How to download the output folder on Kaggle?

Question

I wrote a notebook on Kaggle and imported a dataset.

The main content of the notebook is as follows:

%%bash
pip install xxx # Install dependencies
if [ ! -d "/kaggle/working/latex-ocr-pytorch" ]; then
 echo "Directory does not exist, copying..."
 cp -r /kaggle/input/latex-ocr-pytorch /kaggle/working
else
 echo "Directory already exists, skipping copy"
fi
cd /kaggle/working/latex-ocr-pytorch
python train.py # Train and save the model (save checkpoints)

After that, I ran the notebook using "Save & Run All (Commit)".

After the run completes, multiple checkpoint_xxx.pth.tar files are generated in the /kaggle/working/latex-ocr-pytorch/checkpoints directory. At this point, if I want to download the files in the checkpoints folder, it becomes quite troublesome because there are multiple files, and I have to click download for each one individually.

I tried using the command kaggle kernels output user_name/kaggle-latex-ocr-pytorch -p /path/to/dest to download, but I found that what I downloaded were log files, which is not what I wanted.

So I referred to an online blog and added a code cell at the end of the notebook:

%%bash
cd /kaggle/working/latex-ocr-pytorch/
if [ -d "checkpoints" ]; then
 tar -czf checkpoints.tar.gz checkpoints
 echo "Compression successful"
else
 echo "Warning: checkpoints directory does not exist"
 exit 1
fi

I also noticed that after running the notebook with "Save & Run All (Commit)", when I edit the notebook again, the checkpoints folder under /kaggle/working/latex-ocr-pytorch/ disappears. Therefore, I cannot package the checkpoints folder in a draft session either.

Since the model takes a long time to run and Kaggle has a single-session time limit of 12 hours, the notebook disconnected midway during execution. As a result, the last code cell for packaging the checkpoints did not run. What is the solution to this?

• Sébastien Guarnay: There are different ways to handle a trained model, with save and load, see Kaggle post: kaggle.com/discussions/questions-and-answers/92749#1198429