MMAudio/training/extract_audio_training_latents.py

import logging
import os
from argparse import ArgumentParser
from pathlib import Path

import pandas as pd
import tensordict as td
import torch
import torch.distributed as distributed
import torch.nn.functional as F
from open_clip import create_model_from_pretrained
from torch.utils.data import DataLoader
from tqdm import tqdm

from mmaudio.data.data_setup import error_avoidance_collate
from mmaudio.data.extraction.wav_dataset import WavTextClipsDataset
from mmaudio.ext.autoencoder import AutoEncoderModule
from mmaudio.ext.mel_converter import get_mel_converter

log = logging.getLogger()

torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True

local_rank = int(os.environ['LOCAL_RANK'])
world_size = int(os.environ['WORLD_SIZE'])

# 16k
SAMPLE_RATE = 16_000
NUM_SAMPLES = 16_000 * 8
tod_vae_ckpt = './ext_weights/v1-16.pth'
bigvgan_vocoder_ckpt = './ext_weights/best_netG.pt'
mode = '16k'

# 44k
"""
NOTE: 352800 (8*44100) is not divisible by (STFT hop size * VAE downsampling ratio) which is 1024.
353280 is the next integer divisible by 1024.
"""

# SAMPLE_RATE = 44100
# NUM_SAMPLES = 353280
# tod_vae_ckpt = './ext_weights/v1-44.pth'
# bigvgan_vocoder_ckpt = None
# mode = '44k'


def distributed_setup():
    distributed.init_process_group(backend="nccl")
    local_rank = distributed.get_rank()
    world_size = distributed.get_world_size()
    print(f'Initialized: local_rank={local_rank}, world_size={world_size}')
    return local_rank, world_size


@torch.inference_mode()
def main():
    distributed_setup()

    parser = ArgumentParser()
    parser.add_argument('--data_dir', type=Path, default='./training/example_audios/')
    parser.add_argument('--captions_tsv', type=Path, default='./training/example_audio.tsv')
    parser.add_argument('--clips_tsv', type=Path, default='./training/example_output/clips.tsv')
    parser.add_argument('--latent_dir',
                        type=Path,
                        default='./training/example_output/audio-latents')
    parser.add_argument('--output_dir',
                        type=Path,
                        default='./training/example_output/memmap/audio-example')
    parser.add_argument('--batch_size', type=int, default=32)
    parser.add_argument('--num_workers', type=int, default=8)
    args = parser.parse_args()

    data_dir = args.data_dir
    captions_tsv = args.captions_tsv
    clips_tsv = args.clips_tsv
    latent_dir = args.latent_dir
    output_dir = args.output_dir
    batch_size = args.batch_size
    num_workers = args.num_workers

    clip_model = create_model_from_pretrained('hf-hub:apple/DFN5B-CLIP-ViT-H-14-384',
                                              return_transform=False).eval().cuda()

    # a hack to make it output last hidden states
    def new_encode_text(self, text, normalize: bool = False):
        cast_dtype = self.transformer.get_cast_dtype()

        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]

        x = x + self.positional_embedding.to(cast_dtype)
        x = self.transformer(x, attn_mask=self.attn_mask)
        x = self.ln_final(x)  # [batch_size, n_ctx, transformer.width]
        return F.normalize(x, dim=-1) if normalize else x

    clip_model.encode_text = new_encode_text.__get__(clip_model)

    tod = AutoEncoderModule(vae_ckpt_path=tod_vae_ckpt,
                            vocoder_ckpt_path=bigvgan_vocoder_ckpt,
                            mode=mode).eval().cuda()
    mel_converter = get_mel_converter(mode).eval().cuda()

    dataset = WavTextClipsDataset(data_dir,
                                  captions_tsv=captions_tsv,
                                  clips_tsv=clips_tsv,
                                  sample_rate=SAMPLE_RATE,
                                  num_samples=NUM_SAMPLES,
                                  normalize_audio=True,
                                  reject_silent=True)
    dataloader = DataLoader(dataset,
                            batch_size=batch_size,
                            shuffle=False,
                            num_workers=num_workers,
                            collate_fn=error_avoidance_collate)
    latent_dir.mkdir(exist_ok=True, parents=True)

    # extraction
    for i, batch in tqdm(enumerate(dataloader), total=len(dataloader)):
        ids = batch['id']
        waveforms = batch['waveform'].cuda()
        tokens = batch['tokens'].cuda()

        text_features = clip_model.encode_text(tokens, normalize=True)
        mel = mel_converter(waveforms)
        dist = tod.encode(mel)

        a_mean = dist.mean.detach().cpu().transpose(1, 2)
        a_std = dist.std.detach().cpu().transpose(1, 2)
        text_features = text_features.detach().cpu()

        ids = [id for id in ids]
        captions = [caption for caption in batch['caption']]

        data = {
            'id': ids,
            'caption': captions,
            'mean': a_mean,
            'std': a_std,
            'text_features': text_features,
        }

        torch.save(data, latent_dir / f'r{local_rank}_{i:05d}.pth')

    distributed.barrier()
    # combine the results
    if local_rank == 0:
        print('Extraction done. Combining the results.')

        list_of_ids_and_labels = []
        output_data = {
            'mean': [],
            'std': [],
            'text_features': [],
        }

        latents = sorted(os.listdir(latent_dir))
        latents = [l for l in latents if l.endswith('.pth')]
        for t in tqdm(latents):
            data = torch.load(latent_dir / t, weights_only=True)
            bs = len(data['id'])

            for bi in range(bs):
                this_id = data['id'][bi]
                this_caption = data['caption'][bi]

                list_of_ids_and_labels.append({'id': this_id, 'caption': this_caption})
                output_data['mean'].append(data['mean'][bi])
                output_data['std'].append(data['std'][bi])
                output_data['text_features'].append(data['text_features'][bi])

        output_df = pd.DataFrame(list_of_ids_and_labels)
        output_dir.mkdir(exist_ok=True, parents=True)
        output_name = output_dir.stem
        output_df.to_csv(output_dir.parent / f'{output_name}.tsv', sep='\t', index=False)

        print(f'Output: {len(output_df)}')

        output_data = {k: torch.stack(v) for k, v in output_data.items()}
        td.TensorDict(output_data).memmap_(output_dir)


if __name__ == '__main__':
    main()
    distributed.destroy_process_group()