Datasets: `download_and_prepare` when developing an in-development dataset redownloads data every time

[rsepassi]

Short description
When developing a new dataset, it seems that the download_and_preapre redownloads the data every time.

Environment information

• Operating System: Ubuntu
• Python version: 3.6
• tensorflow-datasets/tfds-nightly version: tfds-nightly==1.0.1.dev201903040105
• tensorflow/tensorflow-gpu/tf-nightly/tf-nightly-gpu version: tensorflow==1.13.1

Reproduction instructions

# Change the MNIST class name in the code so that it seems like it's a new dataset
# class MNISTZ

# Run once to see the data downloading, and kill it before it finishes
# Then run it again and see that it redownloads
python -m tensorflow_datasets.scripts.download_and_prepare --datasets=mnistz

Expected behavior
The downloads to be cached.

[adarob]

Looking at the code, it seems that the sha256 hashes are only stored after the dataset building completes, which causes the files to be re-downloaded every time. It would be very nice if this was relaxed, at least for development.

We could simply add the sha to the INFO file.

[ubershmekel]

This is really slowing me down :(

[Conchylicultor]

The current hack is to locally change the https://github.com/tensorflow/datasets/blob/master/tensorflow_datasets/core/download/download_manager.py#L210:
From:

if (not self._force_download and resource.sha256 and
    resource.exists_locally()):

To:

if not self._force_download and resource.exists_locally():

But it has side effect elsewhere (like not computing checksums anymore).

Waiting for @pierrot0 for a more stable fix.

[jackd]

Is the proposition to cache on the builder level, or the data_dir level? Is there any reason mnistz from above would not want to use downloaded data from mnist, given the url is the same?

[pierrot0]

The proposition is to cache globally.
Path will land soon.