Let torchaudio.load() and torchaudio.save() rely on load_with_torchcodec() and save_with_torchcodec().

[samanklesaria]

This PR wraps the load_with_torchcodec and save_with_torchcodec functions with functions of the name load and save so that code that depends on the old load and save functions can continue to work in the future once we remove backend-specific code.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/audio/4039

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 8 New Failures

As of commit 498ce49 with merge base 02351a6 ():

NEW FAILURES - The following jobs have failed:

• .github/workflows/bandit.yml (gh)
• .github/workflows/integration-test.yml (gh)
• .github/workflows/lint.yml (gh)
• .github/workflows/pr-labels.yml (gh)
• .github/workflows/unittest-macos-cpu.yml (gh)
• .github/workflows/unittest-windows-cpu.yml (gh)
• .github/workflows/unittest-windows-gpu.yml (gh)
• Build documentation / build / Build doc (gh)
  At least one of the pre-conditions you specified did not hold

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[samanklesaria]

Installing ffmpeg>4, which is necessary for torchcodec to be able to load files used during testing, seems to be incompatible with the current CI infrastructure. Perhaps we need a separate PR to install ffmpeg>4, and wait for the infrastructure to improve so that that PR can be merged.

[samanklesaria]

I'm coming across an interesting discrepancy in the load behavior here. In test_load_save_torchcodec.py, the function test_save_channels_first creates a random tensor called waveform. When I try it, I get waveform[0,1345] = -4.0688. Saving and loading the result with scipy gives waveform_ta[0,1345] = -4.0688. But saving with torchcodec and loading with scipy gives waveform[0,1345] = -1.. What exactly is happening here? Are we supposed to be truncating saved values between -1 and 1? Is the normalization that torchcodec is doing different than the normalization scipy does?
I guess the docstring for save_with_torchcodec says "TorchCodec AudioEncoder expects float32 samples in [-1, 1] range". Does this mean the behavior of this test is undefined when our random waveform has larger values?

An easy fix here is just to use rand instead of randn to initialize our random tensor.

[NicolasHug]

Your assessment on test_load_save_torchcodec.py are correct, it's would have been best to ensure that input values of the encoder are in [-1, 1] rather than to use randn.

However, this entire file is meant to test the old torchaudio load() and save() against load_with_torchcodec() and save_with_torchcodec()

Now that we are moving load() and save() to rely on load_with_torchcodec() and save_with_torchcodec() directly, these tests aren't worth running anymore. Even more, we are switching torchcodec for scipy - so these tests are not testing what they were meant to test anymore.

Basically, we can just skip them safely. Just add a big pytest.skip() statement at the top, indicating this test file was only relevant back when load_with_torchcodec() and save_with_torchcodec() were introduced and load() and save() still had their own implementation.

[samanklesaria]

Now we're failing because core dumped) python -c "import torch; import torchaudio; import torchcodec; print(torch.__version__, torchaudio.__version__, torchcodec.__version__)" fails. Because I never defined __version__ in my mock module. But do I need it? It might be easiest to remove this check in the install script.

[samanklesaria]

I'm going to remove the installation of torchcodec during testing, as it shouldn't be used anyway. We should rely on the mock.

[NicolasHug]

sys doesn't seem to be used