mit-han-lab / litepose

[CVPR'22] Lite Pose: Efficient Architecture Design for 2D Human Pose Estimation
https://hanlab.mit.edu
MIT License
304 stars 37 forks source link

missing 1 required argument "num_spatial_dims" #28

Open wanghao-cst opened 1 year ago

wanghao-cst commented 1 year ago

Hi, when I run the first command to train the supermobile model, threre is an error as shown in the pict. Maybe it is caused by the wrong torch version, my version is 1.12, could u ple have a check? 微信图片_20221103095942

jxhnx commented 1 year ago

Try replacing this line by: output_padding = self._output_padding(x, output_size, self.stride, self.padding, self.kernel_size, num_spatial_dims=2)

It's due to a PyTorch update, at some point the argument wasn't required and 2D the default as far as I remember.

wanghao-cst commented 1 year ago

Thx, that works. may I know your torch version?

jxhnx commented 1 year ago

I'm using some nightlies, 1.13.x 1.14.x, that's why I had to adjust that, too. I never trained the supernet though, so please apply that with caution. With subnets, all works as expected on my end.