[Dataset] CC3M

[snkv04]

What does this PR do?

It implements the retrieval script for CC3M and tests that script.

A usage example is at stable_datasets/tests/images/test_cc3m.py.

Note: The training split is somewhat large and can take a while to download even at over 1GB/second of images. Here are the statistics from when I ran it, for reference:

• train split
  • Space:
    • 270GB of space for the downloads
    • 300MB of space for the processed cache
  • Time:
    • 4 hours to download
    • 21 minutes to yield all the examples to build the dataset
• val split
  • Space:
    • 1.4GB of space for the downloads
    • 1.6MB of space for the processed cache
  • Time:
    • 1-2 minutes to download
    • A few seconds to yield all the examples to build the dataset

Before submitting

• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines
• Did you write any new necessary tests?

Who can review?

@RandallBalestriero @Leon-Leyang

[snkv04]

By the way @Leon-Leyang if you want to test this dataset without retrieving the full set of data items so that the tests will run quicker, you can set FIRST_N_IMAGES_PER_SPLIT in cc3m.py to be smaller (e.g., 2000)

[Leon-Leyang]

By the way @Leon-Leyang if you want to test this dataset without retrieving the full set of data items so that the tests will run quicker, you can set FIRST_N_IMAGES_PER_SPLIT in cc3m.py to be smaller (e.g., 2000)

Got it. Thanks!