels, kernel_size=(1, 1))
to
nn.Conv2d(in_channels=dilation_channels, out_channels=skip_channels, kernel_size=(1, 1))

from step.

请问除了降版本没有其他方法了吗

from step.