AudioDecoder imported but not available with documented torchcodec==0.1 install

[julius-richter]

Description

In core/audio_visual_encoder/transforms.py, the code imports:

from torchcodec.decoders import AudioDecoder

However, the documented install instructions specify:

pip install torchcodec==0.1 --index-url=https://download.pytorch.org/whl/cu124

The issue is that torchcodec==0.1 does not provide AudioDecoder.
AudioDecoder was introduced in later versions of TorchCodec, so following the current install instructions leads to an ImportError at runtime when audio processing is enabled.

Workaround / Fix

I replaced TorchCodec-based audio loading with torchaudio, which is already part of the PyTorch stack and avoids the version mismatch.

In core/audio_visual_encoder/transforms.py, I replaced _load_audio with:

import torchaudio

def _load_audio(self, path: str):
    wav, sr = torchaudio.load(path)

    # Convert to mono
    if wav.shape[0] > 1:
        wav = wav.mean(dim=0, keepdim=True)

    # Resample if needed
    if sr != self.sampling_rate:
        wav = torchaudio.functional.resample(wav, sr, self.sampling_rate)

    return wav.contiguous()

This restores compatibility with torchcodec==0.1 while keeping the rest of the audio pipeline unchanged.