I just created a package called torcheck that help you carry out sanity check around your model. It has received several stars, so I wanted to share with the community.
For a general introduction, please check this out: Testing Your PyTorch Models with Torcheck
The GitHub link is: GitHub - pengyan510/torcheck
The major benefit is that you no longer need to write additional testing code for your model training. Just add a few lines of code specifying the checks before training, torcheck will then take over and perform the checks simultaneouly while the training happens.
Another benefit is that torcheck allows you to check your model on different levels. Instead of checking the whole model, you can specify checks for a submodule, a linear layer, or even the weight tensor! This enables more customization around the sanity checks.