add CLEVR dataset

[pmeier]

Addresses #5108.

cc @pmeier @bjuncek

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 35f464c (more details on the Dr. CI page):

---

✅ None of the CI failures appear to be your fault 💚

---

• 5/5 broken upstream at merge base 7120024 since Jan 04

---

🚧 5 ongoing upstream failures:

These were probably caused by upstream breakages that are not fixed yet.

• unittest_linux_cpu_py3.9 since Jan 04 (cc7e856)
  • 🔁 rerun
• unittest_macos_cpu_py3.8 since Jan 04 (cc7e856)
  • 🔁 rerun
• unittest_linux_cpu_py3.8 since Jan 04 (cc7e856)
  • 🔁 rerun
• unittest_linux_cpu_py3.7 since Jan 04 (cc7e856)
  • 🔁 rerun
• unittest_macos_cpu_py3.9 since Jan 04 (cc7e856)
  • 🔁 rerun

---

This comment was automatically generated by Dr. CI (expand for details).

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

[NicolasHug]

Thanks @pmeier , I made a few comments below but this looks good

[NicolasHug]

Could you help me understand why this change is needed?

[pmeier]

This enables chained getattr calls. In turn, this allows me to use this filter function

images_dp = Filter(images_dp, path_comparator("parent.name", config.split))

rather than writing a custom function that extracts the name of the parent folder.

[NicolasHug]

I'll leave it up to you to decide, as the entire logic of path_comparator and path_accessor is honestly too complex for me to handle ATM.

But as rule of thumb, I tend to avoid changes to helper functions when they just only address one single use-case.

[pmeier]

as the entire logic of path_comparator and path_accessor is honestly too complex for me to handle ATM.

The complexity stems from the fact that we cannot use lambdas or local function together with the datapipes, since they cannot be serialized safely. Quick explanation:

• path_accessor: When using datapipes you often have path-handle-tuples for files. path_accessor takes such a tuple, turns the the first item into a pathlib.Path and accesses it according to the input. So path_accessor("parent.name") is equivalent to

  def extract_folder_name(data):
      path = pathlib.Path(data[0])
      return path.parent.name

  This is useful when you want to merge two datapipes and need a function that generates merge keys.

• path_comparator: This is one layer on top of path_accessor by providing an equality comparison for the extract path information. For example Filter(scenes_dp, path_comparator("name", f"CLEVR_{config.split}_scenes.json")) will select all files where the file name matches the second argument. (This is what I used to refactor your suggestion about the regex).

But as rule of thumb, I tend to avoid changes to helper functions when they just only address one single use-case.

It is true, that this currently only addresses this, but I already needed to make the same changes to getitem a while back. So I'm guessing if I don't do it now, I'll have to do it in the future anyway, but then I also have to remember that I need to go back to here and fix this.

[NicolasHug]

Thanks a lot @pmeier , LGTM!