Saving the upproj model fails about sparse-to-dense.pytorch

[fangchangma]

Hi. I've only tested it on PyTorch 0.3.1 and 0.4.0 with Python 3.5. What's your pytorch version?

[4bitadder]

Hi

I tried in CentOS, Pytorch 0.1.12, and python 2.7.

Since this old version does not have the 'nn.Upsample' layer, I use 'nn.UpsamplingBilinear2d' instead. It works well for training and testing with deconv, e.g.,

python main.py -a resnet50 -d deconv3 -m rgbd

[fangchangma]

Unfortunately, it is very hard to maintain backward compatibility for PyTorch 0.1.x. Could you try upgrading your PyTorch to at least 0.3.1 and see if the problem remains?

[4bitadder]

Thanks. After swithing to Python 3.5 and Pytorch 0.4.0, the saving procedure works fine.

But I have to disable the saving csv part in main.py, sometimes the code crashed after writing csv several times (e.g., around epoch 7-10). Not clear for the reason.

[fangchangma]

Do you mind sharing the error messages?