vendored mmaudio + weights from our own namespace + explicit api_names

README.md

@@ -1,13 +1,14 @@
 ---
 title: Rill Lab
-emoji: 🏃
-colorFrom: purple
-colorTo: blue
+emoji: 🎚️
+colorFrom: gray
+colorTo: gray
 sdk: gradio
-sdk_version: 6.22.0
-python_version: '3.12'
+sdk_version: 5.44.1
 app_file: app.py
 pinned: false
+short_description: Audio track lab
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+Generates an audio track for a supplied clip. Endpoints: `v2a` (clip in, clip with track out),
+`v2track` (clip in, track only), `t2a`, `i2a`.

app.py

@@ -0,0 +1,273 @@
+import gc
+import logging
+from argparse import ArgumentParser
+from datetime import datetime
+from fractions import Fraction
+from pathlib import Path
+
+import gradio as gr
+import soundfile as sf
+import spaces
+import torch
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file
+
+from mmaudio.eval_utils import (ModelConfig, VideoInfo, all_model_cfg, generate, load_image,
+                                load_video, make_video, setup_eval_logging)
+from mmaudio.model.flow_matching import FlowMatching
+from mmaudio.model.networks import MMAudio, get_my_mmaudio
+from mmaudio.model.sequence_config import SequenceConfig
+from mmaudio.model.utils.features_utils import FeaturesUtils
+
+logging.getLogger("httpx").setLevel(logging.WARNING)
+logging.getLogger("requests").setLevel(logging.WARNING)
+logging.getLogger("urllib3").setLevel(logging.WARNING)
+
+torch.backends.cuda.matmul.allow_tf32 = True
+torch.backends.cudnn.allow_tf32 = True
+
+log = logging.getLogger()
+
+device = 'cuda' if torch.cuda.is_available() else 'cpu'
+dtype = torch.float32
+
+# Every weight comes from our own namespace; the CLIP encoder and the 44kHz vocoder are repointed
+# inside the vendored `mmaudio` package (features_utils.py / ext/autoencoder.py).
+EXT_REPO = 'thornmaze/mma-ext'
+CORE_REPO = 'thornmaze/mma-core-fp16'
+CORE_FILE = 'mmaudio_large_44k_nsfw_gold_8.5k_final_fp16.safetensors'
+
+# Architecture is large_44k (v1); only the sequence/mode config is taken from the v2 entry.
+MY_MODEL_NAME = 'large_44k'
+
+vae_path = hf_hub_download(repo_id=EXT_REPO, subfolder='ext_weights', filename='v1-44.pth')
+synchformer_path = hf_hub_download(repo_id=EXT_REPO,
+                                   subfolder='ext_weights',
+                                   filename='synchformer_state_dict.pth')
+
+model_cfg_for_params: ModelConfig = all_model_cfg['large_44k_v2']
+
+output_dir = Path('./output/gradio')
+setup_eval_logging()
+
+
+def get_model() -> tuple[MMAudio, FeaturesUtils, SequenceConfig]:
+    seq_cfg = model_cfg_for_params.seq_cfg
+
+    net: MMAudio = get_my_mmaudio(MY_MODEL_NAME).to(device, dtype).eval()
+    weights_path = hf_hub_download(repo_id=CORE_REPO, filename=CORE_FILE)
+    state_dict = load_file(weights_path)
+
+    missing, unexpected = net.load_state_dict(state_dict, strict=False)
+    if missing:
+        log.warning(f"Missing keys: {missing[:8]}{'...' if len(missing) > 8 else ''}")
+    if unexpected:
+        log.warning(f"Unexpected keys: {unexpected[:8]}{'...' if len(unexpected) > 8 else ''}")
+
+    if dtype == torch.float16:
+        net.half()
+    net.to(device).eval()
+
+    # bigvgan_vocoder_ckpt=None -> the 44kHz BigVGANv2 is pulled from our mirror by the vendored
+    # autoencoder; need_vae_encoder=False because inference only decodes.
+    feature_utils = FeaturesUtils(tod_vae_ckpt=vae_path,
+                                  synchformer_ckpt=synchformer_path,
+                                  enable_conditions=True,
+                                  mode=model_cfg_for_params.mode,
+                                  bigvgan_vocoder_ckpt=None,
+                                  need_vae_encoder=False)
+    feature_utils = feature_utils.to(device, dtype).eval()
+
+    return net, feature_utils, seq_cfg
+
+
+net, feature_utils, seq_cfg = get_model()
+
+
+def _rng(seed: int) -> torch.Generator:
+    rng = torch.Generator(device=device)
+    if seed >= 0:
+        rng.manual_seed(seed)
+    else:
+        rng.seed()
+    return rng
+
+
+def _stamp() -> str:
+    return datetime.now().strftime('%Y%m%d_%H%M%S_%f')
+
+
+def _audio_for_video(video, prompt: str, negative_prompt: str, seed: int, num_steps: int,
+                     cfg_strength: float, duration: float):
+    """Shared core of both video paths: decoded video -> generated waveform."""
+    fm = FlowMatching(min_sigma=0, inference_mode='euler', num_steps=num_steps)
+    video_info = load_video(video, duration)
+    clip_frames = video_info.clip_frames.unsqueeze(0)
+    sync_frames = video_info.sync_frames.unsqueeze(0)
+    seq_cfg.duration = video_info.duration_sec
+    net.update_seq_lengths(seq_cfg.latent_seq_len, seq_cfg.clip_seq_len, seq_cfg.sync_seq_len)
+
+    audios = generate(clip_frames,
+                      sync_frames, [prompt],
+                      negative_text=[negative_prompt],
+                      feature_utils=feature_utils,
+                      net=net,
+                      fm=fm,
+                      rng=_rng(seed),
+                      cfg_strength=cfg_strength)
+    return video_info, audios.float().cpu()[0]
+
+
+@spaces.GPU(duration=90)
+@torch.inference_mode()
+def video_to_audio(video: gr.Video, prompt: str, negative_prompt: str, seed: int, num_steps: int,
+                   cfg_strength: float, duration: float):
+    video_info, audio = _audio_for_video(video, prompt, negative_prompt, seed, num_steps,
+                                         cfg_strength, duration)
+    output_dir.mkdir(exist_ok=True, parents=True)
+    video_save_path = output_dir / f'{_stamp()}.mp4'
+    make_video(video_info, video_save_path, audio, sampling_rate=seq_cfg.sampling_rate)
+    gc.collect()
+    return video_save_path
+
+
+@spaces.GPU(duration=90)
+@torch.inference_mode()
+def video_to_track(video: gr.Video, prompt: str, negative_prompt: str, seed: int, num_steps: int,
+                   cfg_strength: float, duration: float):
+    """Same generation, but returns the bare audio track instead of a re-encoded mp4.
+
+    `make_video` re-encodes every frame (measured: a 1.4 MB source came back at 6.5 MB with a
+    second lossy pass). The caller muxes this track onto its own untouched video stream.
+    """
+    _, audio = _audio_for_video(video, prompt, negative_prompt, seed, num_steps, cfg_strength,
+                                duration)
+    output_dir.mkdir(exist_ok=True, parents=True)
+    audio_save_path = output_dir / f'{_stamp()}.flac'
+    # torchaudio.save routes through torchcodec, which is not installed on Spaces
+    sf.write(audio_save_path, audio.transpose(0, 1).numpy(), seq_cfg.sampling_rate)
+    gc.collect()
+    return audio_save_path
+
+
+@spaces.GPU(duration=90)
+@torch.inference_mode()
+def image_to_audio(image: gr.Image, prompt: str, negative_prompt: str, seed: int, num_steps: int,
+                   cfg_strength: float, duration: float):
+    fm = FlowMatching(min_sigma=0, inference_mode='euler', num_steps=num_steps)
+    image_info = load_image(image)
+    clip_frames = image_info.clip_frames.unsqueeze(0)
+    sync_frames = image_info.sync_frames.unsqueeze(0)
+    seq_cfg.duration = duration
+    net.update_seq_lengths(seq_cfg.latent_seq_len, seq_cfg.clip_seq_len, seq_cfg.sync_seq_len)
+
+    audios = generate(clip_frames,
+                      sync_frames, [prompt],
+                      negative_text=[negative_prompt],
+                      feature_utils=feature_utils,
+                      net=net,
+                      fm=fm,
+                      rng=_rng(seed),
+                      cfg_strength=cfg_strength,
+                      image_input=True)
+    audio = audios.float().cpu()[0]
+
+    output_dir.mkdir(exist_ok=True, parents=True)
+    video_save_path = output_dir / f'{_stamp()}.mp4'
+    video_info = VideoInfo.from_image_info(image_info, duration, fps=Fraction(1))
+    make_video(video_info, video_save_path, audio, sampling_rate=seq_cfg.sampling_rate)
+    gc.collect()
+    return video_save_path
+
+
+@spaces.GPU(duration=45)
+@torch.inference_mode()
+def text_to_audio(prompt: str, negative_prompt: str, seed: int, num_steps: int, cfg_strength: float,
+                  duration: float):
+    fm = FlowMatching(min_sigma=0, inference_mode='euler', num_steps=num_steps)
+    seq_cfg.duration = duration
+    net.update_seq_lengths(seq_cfg.latent_seq_len, seq_cfg.clip_seq_len, seq_cfg.sync_seq_len)
+
+    audios = generate(None,
+                      None, [prompt],
+                      negative_text=[negative_prompt],
+                      feature_utils=feature_utils,
+                      net=net,
+                      fm=fm,
+                      rng=_rng(seed),
+                      cfg_strength=cfg_strength)
+    audio = audios.float().cpu()[0]
+
+    output_dir.mkdir(exist_ok=True, parents=True)
+    audio_save_path = output_dir / f'{_stamp()}.flac'
+    sf.write(audio_save_path, audio.transpose(0, 1).numpy(), seq_cfg.sampling_rate)
+    gc.collect()
+    return audio_save_path
+
+
+def _params(with_negative_default: bool) -> list:
+    return [
+        gr.Text(label='Prompt'),
+        gr.Text(label='Negative prompt', value='music' if with_negative_default else ''),
+        gr.Number(label='Seed (-1: random)', value=-1, precision=0, minimum=-1),
+        gr.Number(label='Num steps', value=25, precision=0, minimum=1),
+        gr.Number(label='Guidance Strength', value=4.5, minimum=1),
+        gr.Number(label='Duration (sec)', value=8, minimum=1),
+    ]
+
+
+# api_name is set explicitly on every tab: the default names endpoints by tab ORDER
+# (predict, predict_1, ...), so adding or reordering a tab silently re-points every caller.
+video_to_audio_tab = gr.Interface(
+    fn=video_to_audio,
+    api_name='v2a',
+    inputs=[gr.Video(), *_params(True)],
+    outputs='playable_video',
+    cache_examples=False,
+    title='Video to Audio',
+    description='Resolutions above 384 px on the shorter side cost time without improving output.',
+)
+
+video_to_track_tab = gr.Interface(
+    fn=video_to_track,
+    api_name='v2track',
+    inputs=[gr.Video(), *_params(True)],
+    outputs='audio',
+    cache_examples=False,
+    title='Video to Audio track',
+    description='Returns the generated track only — the source video is never re-encoded.',
+)
+
+text_to_audio_tab = gr.Interface(
+    fn=text_to_audio,
+    api_name='t2a',
+    inputs=_params(False),
+    outputs='audio',
+    cache_examples=False,
+    title='Text to Audio',
+)
+
+image_to_audio_tab = gr.Interface(
+    fn=image_to_audio,
+    api_name='i2a',
+    inputs=[gr.Image(type='filepath'), *_params(False)],
+    outputs='playable_video',
+    cache_examples=False,
+    title='Image to Audio (experimental)',
+)
+
+app = gr.TabbedInterface(
+    [video_to_audio_tab, video_to_track_tab, text_to_audio_tab, image_to_audio_tab],
+    ['Video-to-Audio', 'Video-to-Track', 'Text-to-Audio', 'Image-to-Audio'],
+)
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument('--port', type=int, default=7860)
+    parser.add_argument('--share', action='store_true', help='Create a public link')
+    args = parser.parse_args()
+
+    app.launch(server_name="0.0.0.0",
+               server_port=args.port,
+               share=args.share,
+               allowed_paths=[output_dir])

mmaudio/data/av_utils.py

@@ -0,0 +1,162 @@
+from dataclasses import dataclass
+from fractions import Fraction
+from pathlib import Path
+from typing import Optional
+
+import av
+import numpy as np
+import torch
+from av import AudioFrame
+
+
+@dataclass
+class VideoInfo:
+    duration_sec: float
+    fps: Fraction
+    clip_frames: torch.Tensor
+    sync_frames: torch.Tensor
+    all_frames: Optional[list[np.ndarray]]
+
+    @property
+    def height(self):
+        return self.all_frames[0].shape[0]
+
+    @property
+    def width(self):
+        return self.all_frames[0].shape[1]
+
+    @classmethod
+    def from_image_info(cls, image_info: 'ImageInfo', duration_sec: float,
+                        fps: Fraction) -> 'VideoInfo':
+        num_frames = int(duration_sec * fps)
+        all_frames = [image_info.original_frame] * num_frames
+        return cls(duration_sec=duration_sec,
+                   fps=fps,
+                   clip_frames=image_info.clip_frames,
+                   sync_frames=image_info.sync_frames,
+                   all_frames=all_frames)
+
+
+@dataclass
+class ImageInfo:
+    clip_frames: torch.Tensor
+    sync_frames: torch.Tensor
+    original_frame: Optional[np.ndarray]
+
+    @property
+    def height(self):
+        return self.original_frame.shape[0]
+
+    @property
+    def width(self):
+        return self.original_frame.shape[1]
+
+
+def read_frames(video_path: Path, list_of_fps: list[float], start_sec: float, end_sec: float,
+                need_all_frames: bool) -> tuple[list[np.ndarray], list[np.ndarray], Fraction]:
+    output_frames = [[] for _ in list_of_fps]
+    next_frame_time_for_each_fps = [0.0 for _ in list_of_fps]
+    time_delta_for_each_fps = [1 / fps for fps in list_of_fps]
+    all_frames = []
+
+    # container = av.open(video_path)
+    with av.open(video_path) as container:
+        stream = container.streams.video[0]
+        fps = stream.guessed_rate
+        stream.thread_type = 'AUTO'
+        for packet in container.demux(stream):
+            for frame in packet.decode():
+                frame_time = frame.time
+                if frame_time < start_sec:
+                    continue
+                if frame_time > end_sec:
+                    break
+
+                frame_np = None
+                if need_all_frames:
+                    frame_np = frame.to_ndarray(format='rgb24')
+                    all_frames.append(frame_np)
+
+                for i, _ in enumerate(list_of_fps):
+                    this_time = frame_time
+                    while this_time >= next_frame_time_for_each_fps[i]:
+                        if frame_np is None:
+                            frame_np = frame.to_ndarray(format='rgb24')
+
+                        output_frames[i].append(frame_np)
+                        next_frame_time_for_each_fps[i] += time_delta_for_each_fps[i]
+
+    output_frames = [np.stack(frames) for frames in output_frames]
+    return output_frames, all_frames, fps
+
+
+def reencode_with_audio(video_info: VideoInfo, output_path: Path, audio: torch.Tensor,
+                        sampling_rate: int):
+    container = av.open(output_path, 'w')
+    output_video_stream = container.add_stream('h264', video_info.fps)
+    output_video_stream.codec_context.bit_rate = 10 * 1e6  # 10 Mbps
+    output_video_stream.width = video_info.width
+    output_video_stream.height = video_info.height
+    output_video_stream.pix_fmt = 'yuv420p'
+
+    output_audio_stream = container.add_stream('aac', sampling_rate)
+
+    # encode video
+    for image in video_info.all_frames:
+        image = av.VideoFrame.from_ndarray(image)
+        packet = output_video_stream.encode(image)
+        container.mux(packet)
+
+    for packet in output_video_stream.encode():
+        container.mux(packet)
+
+    # convert float tensor audio to numpy array
+    audio_np = audio.numpy().astype(np.float32)
+    audio_frame = AudioFrame.from_ndarray(audio_np, format='flt', layout='mono')
+    audio_frame.sample_rate = sampling_rate
+
+    for packet in output_audio_stream.encode(audio_frame):
+        container.mux(packet)
+
+    for packet in output_audio_stream.encode():
+        container.mux(packet)
+
+    container.close()
+
+
+def remux_with_audio(video_path: Path, audio: torch.Tensor, output_path: Path, sampling_rate: int):
+    """
+    NOTE: I don't think we can get the exact video duration right without re-encoding
+    so we are not using this but keeping it here for reference
+    """
+    video = av.open(video_path)
+    output = av.open(output_path, 'w')
+    input_video_stream = video.streams.video[0]
+    output_video_stream = output.add_stream(template=input_video_stream)
+    output_audio_stream = output.add_stream('aac', sampling_rate)
+
+    duration_sec = audio.shape[-1] / sampling_rate
+
+    for packet in video.demux(input_video_stream):
+        # We need to skip the "flushing" packets that `demux` generates.
+        if packet.dts is None:
+            continue
+        # We need to assign the packet to the new stream.
+        packet.stream = output_video_stream
+        output.mux(packet)
+
+    # convert float tensor audio to numpy array
+    audio_np = audio.numpy().astype(np.float32)
+    audio_frame = av.AudioFrame.from_ndarray(audio_np, format='flt', layout='mono')
+    audio_frame.sample_rate = sampling_rate
+
+    for packet in output_audio_stream.encode(audio_frame):
+        output.mux(packet)
+
+    for packet in output_audio_stream.encode():
+        output.mux(packet)
+
+    video.close()
+    output.close()
+
+    output.close()

mmaudio/data/data_setup.py

@@ -0,0 +1,174 @@
+import logging
+import random
+
+import numpy as np
+import torch
+from omegaconf import DictConfig
+from torch.utils.data import DataLoader, Dataset
+from torch.utils.data.dataloader import default_collate
+from torch.utils.data.distributed import DistributedSampler
+
+from mmaudio.data.eval.audiocaps import AudioCapsData
+from mmaudio.data.eval.video_dataset import MovieGen, VGGSound
+from mmaudio.data.extracted_audio import ExtractedAudio
+from mmaudio.data.extracted_vgg import ExtractedVGG
+from mmaudio.data.mm_dataset import MultiModalDataset
+from mmaudio.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+
+# Re-seed randomness every time we start a worker
+def worker_init_fn(worker_id: int):
+    worker_seed = torch.initial_seed() % (2**31) + worker_id + local_rank * 1000
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)
+    log.debug(f'Worker {worker_id} re-seeded with seed {worker_seed} in rank {local_rank}')
+
+
+def load_vgg_data(cfg: DictConfig, data_cfg: DictConfig) -> Dataset:
+    dataset = ExtractedVGG(tsv_path=data_cfg.tsv,
+                           data_dim=cfg.data_dim,
+                           premade_mmap_dir=data_cfg.memmap_dir)
+
+    return dataset
+
+
+def load_audio_data(cfg: DictConfig, data_cfg: DictConfig) -> Dataset:
+    dataset = ExtractedAudio(tsv_path=data_cfg.tsv,
+                             data_dim=cfg.data_dim,
+                             premade_mmap_dir=data_cfg.memmap_dir)
+
+    return dataset
+
+
+def setup_training_datasets(cfg: DictConfig) -> tuple[Dataset, DistributedSampler, DataLoader]:
+    if cfg.mini_train:
+        vgg = load_vgg_data(cfg, cfg.data.ExtractedVGG_val)
+        audiocaps = load_audio_data(cfg, cfg.data.AudioCaps)
+        dataset = MultiModalDataset([vgg], [audiocaps])
+    if cfg.example_train:
+        video = load_vgg_data(cfg, cfg.data.Example_video)
+        audio = load_audio_data(cfg, cfg.data.Example_audio)
+        dataset = MultiModalDataset([video], [audio])
+    else:
+        # load the largest one first
+        freesound = load_audio_data(cfg, cfg.data.FreeSound)
+        vgg = load_vgg_data(cfg, cfg.data.ExtractedVGG)
+        audiocaps = load_audio_data(cfg, cfg.data.AudioCaps)
+        audioset_sl = load_audio_data(cfg, cfg.data.AudioSetSL)
+        bbcsound = load_audio_data(cfg, cfg.data.BBCSound)
+        clotho = load_audio_data(cfg, cfg.data.Clotho)
+        dataset = MultiModalDataset([vgg] * cfg.vgg_oversample_rate,
+                                    [audiocaps, audioset_sl, bbcsound, freesound, clotho])
+
+    batch_size = cfg.batch_size
+    num_workers = cfg.num_workers
+    pin_memory = cfg.pin_memory
+    sampler, loader = construct_loader(dataset,
+                                       batch_size,
+                                       num_workers,
+                                       shuffle=True,
+                                       drop_last=True,
+                                       pin_memory=pin_memory)
+
+    return dataset, sampler, loader
+
+
+def setup_test_datasets(cfg):
+    dataset = load_vgg_data(cfg, cfg.data.ExtractedVGG_test)
+
+    batch_size = cfg.batch_size
+    num_workers = cfg.num_workers
+    pin_memory = cfg.pin_memory
+    sampler, loader = construct_loader(dataset,
+                                       batch_size,
+                                       num_workers,
+                                       shuffle=False,
+                                       drop_last=False,
+                                       pin_memory=pin_memory)
+
+    return dataset, sampler, loader
+
+
+def setup_val_datasets(cfg: DictConfig) -> tuple[Dataset, DataLoader, DataLoader]:
+    if cfg.example_train:
+        dataset = load_vgg_data(cfg, cfg.data.Example_video)
+    else:
+        dataset = load_vgg_data(cfg, cfg.data.ExtractedVGG_val)
+
+    val_batch_size = cfg.batch_size
+    val_eval_batch_size = cfg.eval_batch_size
+    num_workers = cfg.num_workers
+    pin_memory = cfg.pin_memory
+    _, val_loader = construct_loader(dataset,
+                                     val_batch_size,
+                                     num_workers,
+                                     shuffle=False,
+                                     drop_last=False,
+                                     pin_memory=pin_memory)
+    _, eval_loader = construct_loader(dataset,
+                                      val_eval_batch_size,
+                                      num_workers,
+                                      shuffle=False,
+                                      drop_last=False,
+                                      pin_memory=pin_memory)
+
+    return dataset, val_loader, eval_loader
+
+
+def setup_eval_dataset(dataset_name: str, cfg: DictConfig) -> tuple[Dataset, DataLoader]:
+    if dataset_name.startswith('audiocaps_full'):
+        dataset = AudioCapsData(cfg.eval_data.AudioCaps_full.audio_path,
+                                cfg.eval_data.AudioCaps_full.csv_path)
+    elif dataset_name.startswith('audiocaps'):
+        dataset = AudioCapsData(cfg.eval_data.AudioCaps.audio_path,
+                                cfg.eval_data.AudioCaps.csv_path)
+    elif dataset_name.startswith('moviegen'):
+        dataset = MovieGen(cfg.eval_data.MovieGen.video_path,
+                           cfg.eval_data.MovieGen.jsonl_path,
+                           duration_sec=cfg.duration_s)
+    elif dataset_name.startswith('vggsound'):
+        dataset = VGGSound(cfg.eval_data.VGGSound.video_path,
+                           cfg.eval_data.VGGSound.csv_path,
+                           duration_sec=cfg.duration_s)
+    else:
+        raise ValueError(f'Invalid dataset name: {dataset_name}')
+
+    batch_size = cfg.batch_size
+    num_workers = cfg.num_workers
+    pin_memory = cfg.pin_memory
+    _, loader = construct_loader(dataset,
+                                 batch_size,
+                                 num_workers,
+                                 shuffle=False,
+                                 drop_last=False,
+                                 pin_memory=pin_memory,
+                                 error_avoidance=True)
+    return dataset, loader
+
+
+def error_avoidance_collate(batch):
+    batch = list(filter(lambda x: x is not None, batch))
+    return default_collate(batch)
+
+
+def construct_loader(dataset: Dataset,
+                     batch_size: int,
+                     num_workers: int,
+                     *,
+                     shuffle: bool = True,
+                     drop_last: bool = True,
+                     pin_memory: bool = False,
+                     error_avoidance: bool = False) -> tuple[DistributedSampler, DataLoader]:
+    train_sampler = DistributedSampler(dataset, rank=local_rank, shuffle=shuffle)
+    train_loader = DataLoader(dataset,
+                              batch_size,
+                              sampler=train_sampler,
+                              num_workers=num_workers,
+                              worker_init_fn=worker_init_fn,
+                              drop_last=drop_last,
+                              persistent_workers=num_workers > 0,
+                              pin_memory=pin_memory,
+                              collate_fn=error_avoidance_collate if error_avoidance else None)
+    return train_sampler, train_loader

mmaudio/data/eval/audiocaps.py

@@ -0,0 +1,39 @@
+import logging
+import os
+from collections import defaultdict
+from pathlib import Path
+from typing import Union
+
+import pandas as pd
+import torch
+from torch.utils.data.dataset import Dataset
+
+log = logging.getLogger()
+
+
+class AudioCapsData(Dataset):
+
+    def __init__(self, audio_path: Union[str, Path], csv_path: Union[str, Path]):
+        df = pd.read_csv(csv_path).to_dict(orient='records')
+
+        audio_files = sorted(os.listdir(audio_path))
+        audio_files = set(
+            [Path(f).stem for f in audio_files if f.endswith('.wav') or f.endswith('.flac')])
+
+        self.data = []
+        for row in df:
+            self.data.append({
+                'name': row['name'],
+                'caption': row['caption'],
+            })
+
+        self.audio_path = Path(audio_path)
+        self.csv_path = Path(csv_path)
+
+        log.info(f'Found {len(self.data)} matching audio files in {self.audio_path}')
+
+    def __getitem__(self, idx: int) -> torch.Tensor:
+        return self.data[idx]
+
+    def __len__(self):
+        return len(self.data)

mmaudio/data/eval/moviegen.py

@@ -0,0 +1,131 @@
+import json
+import logging
+import os
+from pathlib import Path
+from typing import Union
+
+import torch
+from torch.utils.data.dataset import Dataset
+from torchvision.transforms import v2
+from torio.io import StreamingMediaDecoder
+
+from mmaudio.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+class MovieGenData(Dataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        sync_root: Union[str, Path],
+        jsonl_root: Union[str, Path],
+        *,
+        duration_sec: float = 10.0,
+        read_clip: bool = True,
+    ):
+        self.video_root = Path(video_root)
+        self.sync_root = Path(sync_root)
+        self.jsonl_root = Path(jsonl_root)
+        self.read_clip = read_clip
+
+        videos = sorted(os.listdir(self.video_root))
+        videos = [v[:-4] for v in videos]  # remove extensions
+        self.captions = {}
+
+        for v in videos:
+            with open(self.jsonl_root / (v + '.jsonl')) as f:
+                data = json.load(f)
+                self.captions[v] = data['audio_prompt']
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+
+        self.duration_sec = duration_sec
+
+        self.clip_expected_length = int(_CLIP_FPS * self.duration_sec)
+        self.sync_expected_length = int(_SYNC_FPS * self.duration_sec)
+
+        self.clip_augment = v2.Compose([
+            v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+        ])
+
+        self.sync_augment = v2.Compose([
+            v2.Resize((_SYNC_SIZE, _SYNC_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+
+        self.videos = videos
+
+    def sample(self, idx: int) -> dict[str, torch.Tensor]:
+        video_id = self.videos[idx]
+        caption = self.captions[video_id]
+
+        reader = StreamingMediaDecoder(self.video_root / (video_id + '.mp4'))
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_CLIP_FPS * self.duration_sec),
+            frame_rate=_CLIP_FPS,
+            format='rgb24',
+        )
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_SYNC_FPS * self.duration_sec),
+            frame_rate=_SYNC_FPS,
+            format='rgb24',
+        )
+
+        reader.fill_buffer()
+        data_chunk = reader.pop_chunks()
+
+        clip_chunk = data_chunk[0]
+        sync_chunk = data_chunk[1]
+        if clip_chunk is None:
+            raise RuntimeError(f'CLIP video returned None {video_id}')
+        if clip_chunk.shape[0] < self.clip_expected_length:
+            raise RuntimeError(f'CLIP video too short {video_id}')
+
+        if sync_chunk is None:
+            raise RuntimeError(f'Sync video returned None {video_id}')
+        if sync_chunk.shape[0] < self.sync_expected_length:
+            raise RuntimeError(f'Sync video too short {video_id}')
+
+        # truncate the video
+        clip_chunk = clip_chunk[:self.clip_expected_length]
+        if clip_chunk.shape[0] != self.clip_expected_length:
+            raise RuntimeError(f'CLIP video wrong length {video_id}, '
+                               f'expected {self.clip_expected_length}, '
+                               f'got {clip_chunk.shape[0]}')
+        clip_chunk = self.clip_augment(clip_chunk)
+
+        sync_chunk = sync_chunk[:self.sync_expected_length]
+        if sync_chunk.shape[0] != self.sync_expected_length:
+            raise RuntimeError(f'Sync video wrong length {video_id}, '
+                               f'expected {self.sync_expected_length}, '
+                               f'got {sync_chunk.shape[0]}')
+        sync_chunk = self.sync_augment(sync_chunk)
+
+        data = {
+            'name': video_id,
+            'caption': caption,
+            'clip_video': clip_chunk,
+            'sync_video': sync_chunk,
+        }
+
+        return data
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        return self.sample(idx)
+
+    def __len__(self):
+        return len(self.captions)

mmaudio/data/eval/video_dataset.py

@@ -0,0 +1,197 @@
+import json
+import logging
+import os
+from pathlib import Path
+from typing import Union
+
+import pandas as pd
+import torch
+from torch.utils.data.dataset import Dataset
+from torchvision.transforms import v2
+from torio.io import StreamingMediaDecoder
+
+from mmaudio.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+class VideoDataset(Dataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        *,
+        duration_sec: float = 8.0,
+    ):
+        self.video_root = Path(video_root)
+
+        self.duration_sec = duration_sec
+
+        self.clip_expected_length = int(_CLIP_FPS * self.duration_sec)
+        self.sync_expected_length = int(_SYNC_FPS * self.duration_sec)
+
+        self.clip_transform = v2.Compose([
+            v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+        ])
+
+        self.sync_transform = v2.Compose([
+            v2.Resize(_SYNC_SIZE, interpolation=v2.InterpolationMode.BICUBIC),
+            v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+
+        # to be implemented by subclasses
+        self.captions = {}
+        self.videos = sorted(list(self.captions.keys()))
+
+    def sample(self, idx: int) -> dict[str, torch.Tensor]:
+        video_id = self.videos[idx]
+        caption = self.captions[video_id]
+
+        reader = StreamingMediaDecoder(self.video_root / (video_id + '.mp4'))
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_CLIP_FPS * self.duration_sec),
+            frame_rate=_CLIP_FPS,
+            format='rgb24',
+        )
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_SYNC_FPS * self.duration_sec),
+            frame_rate=_SYNC_FPS,
+            format='rgb24',
+        )
+
+        reader.fill_buffer()
+        data_chunk = reader.pop_chunks()
+
+        clip_chunk = data_chunk[0]
+        sync_chunk = data_chunk[1]
+        if clip_chunk is None:
+            raise RuntimeError(f'CLIP video returned None {video_id}')
+        if clip_chunk.shape[0] < self.clip_expected_length:
+            raise RuntimeError(
+                f'CLIP video too short {video_id}, expected {self.clip_expected_length}, got {clip_chunk.shape[0]}'
+            )
+
+        if sync_chunk is None:
+            raise RuntimeError(f'Sync video returned None {video_id}')
+        if sync_chunk.shape[0] < self.sync_expected_length:
+            raise RuntimeError(
+                f'Sync video too short {video_id}, expected {self.sync_expected_length}, got {sync_chunk.shape[0]}'
+            )
+
+        # truncate the video
+        clip_chunk = clip_chunk[:self.clip_expected_length]
+        if clip_chunk.shape[0] != self.clip_expected_length:
+            raise RuntimeError(f'CLIP video wrong length {video_id}, '
+                               f'expected {self.clip_expected_length}, '
+                               f'got {clip_chunk.shape[0]}')
+        clip_chunk = self.clip_transform(clip_chunk)
+
+        sync_chunk = sync_chunk[:self.sync_expected_length]
+        if sync_chunk.shape[0] != self.sync_expected_length:
+            raise RuntimeError(f'Sync video wrong length {video_id}, '
+                               f'expected {self.sync_expected_length}, '
+                               f'got {sync_chunk.shape[0]}')
+        sync_chunk = self.sync_transform(sync_chunk)
+
+        data = {
+            'name': video_id,
+            'caption': caption,
+            'clip_video': clip_chunk,
+            'sync_video': sync_chunk,
+        }
+
+        return data
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        try:
+            return self.sample(idx)
+        except Exception as e:
+            log.error(f'Error loading video {self.videos[idx]}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.captions)
+
+
+class VGGSound(VideoDataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        csv_path: Union[str, Path],
+        *,
+        duration_sec: float = 8.0,
+    ):
+        super().__init__(video_root, duration_sec=duration_sec)
+        self.video_root = Path(video_root)
+        self.csv_path = Path(csv_path)
+
+        videos = sorted(os.listdir(self.video_root))
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+        self.captions = {}
+
+        df = pd.read_csv(csv_path, header=None, names=['id', 'sec', 'caption',
+                                                       'split']).to_dict(orient='records')
+
+        videos_no_found = []
+        for row in df:
+            if row['split'] == 'test':
+                start_sec = int(row['sec'])
+                video_id = str(row['id'])
+                # this is how our videos are named
+                video_name = f'{video_id}_{start_sec:06d}'
+                if video_name + '.mp4' not in videos:
+                    videos_no_found.append(video_name)
+                    continue
+
+                self.captions[video_name] = row['caption']
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+            log.info(f'{len(self.captions)} useable videos found')
+            if videos_no_found:
+                log.info(f'{len(videos_no_found)} found in {csv_path} but not in {video_root}')
+                log.info(
+                    'A small amount is expected, as not all videos are still available on YouTube')
+
+        self.videos = sorted(list(self.captions.keys()))
+
+
+class MovieGen(VideoDataset):
+
+    def __init__(
+        self,
+        video_root: Union[str, Path],
+        jsonl_root: Union[str, Path],
+        *,
+        duration_sec: float = 10.0,
+    ):
+        super().__init__(video_root, duration_sec=duration_sec)
+        self.video_root = Path(video_root)
+        self.jsonl_root = Path(jsonl_root)
+
+        videos = sorted(os.listdir(self.video_root))
+        videos = [v[:-4] for v in videos]  # remove extensions
+        self.captions = {}
+
+        for v in videos:
+            with open(self.jsonl_root / (v + '.jsonl')) as f:
+                data = json.load(f)
+                self.captions[v] = data['audio_prompt']
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {video_root}')
+
+        self.videos = videos

mmaudio/data/extracted_audio.py

@@ -0,0 +1,88 @@
+import logging
+from pathlib import Path
+from typing import Union
+
+import pandas as pd
+import torch
+from tensordict import TensorDict
+from torch.utils.data.dataset import Dataset
+
+from mmaudio.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+
+class ExtractedAudio(Dataset):
+
+    def __init__(
+        self,
+        tsv_path: Union[str, Path],
+        *,
+        premade_mmap_dir: Union[str, Path],
+        data_dim: dict[str, int],
+    ):
+        super().__init__()
+
+        self.data_dim = data_dim
+        self.df_list = pd.read_csv(tsv_path, sep='\t').to_dict('records')
+        self.ids = [str(d['id']) for d in self.df_list]
+
+        log.info(f'Loading precomputed mmap from {premade_mmap_dir}')
+        # load precomputed memory mapped tensors
+        premade_mmap_dir = Path(premade_mmap_dir)
+        td = TensorDict.load_memmap(premade_mmap_dir)
+        log.info(f'Loaded precomputed mmap from {premade_mmap_dir}')
+        self.mean = td['mean']
+        self.std = td['std']
+        self.text_features = td['text_features']
+
+        log.info(f'Loaded {len(self)} samples from {premade_mmap_dir}.')
+        log.info(f'Loaded mean: {self.mean.shape}.')
+        log.info(f'Loaded std: {self.std.shape}.')
+        log.info(f'Loaded text features: {self.text_features.shape}.')
+
+        assert self.mean.shape[1] == self.data_dim['latent_seq_len'], \
+            f'{self.mean.shape[1]} != {self.data_dim["latent_seq_len"]}'
+        assert self.std.shape[1] == self.data_dim['latent_seq_len'], \
+            f'{self.std.shape[1]} != {self.data_dim["latent_seq_len"]}'
+
+        assert self.text_features.shape[1] == self.data_dim['text_seq_len'], \
+            f'{self.text_features.shape[1]} != {self.data_dim["text_seq_len"]}'
+        assert self.text_features.shape[-1] == self.data_dim['text_dim'], \
+            f'{self.text_features.shape[-1]} != {self.data_dim["text_dim"]}'
+
+        self.fake_clip_features = torch.zeros(self.data_dim['clip_seq_len'],
+                                              self.data_dim['clip_dim'])
+        self.fake_sync_features = torch.zeros(self.data_dim['sync_seq_len'],
+                                              self.data_dim['sync_dim'])
+        self.video_exist = torch.tensor(0, dtype=torch.bool)
+        self.text_exist = torch.tensor(1, dtype=torch.bool)
+
+    def compute_latent_stats(self) -> tuple[torch.Tensor, torch.Tensor]:
+        latents = self.mean
+        return latents.mean(dim=(0, 1)), latents.std(dim=(0, 1))
+
+    def get_memory_mapped_tensor(self) -> TensorDict:
+        td = TensorDict({
+            'mean': self.mean,
+            'std': self.std,
+            'text_features': self.text_features,
+        })
+        return td
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        data = {
+            'id': str(self.df_list[idx]['id']),
+            'a_mean': self.mean[idx],
+            'a_std': self.std[idx],
+            'clip_features': self.fake_clip_features,
+            'sync_features': self.fake_sync_features,
+            'text_features': self.text_features[idx],
+            'caption': self.df_list[idx]['caption'],
+            'video_exist': self.video_exist,
+            'text_exist': self.text_exist,
+        }
+        return data
+
+    def __len__(self):
+        return len(self.ids)

mmaudio/data/extracted_vgg.py

@@ -0,0 +1,101 @@
+import logging
+from pathlib import Path
+from typing import Union
+
+import pandas as pd
+import torch
+from tensordict import TensorDict
+from torch.utils.data.dataset import Dataset
+
+from mmaudio.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+
+class ExtractedVGG(Dataset):
+
+    def __init__(
+        self,
+        tsv_path: Union[str, Path],
+        *,
+        premade_mmap_dir: Union[str, Path],
+        data_dim: dict[str, int],
+    ):
+        super().__init__()
+
+        self.data_dim = data_dim
+        self.df_list = pd.read_csv(tsv_path, sep='\t').to_dict('records')
+        self.ids = [d['id'] for d in self.df_list]
+
+        log.info(f'Loading precomputed mmap from {premade_mmap_dir}')
+        # load precomputed memory mapped tensors
+        premade_mmap_dir = Path(premade_mmap_dir)
+        td = TensorDict.load_memmap(premade_mmap_dir)
+        log.info(f'Loaded precomputed mmap from {premade_mmap_dir}')
+        self.mean = td['mean']
+        self.std = td['std']
+        self.clip_features = td['clip_features']
+        self.sync_features = td['sync_features']
+        self.text_features = td['text_features']
+
+        if local_rank == 0:
+            log.info(f'Loaded {len(self)} samples.')
+            log.info(f'Loaded mean: {self.mean.shape}.')
+            log.info(f'Loaded std: {self.std.shape}.')
+            log.info(f'Loaded clip_features: {self.clip_features.shape}.')
+            log.info(f'Loaded sync_features: {self.sync_features.shape}.')
+            log.info(f'Loaded text_features: {self.text_features.shape}.')
+
+        assert self.mean.shape[1] == self.data_dim['latent_seq_len'], \
+            f'{self.mean.shape[1]} != {self.data_dim["latent_seq_len"]}'
+        assert self.std.shape[1] == self.data_dim['latent_seq_len'], \
+            f'{self.std.shape[1]} != {self.data_dim["latent_seq_len"]}'
+
+        assert self.clip_features.shape[1] == self.data_dim['clip_seq_len'], \
+            f'{self.clip_features.shape[1]} != {self.data_dim["clip_seq_len"]}'
+        assert self.sync_features.shape[1] == self.data_dim['sync_seq_len'], \
+            f'{self.sync_features.shape[1]} != {self.data_dim["sync_seq_len"]}'
+        assert self.text_features.shape[1] == self.data_dim['text_seq_len'], \
+            f'{self.text_features.shape[1]} != {self.data_dim["text_seq_len"]}'
+
+        assert self.clip_features.shape[-1] == self.data_dim['clip_dim'], \
+            f'{self.clip_features.shape[-1]} != {self.data_dim["clip_dim"]}'
+        assert self.sync_features.shape[-1] == self.data_dim['sync_dim'], \
+            f'{self.sync_features.shape[-1]} != {self.data_dim["sync_dim"]}'
+        assert self.text_features.shape[-1] == self.data_dim['text_dim'], \
+            f'{self.text_features.shape[-1]} != {self.data_dim["text_dim"]}'
+
+        self.video_exist = torch.tensor(1, dtype=torch.bool)
+        self.text_exist = torch.tensor(1, dtype=torch.bool)
+
+    def compute_latent_stats(self) -> tuple[torch.Tensor, torch.Tensor]:
+        latents = self.mean
+        return latents.mean(dim=(0, 1)), latents.std(dim=(0, 1))
+
+    def get_memory_mapped_tensor(self) -> TensorDict:
+        td = TensorDict({
+            'mean': self.mean,
+            'std': self.std,
+            'clip_features': self.clip_features,
+            'sync_features': self.sync_features,
+            'text_features': self.text_features,
+        })
+        return td
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        data = {
+            'id': self.df_list[idx]['id'],
+            'a_mean': self.mean[idx],
+            'a_std': self.std[idx],
+            'clip_features': self.clip_features[idx],
+            'sync_features': self.sync_features[idx],
+            'text_features': self.text_features[idx],
+            'caption': self.df_list[idx]['label'],
+            'video_exist': self.video_exist,
+            'text_exist': self.text_exist,
+        }
+
+        return data
+
+    def __len__(self):
+        return len(self.ids)

mmaudio/data/extraction/vgg_sound.py

@@ -0,0 +1,193 @@
+import logging
+import os
+from pathlib import Path
+from typing import Optional, Union
+
+import pandas as pd
+import torch
+import torchaudio
+from torch.utils.data.dataset import Dataset
+from torchvision.transforms import v2
+from torio.io import StreamingMediaDecoder
+
+from mmaudio.utils.dist_utils import local_rank
+
+log = logging.getLogger()
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+class VGGSound(Dataset):
+
+    def __init__(
+        self,
+        root: Union[str, Path],
+        *,
+        tsv_path: Union[str, Path] = 'sets/vgg3-train.tsv',
+        sample_rate: int = 16_000,
+        duration_sec: float = 8.0,
+        audio_samples: Optional[int] = None,
+        normalize_audio: bool = False,
+    ):
+        self.root = Path(root)
+        self.normalize_audio = normalize_audio
+        if audio_samples is None:
+            self.audio_samples = int(sample_rate * duration_sec)
+        else:
+            self.audio_samples = audio_samples
+            effective_duration = audio_samples / sample_rate
+            # make sure the duration is close enough, within 15ms
+            assert abs(effective_duration - duration_sec) < 0.015, \
+                f'audio_samples {audio_samples} does not match duration_sec {duration_sec}'
+
+        videos = sorted(os.listdir(self.root))
+        videos = set([Path(v).stem for v in videos])  # remove extensions
+        self.labels = {}
+        self.videos = []
+        missing_videos = []
+
+        # read the tsv for subset information
+        df_list = pd.read_csv(tsv_path, sep='\t', dtype={'id': str}).to_dict('records')
+        for record in df_list:
+            id = record['id']
+            label = record['label']
+            if id in videos:
+                self.labels[id] = label
+                self.videos.append(id)
+            else:
+                missing_videos.append(id)
+
+        if local_rank == 0:
+            log.info(f'{len(videos)} videos found in {root}')
+            log.info(f'{len(self.videos)} videos found in {tsv_path}')
+            log.info(f'{len(missing_videos)} videos missing in {root}')
+
+        self.sample_rate = sample_rate
+        self.duration_sec = duration_sec
+
+        self.expected_audio_length = audio_samples
+        self.clip_expected_length = int(_CLIP_FPS * self.duration_sec)
+        self.sync_expected_length = int(_SYNC_FPS * self.duration_sec)
+
+        self.clip_transform = v2.Compose([
+            v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+        ])
+
+        self.sync_transform = v2.Compose([
+            v2.Resize(_SYNC_SIZE, interpolation=v2.InterpolationMode.BICUBIC),
+            v2.CenterCrop(_SYNC_SIZE),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+
+        self.resampler = {}
+
+    def sample(self, idx: int) -> dict[str, torch.Tensor]:
+        video_id = self.videos[idx]
+        label = self.labels[video_id]
+
+        reader = StreamingMediaDecoder(self.root / (video_id + '.mp4'))
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_CLIP_FPS * self.duration_sec),
+            frame_rate=_CLIP_FPS,
+            format='rgb24',
+        )
+        reader.add_basic_video_stream(
+            frames_per_chunk=int(_SYNC_FPS * self.duration_sec),
+            frame_rate=_SYNC_FPS,
+            format='rgb24',
+        )
+        reader.add_basic_audio_stream(frames_per_chunk=2**30, )
+
+        reader.fill_buffer()
+        data_chunk = reader.pop_chunks()
+
+        clip_chunk = data_chunk[0]
+        sync_chunk = data_chunk[1]
+        audio_chunk = data_chunk[2]
+
+        if clip_chunk is None:
+            raise RuntimeError(f'CLIP video returned None {video_id}')
+        if clip_chunk.shape[0] < self.clip_expected_length:
+            raise RuntimeError(
+                f'CLIP video too short {video_id}, expected {self.clip_expected_length}, got {clip_chunk.shape[0]}'
+            )
+
+        if sync_chunk is None:
+            raise RuntimeError(f'Sync video returned None {video_id}')
+        if sync_chunk.shape[0] < self.sync_expected_length:
+            raise RuntimeError(
+                f'Sync video too short {video_id}, expected {self.sync_expected_length}, got {sync_chunk.shape[0]}'
+            )
+
+        # process audio
+        sample_rate = int(reader.get_out_stream_info(2).sample_rate)
+        audio_chunk = audio_chunk.transpose(0, 1)
+        audio_chunk = audio_chunk.mean(dim=0)  # mono
+        if self.normalize_audio:
+            abs_max = audio_chunk.abs().max()
+            audio_chunk = audio_chunk / abs_max * 0.95
+            if abs_max <= 1e-6:
+                raise RuntimeError(f'Audio is silent {video_id}')
+
+        # resample
+        if sample_rate == self.sample_rate:
+            audio_chunk = audio_chunk
+        else:
+            if sample_rate not in self.resampler:
+                # https://pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html#kaiser-best
+                self.resampler[sample_rate] = torchaudio.transforms.Resample(
+                    sample_rate,
+                    self.sample_rate,
+                    lowpass_filter_width=64,
+                    rolloff=0.9475937167399596,
+                    resampling_method='sinc_interp_kaiser',
+                    beta=14.769656459379492,
+                )
+            audio_chunk = self.resampler[sample_rate](audio_chunk)
+
+        if audio_chunk.shape[0] < self.expected_audio_length:
+            raise RuntimeError(f'Audio too short {video_id}')
+        audio_chunk = audio_chunk[:self.expected_audio_length]
+
+        # truncate the video
+        clip_chunk = clip_chunk[:self.clip_expected_length]
+        if clip_chunk.shape[0] != self.clip_expected_length:
+            raise RuntimeError(f'CLIP video wrong length {video_id}, '
+                               f'expected {self.clip_expected_length}, '
+                               f'got {clip_chunk.shape[0]}')
+        clip_chunk = self.clip_transform(clip_chunk)
+
+        sync_chunk = sync_chunk[:self.sync_expected_length]
+        if sync_chunk.shape[0] != self.sync_expected_length:
+            raise RuntimeError(f'Sync video wrong length {video_id}, '
+                               f'expected {self.sync_expected_length}, '
+                               f'got {sync_chunk.shape[0]}')
+        sync_chunk = self.sync_transform(sync_chunk)
+
+        data = {
+            'id': video_id,
+            'caption': label,
+            'audio': audio_chunk,
+            'clip_video': clip_chunk,
+            'sync_video': sync_chunk,
+        }
+
+        return data
+
+    def __getitem__(self, idx: int) -> dict[str, torch.Tensor]:
+        try:
+            return self.sample(idx)
+        except Exception as e:
+            log.error(f'Error loading video {self.videos[idx]}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.labels)

mmaudio/data/extraction/wav_dataset.py

@@ -0,0 +1,132 @@
+import logging
+import os
+from pathlib import Path
+from typing import Union
+
+import open_clip
+import pandas as pd
+import torch
+import torchaudio
+from torch.utils.data.dataset import Dataset
+
+log = logging.getLogger()
+
+
+class WavTextClipsDataset(Dataset):
+
+    def __init__(
+        self,
+        root: Union[str, Path],
+        *,
+        captions_tsv: Union[str, Path],
+        clips_tsv: Union[str, Path],
+        sample_rate: int,
+        num_samples: int,
+        normalize_audio: bool = False,
+        reject_silent: bool = False,
+        tokenizer_id: str = 'ViT-H-14-378-quickgelu',
+    ):
+        self.root = Path(root)
+        self.sample_rate = sample_rate
+        self.num_samples = num_samples
+        self.normalize_audio = normalize_audio
+        self.reject_silent = reject_silent
+        self.tokenizer = open_clip.get_tokenizer(tokenizer_id)
+
+        audios = sorted(os.listdir(self.root))
+        audios = set([
+            Path(audio).stem for audio in audios
+            if audio.endswith('.wav') or audio.endswith('.flac')
+        ])
+        self.captions = {}
+
+        # read the caption tsv
+        df_list = pd.read_csv(captions_tsv, sep='\t', dtype={'id': str}).to_dict('records')
+        for record in df_list:
+            id = record['id']
+            caption = record['caption']
+            self.captions[id] = caption
+
+        # read the clip tsv
+        df_list = pd.read_csv(clips_tsv, sep='\t', dtype={
+            'id': str,
+            'name': str
+        }).to_dict('records')
+        self.clips = []
+        for record in df_list:
+            record['id'] = record['id']
+            record['name'] = record['name']
+            id = record['id']
+            name = record['name']
+            if name not in self.captions:
+                log.warning(f'Audio {name} not found in {captions_tsv}')
+                continue
+            record['caption'] = self.captions[name]
+            self.clips.append(record)
+
+        log.info(f'Found {len(self.clips)} audio files in {self.root}')
+
+        self.resampler = {}
+
+    def __getitem__(self, idx: int) -> torch.Tensor:
+        try:
+            clip = self.clips[idx]
+            audio_name = clip['name']
+            audio_id = clip['id']
+            caption = clip['caption']
+            start_sample = clip['start_sample']
+            end_sample = clip['end_sample']
+
+            audio_path = self.root / f'{audio_name}.flac'
+            if not audio_path.exists():
+                audio_path = self.root / f'{audio_name}.wav'
+                assert audio_path.exists()
+
+            audio_chunk, sample_rate = torchaudio.load(audio_path)
+            audio_chunk = audio_chunk.mean(dim=0)  # mono
+            abs_max = audio_chunk.abs().max()
+            if self.normalize_audio:
+                audio_chunk = audio_chunk / abs_max * 0.95
+
+            if self.reject_silent and abs_max < 1e-6:
+                log.warning(f'Rejecting silent audio')
+                return None
+
+            audio_chunk = audio_chunk[start_sample:end_sample]
+
+            # resample
+            if sample_rate == self.sample_rate:
+                audio_chunk = audio_chunk
+            else:
+                if sample_rate not in self.resampler:
+                    # https://pytorch.org/audio/stable/tutorials/audio_resampling_tutorial.html#kaiser-best
+                    self.resampler[sample_rate] = torchaudio.transforms.Resample(
+                        sample_rate,
+                        self.sample_rate,
+                        lowpass_filter_width=64,
+                        rolloff=0.9475937167399596,
+                        resampling_method='sinc_interp_kaiser',
+                        beta=14.769656459379492,
+                    )
+                audio_chunk = self.resampler[sample_rate](audio_chunk)
+
+            if audio_chunk.shape[0] < self.num_samples:
+                raise ValueError('Audio is too short')
+            audio_chunk = audio_chunk[:self.num_samples]
+
+            tokens = self.tokenizer([caption])[0]
+
+            output = {
+                'waveform': audio_chunk,
+                'id': audio_id,
+                'caption': caption,
+                'tokens': tokens,
+            }
+
+            return output
+        except Exception as e:
+            log.error(f'Error reading {audio_path}: {e}')
+            return None
+
+    def __len__(self):
+        return len(self.clips)

mmaudio/data/mm_dataset.py

@@ -0,0 +1,45 @@
+import bisect
+
+import torch
+from torch.utils.data.dataset import Dataset
+
+
+# modified from https://pytorch.org/docs/stable/_modules/torch/utils/data/dataset.html#ConcatDataset
+class MultiModalDataset(Dataset):
+    datasets: list[Dataset]
+    cumulative_sizes: list[int]
+
+    @staticmethod
+    def cumsum(sequence):
+        r, s = [], 0
+        for e in sequence:
+            l = len(e)
+            r.append(l + s)
+            s += l
+        return r
+
+    def __init__(self, video_datasets: list[Dataset], audio_datasets: list[Dataset]):
+        super().__init__()
+        self.video_datasets = list(video_datasets)
+        self.audio_datasets = list(audio_datasets)
+        self.datasets = self.video_datasets + self.audio_datasets
+
+        self.cumulative_sizes = self.cumsum(self.datasets)
+
+    def __len__(self):
+        return self.cumulative_sizes[-1]
+
+    def __getitem__(self, idx):
+        if idx < 0:
+            if -idx > len(self):
+                raise ValueError("absolute value of index should not exceed dataset length")
+            idx = len(self) + idx
+        dataset_idx = bisect.bisect_right(self.cumulative_sizes, idx)
+        if dataset_idx == 0:
+            sample_idx = idx
+        else:
+            sample_idx = idx - self.cumulative_sizes[dataset_idx - 1]
+        return self.datasets[dataset_idx][sample_idx]
+
+    def compute_latent_stats(self) -> tuple[torch.Tensor, torch.Tensor]:
+        return self.video_datasets[0].compute_latent_stats()

mmaudio/data/utils.py

@@ -0,0 +1,148 @@
+import logging
+import os
+import random
+import tempfile
+from pathlib import Path
+from typing import Any, Optional, Union
+
+import torch
+import torch.distributed as dist
+from tensordict import MemoryMappedTensor
+from torch.utils.data import DataLoader
+from torch.utils.data.dataset import Dataset
+from tqdm import tqdm
+
+from mmaudio.utils.dist_utils import local_rank, world_size
+
+scratch_path = Path(os.environ['SLURM_SCRATCH'] if 'SLURM_SCRATCH' in os.environ else '/dev/shm')
+shm_path = Path('/dev/shm')
+
+log = logging.getLogger()
+
+
+def reseed(seed):
+    random.seed(seed)
+    torch.manual_seed(seed)
+
+
+def local_scatter_torch(obj: Optional[Any]):
+    if world_size == 1:
+        # Just one worker. Do nothing.
+        return obj
+
+    array = [obj] * world_size
+    target_array = [None]
+    if local_rank == 0:
+        dist.scatter_object_list(target_array, scatter_object_input_list=array, src=0)
+    else:
+        dist.scatter_object_list(target_array, scatter_object_input_list=None, src=0)
+    return target_array[0]
+
+
+class ShardDataset(Dataset):
+
+    def __init__(self, root):
+        self.root = root
+        self.shards = sorted(os.listdir(root))
+
+    def __len__(self):
+        return len(self.shards)
+
+    def __getitem__(self, idx):
+        return torch.load(os.path.join(self.root, self.shards[idx]), weights_only=True)
+
+
+def get_tmp_dir(in_memory: bool) -> Path:
+    return shm_path if in_memory else scratch_path
+
+
+def load_shards_and_share(data_path: Union[str, Path], ids: list[int],
+                          in_memory: bool) -> MemoryMappedTensor:
+    if local_rank == 0:
+        with tempfile.NamedTemporaryFile(prefix='shared-tensor-', dir=get_tmp_dir(in_memory)) as f:
+            log.info(f'Loading shards from {data_path} into {f.name}...')
+            data = load_shards(data_path, ids=ids, tmp_file_path=f.name)
+            data = share_tensor_to_all(data)
+            torch.distributed.barrier()
+            f.close()  # why does the context manager not close the file for me?
+    else:
+        log.info('Waiting for the data to be shared with me...')
+        data = share_tensor_to_all(None)
+        torch.distributed.barrier()
+
+    return data
+
+
+def load_shards(
+    data_path: Union[str, Path],
+    ids: list[int],
+    *,
+    tmp_file_path: str,
+) -> Union[torch.Tensor, dict[str, torch.Tensor]]:
+
+    id_set = set(ids)
+    shards = sorted(os.listdir(data_path))
+    log.info(f'Found {len(shards)} shards in {data_path}.')
+    first_shard = torch.load(os.path.join(data_path, shards[0]), weights_only=True)
+
+    log.info(f'Rank {local_rank} created file {tmp_file_path}')
+    first_item = next(iter(first_shard.values()))
+    log.info(f'First item shape: {first_item.shape}')
+    mm_tensor = MemoryMappedTensor.empty(shape=(len(ids), *first_item.shape),
+                                         dtype=torch.float32,
+                                         filename=tmp_file_path,
+                                         existsok=True)
+    total_count = 0
+    used_index = set()
+    id_indexing = {i: idx for idx, i in enumerate(ids)}
+    # faster with no workers; otherwise we need to set_sharing_strategy('file_system')
+    loader = DataLoader(ShardDataset(data_path), batch_size=1, num_workers=0)
+    for data in tqdm(loader, desc='Loading shards'):
+        for i, v in data.items():
+            if i not in id_set:
+                continue
+
+            # tensor_index = ids.index(i)
+            tensor_index = id_indexing[i]
+            if tensor_index in used_index:
+                raise ValueError(f'Duplicate id {i} found in {data_path}.')
+            used_index.add(tensor_index)
+            mm_tensor[tensor_index] = v
+            total_count += 1
+
+    assert total_count == len(ids), f'Expected {len(ids)} tensors, got {total_count}.'
+    log.info(f'Loaded {total_count} tensors from {data_path}.')
+
+    return mm_tensor
+
+
+def share_tensor_to_all(x: Optional[MemoryMappedTensor]) -> MemoryMappedTensor:
+    """
+    x: the tensor to be shared; None if local_rank != 0
+    return: the shared tensor
+    """
+
+    # there is no need to share your stuff with anyone if you are alone; must be in memory
+    if world_size == 1:
+        return x
+
+    if local_rank == 0:
+        assert x is not None, 'x must not be None if local_rank == 0'
+    else:
+        assert x is None, 'x must be None if local_rank != 0'
+
+    if local_rank == 0:
+        filename = x.filename
+        meta_information = (filename, x.shape, x.dtype)
+    else:
+        meta_information = None
+
+    filename, data_shape, data_type = local_scatter_torch(meta_information)
+    if local_rank == 0:
+        data = x
+    else:
+        data = MemoryMappedTensor.from_filename(filename=filename,
+                                                dtype=data_type,
+                                                shape=data_shape)
+
+    return data

mmaudio/eval_utils.py

@@ -0,0 +1,255 @@
+import dataclasses
+import logging
+from pathlib import Path
+from typing import Optional
+
+import numpy as np
+import torch
+from colorlog import ColoredFormatter
+from PIL import Image
+from torchvision.transforms import v2
+
+from mmaudio.data.av_utils import ImageInfo, VideoInfo, read_frames, reencode_with_audio
+from mmaudio.model.flow_matching import FlowMatching
+from mmaudio.model.networks import MMAudio
+from mmaudio.model.sequence_config import CONFIG_16K, CONFIG_44K, SequenceConfig
+from mmaudio.model.utils.features_utils import FeaturesUtils
+from mmaudio.utils.download_utils import download_model_if_needed
+
+log = logging.getLogger()
+
+
+@dataclasses.dataclass
+class ModelConfig:
+    model_name: str
+    model_path: Path
+    vae_path: Path
+    bigvgan_16k_path: Optional[Path]
+    mode: str
+    synchformer_ckpt: Path = Path('./ext_weights/synchformer_state_dict.pth')
+
+    @property
+    def seq_cfg(self) -> SequenceConfig:
+        if self.mode == '16k':
+            return CONFIG_16K
+        elif self.mode == '44k':
+            return CONFIG_44K
+
+    def download_if_needed(self):
+        download_model_if_needed(self.model_path)
+        download_model_if_needed(self.vae_path)
+        if self.bigvgan_16k_path is not None:
+            download_model_if_needed(self.bigvgan_16k_path)
+        download_model_if_needed(self.synchformer_ckpt)
+
+
+small_16k = ModelConfig(model_name='small_16k',
+                        model_path=Path('./weights/mmaudio_small_16k.pth'),
+                        vae_path=Path('./ext_weights/v1-16.pth'),
+                        bigvgan_16k_path=Path('./ext_weights/best_netG.pt'),
+                        mode='16k')
+small_44k = ModelConfig(model_name='small_44k',
+                        model_path=Path('./weights/mmaudio_small_44k.pth'),
+                        vae_path=Path('./ext_weights/v1-44.pth'),
+                        bigvgan_16k_path=None,
+                        mode='44k')
+medium_44k = ModelConfig(model_name='medium_44k',
+                         model_path=Path('./weights/mmaudio_medium_44k.pth'),
+                         vae_path=Path('./ext_weights/v1-44.pth'),
+                         bigvgan_16k_path=None,
+                         mode='44k')
+large_44k = ModelConfig(model_name='large_44k',
+                        model_path=Path('./weights/mmaudio_large_44k.pth'),
+                        vae_path=Path('./ext_weights/v1-44.pth'),
+                        bigvgan_16k_path=None,
+                        mode='44k')
+large_44k_v2 = ModelConfig(model_name='large_44k_v2',
+                           model_path=Path('./weights/mmaudio_large_44k_v2.pth'),
+                           vae_path=Path('./ext_weights/v1-44.pth'),
+                           bigvgan_16k_path=None,
+                           mode='44k')
+all_model_cfg: dict[str, ModelConfig] = {
+    'small_16k': small_16k,
+    'small_44k': small_44k,
+    'medium_44k': medium_44k,
+    'large_44k': large_44k,
+    'large_44k_v2': large_44k_v2,
+}
+
+
+def generate(
+    clip_video: Optional[torch.Tensor],
+    sync_video: Optional[torch.Tensor],
+    text: Optional[list[str]],
+    *,
+    negative_text: Optional[list[str]] = None,
+    feature_utils: FeaturesUtils,
+    net: MMAudio,
+    fm: FlowMatching,
+    rng: torch.Generator,
+    cfg_strength: float,
+    clip_batch_size_multiplier: int = 40,
+    sync_batch_size_multiplier: int = 40,
+    image_input: bool = False,
+) -> torch.Tensor:
+    device = feature_utils.device
+    dtype = feature_utils.dtype
+
+    bs = len(text)
+    if clip_video is not None:
+        clip_video = clip_video.to(device, dtype, non_blocking=True)
+        clip_features = feature_utils.encode_video_with_clip(clip_video,
+                                                             batch_size=bs *
+                                                             clip_batch_size_multiplier)
+        if image_input:
+            clip_features = clip_features.expand(-1, net.clip_seq_len, -1)
+    else:
+        clip_features = net.get_empty_clip_sequence(bs)
+
+    if sync_video is not None and not image_input:
+        sync_video = sync_video.to(device, dtype, non_blocking=True)
+        sync_features = feature_utils.encode_video_with_sync(sync_video,
+                                                             batch_size=bs *
+                                                             sync_batch_size_multiplier)
+    else:
+        sync_features = net.get_empty_sync_sequence(bs)
+
+    if text is not None:
+        text_features = feature_utils.encode_text(text)
+    else:
+        text_features = net.get_empty_string_sequence(bs)
+
+    if negative_text is not None:
+        assert len(negative_text) == bs
+        negative_text_features = feature_utils.encode_text(negative_text)
+    else:
+        negative_text_features = net.get_empty_string_sequence(bs)
+
+    x0 = torch.randn(bs,
+                     net.latent_seq_len,
+                     net.latent_dim,
+                     device=device,
+                     dtype=dtype,
+                     generator=rng)
+    preprocessed_conditions = net.preprocess_conditions(clip_features, sync_features, text_features)
+    empty_conditions = net.get_empty_conditions(
+        bs, negative_text_features=negative_text_features if negative_text is not None else None)
+
+    cfg_ode_wrapper = lambda t, x: net.ode_wrapper(t, x, preprocessed_conditions, empty_conditions,
+                                                   cfg_strength)
+    x1 = fm.to_data(cfg_ode_wrapper, x0)
+    x1 = net.unnormalize(x1)
+    spec = feature_utils.decode(x1)
+    audio = feature_utils.vocode(spec)
+    return audio
+
+
+LOGFORMAT = "[%(log_color)s%(levelname)-8s%(reset)s]: %(log_color)s%(message)s%(reset)s"
+
+
+def setup_eval_logging(log_level: int = logging.INFO):
+    logging.root.setLevel(log_level)
+    formatter = ColoredFormatter(LOGFORMAT)
+    stream = logging.StreamHandler()
+    stream.setLevel(log_level)
+    stream.setFormatter(formatter)
+    log = logging.getLogger()
+    log.setLevel(log_level)
+    log.addHandler(stream)
+
+
+_CLIP_SIZE = 384
+_CLIP_FPS = 8.0
+
+_SYNC_SIZE = 224
+_SYNC_FPS = 25.0
+
+
+def load_video(video_path: Path, duration_sec: float, load_all_frames: bool = True) -> VideoInfo:
+
+    clip_transform = v2.Compose([
+        v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+    ])
+
+    sync_transform = v2.Compose([
+        v2.Resize(_SYNC_SIZE, interpolation=v2.InterpolationMode.BICUBIC),
+        v2.CenterCrop(_SYNC_SIZE),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+        v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+    ])
+
+    output_frames, all_frames, orig_fps = read_frames(video_path,
+                                                      list_of_fps=[_CLIP_FPS, _SYNC_FPS],
+                                                      start_sec=0,
+                                                      end_sec=duration_sec,
+                                                      need_all_frames=load_all_frames)
+
+    clip_chunk, sync_chunk = output_frames
+    clip_chunk = torch.from_numpy(clip_chunk).permute(0, 3, 1, 2)
+    sync_chunk = torch.from_numpy(sync_chunk).permute(0, 3, 1, 2)
+
+    clip_frames = clip_transform(clip_chunk)
+    sync_frames = sync_transform(sync_chunk)
+
+    clip_length_sec = clip_frames.shape[0] / _CLIP_FPS
+    sync_length_sec = sync_frames.shape[0] / _SYNC_FPS
+
+    if clip_length_sec < duration_sec:
+        log.warning(f'Clip video is too short: {clip_length_sec:.2f} < {duration_sec:.2f}')
+        log.warning(f'Truncating to {clip_length_sec:.2f} sec')
+        duration_sec = clip_length_sec
+
+    if sync_length_sec < duration_sec:
+        log.warning(f'Sync video is too short: {sync_length_sec:.2f} < {duration_sec:.2f}')
+        log.warning(f'Truncating to {sync_length_sec:.2f} sec')
+        duration_sec = sync_length_sec
+
+    clip_frames = clip_frames[:int(_CLIP_FPS * duration_sec)]
+    sync_frames = sync_frames[:int(_SYNC_FPS * duration_sec)]
+
+    video_info = VideoInfo(
+        duration_sec=duration_sec,
+        fps=orig_fps,
+        clip_frames=clip_frames,
+        sync_frames=sync_frames,
+        all_frames=all_frames if load_all_frames else None,
+    )
+    return video_info
+
+
+def load_image(image_path: Path) -> VideoInfo:
+    clip_transform = v2.Compose([
+        v2.Resize((_CLIP_SIZE, _CLIP_SIZE), interpolation=v2.InterpolationMode.BICUBIC),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+    ])
+
+    sync_transform = v2.Compose([
+        v2.Resize(_SYNC_SIZE, interpolation=v2.InterpolationMode.BICUBIC),
+        v2.CenterCrop(_SYNC_SIZE),
+        v2.ToImage(),
+        v2.ToDtype(torch.float32, scale=True),
+        v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+    ])
+
+    frame = np.array(Image.open(image_path))
+
+    clip_chunk = torch.from_numpy(frame).unsqueeze(0).permute(0, 3, 1, 2)
+    sync_chunk = torch.from_numpy(frame).unsqueeze(0).permute(0, 3, 1, 2)
+
+    clip_frames = clip_transform(clip_chunk)
+    sync_frames = sync_transform(sync_chunk)
+
+    video_info = ImageInfo(
+        clip_frames=clip_frames,
+        sync_frames=sync_frames,
+        original_frame=frame,
+    )
+    return video_info
+
+
+def make_video(video_info: VideoInfo, output_path: Path, audio: torch.Tensor, sampling_rate: int):
+    reencode_with_audio(video_info, output_path, audio, sampling_rate)

mmaudio/ext/__init__.py

@@ -0,0 +1 @@
+

mmaudio/ext/autoencoder/__init__.py

@@ -0,0 +1 @@
+from .autoencoder import AutoEncoderModule

mmaudio/ext/autoencoder/autoencoder.py

@@ -0,0 +1,52 @@
+from typing import Literal, Optional
+
+import torch
+import torch.nn as nn
+
+from mmaudio.ext.autoencoder.vae import VAE, get_my_vae
+from mmaudio.ext.bigvgan import BigVGAN
+from mmaudio.ext.bigvgan_v2.bigvgan import BigVGAN as BigVGANv2
+from mmaudio.model.utils.distributions import DiagonalGaussianDistribution
+
+
+class AutoEncoderModule(nn.Module):
+
+    def __init__(self,
+                 *,
+                 vae_ckpt_path,
+                 vocoder_ckpt_path: Optional[str] = None,
+                 mode: Literal['16k', '44k'],
+                 need_vae_encoder: bool = True):
+        super().__init__()
+        self.vae: VAE = get_my_vae(mode).eval()
+        vae_state_dict = torch.load(vae_ckpt_path, weights_only=True, map_location='cpu')
+        self.vae.load_state_dict(vae_state_dict)
+        self.vae.remove_weight_norm()
+
+        if mode == '16k':
+            assert vocoder_ckpt_path is not None
+            self.vocoder = BigVGAN(vocoder_ckpt_path).eval()
+        elif mode == '44k':
+            self.vocoder = BigVGANv2.from_pretrained('thornmaze/mma-vocoder-44k',
+                                                     use_cuda_kernel=False)
+            self.vocoder.remove_weight_norm()
+        else:
+            raise ValueError(f'Unknown mode: {mode}')
+
+        for param in self.parameters():
+            param.requires_grad = False
+
+        if not need_vae_encoder:
+            del self.vae.encoder
+
+    @torch.inference_mode()
+    def encode(self, x: torch.Tensor) -> DiagonalGaussianDistribution:
+        return self.vae.encode(x)
+
+    @torch.inference_mode()
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        return self.vae.decode(z)
+
+    @torch.inference_mode()
+    def vocode(self, spec: torch.Tensor) -> torch.Tensor:
+        return self.vocoder(spec)

mmaudio/ext/autoencoder/edm2_utils.py

@@ -0,0 +1,168 @@
+# Copyright (c) 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# This work is licensed under a Creative Commons
+# Attribution-NonCommercial-ShareAlike 4.0 International License.
+# You should have received a copy of the license along with this
+# work. If not, see http://creativecommons.org/licenses/by-nc-sa/4.0/
+"""Improved diffusion model architecture proposed in the paper
+"Analyzing and Improving the Training Dynamics of Diffusion Models"."""
+
+import numpy as np
+import torch
+
+#----------------------------------------------------------------------------
+# Variant of constant() that inherits dtype and device from the given
+# reference tensor by default.
+
+_constant_cache = dict()
+
+
+def constant(value, shape=None, dtype=None, device=None, memory_format=None):
+    value = np.asarray(value)
+    if shape is not None:
+        shape = tuple(shape)
+    if dtype is None:
+        dtype = torch.get_default_dtype()
+    if device is None:
+        device = torch.device('cpu')
+    if memory_format is None:
+        memory_format = torch.contiguous_format
+
+    key = (value.shape, value.dtype, value.tobytes(), shape, dtype, device, memory_format)
+    tensor = _constant_cache.get(key, None)
+    if tensor is None:
+        tensor = torch.as_tensor(value.copy(), dtype=dtype, device=device)
+        if shape is not None:
+            tensor, _ = torch.broadcast_tensors(tensor, torch.empty(shape))
+        tensor = tensor.contiguous(memory_format=memory_format)
+        _constant_cache[key] = tensor
+    return tensor
+
+
+def const_like(ref, value, shape=None, dtype=None, device=None, memory_format=None):
+    if dtype is None:
+        dtype = ref.dtype
+    if device is None:
+        device = ref.device
+    return constant(value, shape=shape, dtype=dtype, device=device, memory_format=memory_format)
+
+
+#----------------------------------------------------------------------------
+# Normalize given tensor to unit magnitude with respect to the given
+# dimensions. Default = all dimensions except the first.
+
+
+def normalize(x, dim=None, eps=1e-4):
+    if dim is None:
+        dim = list(range(1, x.ndim))
+    norm = torch.linalg.vector_norm(x, dim=dim, keepdim=True, dtype=torch.float32)
+    norm = torch.add(eps, norm, alpha=np.sqrt(norm.numel() / x.numel()))
+    return x / norm.to(x.dtype)
+
+
+class Normalize(torch.nn.Module):
+
+    def __init__(self, dim=None, eps=1e-4):
+        super().__init__()
+        self.dim = dim
+        self.eps = eps
+
+    def forward(self, x):
+        return normalize(x, dim=self.dim, eps=self.eps)
+
+
+#----------------------------------------------------------------------------
+# Upsample or downsample the given tensor with the given filter,
+# or keep it as is.
+
+
+def resample(x, f=[1, 1], mode='keep'):
+    if mode == 'keep':
+        return x
+    f = np.float32(f)
+    assert f.ndim == 1 and len(f) % 2 == 0
+    pad = (len(f) - 1) // 2
+    f = f / f.sum()
+    f = np.outer(f, f)[np.newaxis, np.newaxis, :, :]
+    f = const_like(x, f)
+    c = x.shape[1]
+    if mode == 'down':
+        return torch.nn.functional.conv2d(x,
+                                          f.tile([c, 1, 1, 1]),
+                                          groups=c,
+                                          stride=2,
+                                          padding=(pad, ))
+    assert mode == 'up'
+    return torch.nn.functional.conv_transpose2d(x, (f * 4).tile([c, 1, 1, 1]),
+                                                groups=c,
+                                                stride=2,
+                                                padding=(pad, ))
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving SiLU (Equation 81).
+
+
+def mp_silu(x):
+    return torch.nn.functional.silu(x) / 0.596
+
+
+class MPSiLU(torch.nn.Module):
+
+    def forward(self, x):
+        return mp_silu(x)
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving sum (Equation 88).
+
+
+def mp_sum(a, b, t=0.5):
+    return a.lerp(b, t) / np.sqrt((1 - t)**2 + t**2)
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving concatenation (Equation 103).
+
+
+def mp_cat(a, b, dim=1, t=0.5):
+    Na = a.shape[dim]
+    Nb = b.shape[dim]
+    C = np.sqrt((Na + Nb) / ((1 - t)**2 + t**2))
+    wa = C / np.sqrt(Na) * (1 - t)
+    wb = C / np.sqrt(Nb) * t
+    return torch.cat([wa * a, wb * b], dim=dim)
+
+
+#----------------------------------------------------------------------------
+# Magnitude-preserving convolution or fully-connected layer (Equation 47)
+# with force weight normalization (Equation 66).
+
+
+class MPConv1D(torch.nn.Module):
+
+    def __init__(self, in_channels, out_channels, kernel_size):
+        super().__init__()
+        self.out_channels = out_channels
+        self.weight = torch.nn.Parameter(torch.randn(out_channels, in_channels, kernel_size))
+
+        self.weight_norm_removed = False
+
+    def forward(self, x, gain=1):
+        assert self.weight_norm_removed, 'call remove_weight_norm() before inference'
+
+        w = self.weight * gain
+        if w.ndim == 2:
+            return x @ w.t()
+        assert w.ndim == 3
+        return torch.nn.functional.conv1d(x, w, padding=(w.shape[-1] // 2, ))
+
+    def remove_weight_norm(self):
+        w = self.weight.to(torch.float32)
+        w = normalize(w)  # traditional weight normalization
+        w = w / np.sqrt(w[0].numel())
+        w = w.to(self.weight.dtype)
+        self.weight.data.copy_(w)
+
+        self.weight_norm_removed = True
+        return self

mmaudio/ext/autoencoder/vae.py

@@ -0,0 +1,369 @@
+import logging
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from mmaudio.ext.autoencoder.edm2_utils import MPConv1D
+from mmaudio.ext.autoencoder.vae_modules import (AttnBlock1D, Downsample1D, ResnetBlock1D,
+                                                 Upsample1D, nonlinearity)
+from mmaudio.model.utils.distributions import DiagonalGaussianDistribution
+
+log = logging.getLogger()
+
+DATA_MEAN_80D = [
+    -1.6058, -1.3676, -1.2520, -1.2453, -1.2078, -1.2224, -1.2419, -1.2439, -1.2922, -1.2927,
+    -1.3170, -1.3543, -1.3401, -1.3836, -1.3907, -1.3912, -1.4313, -1.4152, -1.4527, -1.4728,
+    -1.4568, -1.5101, -1.5051, -1.5172, -1.5623, -1.5373, -1.5746, -1.5687, -1.6032, -1.6131,
+    -1.6081, -1.6331, -1.6489, -1.6489, -1.6700, -1.6738, -1.6953, -1.6969, -1.7048, -1.7280,
+    -1.7361, -1.7495, -1.7658, -1.7814, -1.7889, -1.8064, -1.8221, -1.8377, -1.8417, -1.8643,
+    -1.8857, -1.8929, -1.9173, -1.9379, -1.9531, -1.9673, -1.9824, -2.0042, -2.0215, -2.0436,
+    -2.0766, -2.1064, -2.1418, -2.1855, -2.2319, -2.2767, -2.3161, -2.3572, -2.3954, -2.4282,
+    -2.4659, -2.5072, -2.5552, -2.6074, -2.6584, -2.7107, -2.7634, -2.8266, -2.8981, -2.9673
+]
+
+DATA_STD_80D = [
+    1.0291, 1.0411, 1.0043, 0.9820, 0.9677, 0.9543, 0.9450, 0.9392, 0.9343, 0.9297, 0.9276, 0.9263,
+    0.9242, 0.9254, 0.9232, 0.9281, 0.9263, 0.9315, 0.9274, 0.9247, 0.9277, 0.9199, 0.9188, 0.9194,
+    0.9160, 0.9161, 0.9146, 0.9161, 0.9100, 0.9095, 0.9145, 0.9076, 0.9066, 0.9095, 0.9032, 0.9043,
+    0.9038, 0.9011, 0.9019, 0.9010, 0.8984, 0.8983, 0.8986, 0.8961, 0.8962, 0.8978, 0.8962, 0.8973,
+    0.8993, 0.8976, 0.8995, 0.9016, 0.8982, 0.8972, 0.8974, 0.8949, 0.8940, 0.8947, 0.8936, 0.8939,
+    0.8951, 0.8956, 0.9017, 0.9167, 0.9436, 0.9690, 1.0003, 1.0225, 1.0381, 1.0491, 1.0545, 1.0604,
+    1.0761, 1.0929, 1.1089, 1.1196, 1.1176, 1.1156, 1.1117, 1.1070
+]
+
+DATA_MEAN_128D = [
+    -3.3462, -2.6723, -2.4893, -2.3143, -2.2664, -2.3317, -2.1802, -2.4006, -2.2357, -2.4597,
+    -2.3717, -2.4690, -2.5142, -2.4919, -2.6610, -2.5047, -2.7483, -2.5926, -2.7462, -2.7033,
+    -2.7386, -2.8112, -2.7502, -2.9594, -2.7473, -3.0035, -2.8891, -2.9922, -2.9856, -3.0157,
+    -3.1191, -2.9893, -3.1718, -3.0745, -3.1879, -3.2310, -3.1424, -3.2296, -3.2791, -3.2782,
+    -3.2756, -3.3134, -3.3509, -3.3750, -3.3951, -3.3698, -3.4505, -3.4509, -3.5089, -3.4647,
+    -3.5536, -3.5788, -3.5867, -3.6036, -3.6400, -3.6747, -3.7072, -3.7279, -3.7283, -3.7795,
+    -3.8259, -3.8447, -3.8663, -3.9182, -3.9605, -3.9861, -4.0105, -4.0373, -4.0762, -4.1121,
+    -4.1488, -4.1874, -4.2461, -4.3170, -4.3639, -4.4452, -4.5282, -4.6297, -4.7019, -4.7960,
+    -4.8700, -4.9507, -5.0303, -5.0866, -5.1634, -5.2342, -5.3242, -5.4053, -5.4927, -5.5712,
+    -5.6464, -5.7052, -5.7619, -5.8410, -5.9188, -6.0103, -6.0955, -6.1673, -6.2362, -6.3120,
+    -6.3926, -6.4797, -6.5565, -6.6511, -6.8130, -6.9961, -7.1275, -7.2457, -7.3576, -7.4663,
+    -7.6136, -7.7469, -7.8815, -8.0132, -8.1515, -8.3071, -8.4722, -8.7418, -9.3975, -9.6628,
+    -9.7671, -9.8863, -9.9992, -10.0860, -10.1709, -10.5418, -11.2795, -11.3861
+]
+
+DATA_STD_128D = [
+    2.3804, 2.4368, 2.3772, 2.3145, 2.2803, 2.2510, 2.2316, 2.2083, 2.1996, 2.1835, 2.1769, 2.1659,
+    2.1631, 2.1618, 2.1540, 2.1606, 2.1571, 2.1567, 2.1612, 2.1579, 2.1679, 2.1683, 2.1634, 2.1557,
+    2.1668, 2.1518, 2.1415, 2.1449, 2.1406, 2.1350, 2.1313, 2.1415, 2.1281, 2.1352, 2.1219, 2.1182,
+    2.1327, 2.1195, 2.1137, 2.1080, 2.1179, 2.1036, 2.1087, 2.1036, 2.1015, 2.1068, 2.0975, 2.0991,
+    2.0902, 2.1015, 2.0857, 2.0920, 2.0893, 2.0897, 2.0910, 2.0881, 2.0925, 2.0873, 2.0960, 2.0900,
+    2.0957, 2.0958, 2.0978, 2.0936, 2.0886, 2.0905, 2.0845, 2.0855, 2.0796, 2.0840, 2.0813, 2.0817,
+    2.0838, 2.0840, 2.0917, 2.1061, 2.1431, 2.1976, 2.2482, 2.3055, 2.3700, 2.4088, 2.4372, 2.4609,
+    2.4731, 2.4847, 2.5072, 2.5451, 2.5772, 2.6147, 2.6529, 2.6596, 2.6645, 2.6726, 2.6803, 2.6812,
+    2.6899, 2.6916, 2.6931, 2.6998, 2.7062, 2.7262, 2.7222, 2.7158, 2.7041, 2.7485, 2.7491, 2.7451,
+    2.7485, 2.7233, 2.7297, 2.7233, 2.7145, 2.6958, 2.6788, 2.6439, 2.6007, 2.4786, 2.2469, 2.1877,
+    2.1392, 2.0717, 2.0107, 1.9676, 1.9140, 1.7102, 0.9101, 0.7164
+]
+
+
+class VAE(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        data_dim: int,
+        embed_dim: int,
+        hidden_dim: int,
+    ):
+        super().__init__()
+
+        if data_dim == 80:
+            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_80D, dtype=torch.float32))
+            self.data_std = nn.Buffer(torch.tensor(DATA_STD_80D, dtype=torch.float32))
+        elif data_dim == 128:
+            self.data_mean = nn.Buffer(torch.tensor(DATA_MEAN_128D, dtype=torch.float32))
+            self.data_std = nn.Buffer(torch.tensor(DATA_STD_128D, dtype=torch.float32))
+
+        self.data_mean = self.data_mean.view(1, -1, 1)
+        self.data_std = self.data_std.view(1, -1, 1)
+
+        self.encoder = Encoder1D(
+            dim=hidden_dim,
+            ch_mult=(1, 2, 4),
+            num_res_blocks=2,
+            attn_layers=[3],
+            down_layers=[0],
+            in_dim=data_dim,
+            embed_dim=embed_dim,
+        )
+        self.decoder = Decoder1D(
+            dim=hidden_dim,
+            ch_mult=(1, 2, 4),
+            num_res_blocks=2,
+            attn_layers=[3],
+            down_layers=[0],
+            in_dim=data_dim,
+            out_dim=data_dim,
+            embed_dim=embed_dim,
+        )
+
+        self.embed_dim = embed_dim
+        # self.quant_conv = nn.Conv1d(2 * embed_dim, 2 * embed_dim, 1)
+        # self.post_quant_conv = nn.Conv1d(embed_dim, embed_dim, 1)
+
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        pass
+
+    def encode(self, x: torch.Tensor, normalize: bool = True) -> DiagonalGaussianDistribution:
+        if normalize:
+            x = self.normalize(x)
+        moments = self.encoder(x)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z: torch.Tensor, unnormalize: bool = True) -> torch.Tensor:
+        dec = self.decoder(z)
+        if unnormalize:
+            dec = self.unnormalize(dec)
+        return dec
+
+    def normalize(self, x: torch.Tensor) -> torch.Tensor:
+        return (x - self.data_mean) / self.data_std
+
+    def unnormalize(self, x: torch.Tensor) -> torch.Tensor:
+        return x * self.data_std + self.data_mean
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        sample_posterior: bool = True,
+        rng: Optional[torch.Generator] = None,
+        normalize: bool = True,
+        unnormalize: bool = True,
+    ) -> tuple[torch.Tensor, DiagonalGaussianDistribution]:
+
+        posterior = self.encode(x, normalize=normalize)
+        if sample_posterior:
+            z = posterior.sample(rng)
+        else:
+            z = posterior.mode()
+        dec = self.decode(z, unnormalize=unnormalize)
+        return dec, posterior
+
+    def load_weights(self, src_dict) -> None:
+        self.load_state_dict(src_dict, strict=True)
+
+    @property
+    def device(self) -> torch.device:
+        return next(self.parameters()).device
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def remove_weight_norm(self):
+        for name, m in self.named_modules():
+            if isinstance(m, MPConv1D):
+                m.remove_weight_norm()
+                log.debug(f"Removed weight norm from {name}")
+        return self
+
+
+class Encoder1D(nn.Module):
+
+    def __init__(self,
+                 *,
+                 dim: int,
+                 ch_mult: tuple[int] = (1, 2, 4, 8),
+                 num_res_blocks: int,
+                 attn_layers: list[int] = [],
+                 down_layers: list[int] = [],
+                 resamp_with_conv: bool = True,
+                 in_dim: int,
+                 embed_dim: int,
+                 double_z: bool = True,
+                 kernel_size: int = 3,
+                 clip_act: float = 256.0):
+        super().__init__()
+        self.dim = dim
+        self.num_layers = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.in_channels = in_dim
+        self.clip_act = clip_act
+        self.down_layers = down_layers
+        self.attn_layers = attn_layers
+        self.conv_in = MPConv1D(in_dim, self.dim, kernel_size=kernel_size)
+
+        in_ch_mult = (1, ) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        # downsampling
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_layers):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = dim * in_ch_mult[i_level]
+            block_out = dim * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock1D(in_dim=block_in,
+                                  out_dim=block_out,
+                                  kernel_size=kernel_size,
+                                  use_norm=True))
+                block_in = block_out
+                if i_level in attn_layers:
+                    attn.append(AttnBlock1D(block_in))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level in down_layers:
+                down.downsample = Downsample1D(block_in, resamp_with_conv)
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock1D(in_dim=block_in,
+                                         out_dim=block_in,
+                                         kernel_size=kernel_size,
+                                         use_norm=True)
+        self.mid.attn_1 = AttnBlock1D(block_in)
+        self.mid.block_2 = ResnetBlock1D(in_dim=block_in,
+                                         out_dim=block_in,
+                                         kernel_size=kernel_size,
+                                         use_norm=True)
+
+        # end
+        self.conv_out = MPConv1D(block_in,
+                                 2 * embed_dim if double_z else embed_dim,
+                                 kernel_size=kernel_size)
+
+        self.learnable_gain = nn.Parameter(torch.zeros([]))
+
+    def forward(self, x):
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_layers):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1])
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                h = h.clamp(-self.clip_act, self.clip_act)
+                hs.append(h)
+            if i_level in self.down_layers:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        h = h.clamp(-self.clip_act, self.clip_act)
+
+        # end
+        h = nonlinearity(h)
+        h = self.conv_out(h, gain=(self.learnable_gain + 1))
+        return h
+
+
+class Decoder1D(nn.Module):
+
+    def __init__(self,
+                 *,
+                 dim: int,
+                 out_dim: int,
+                 ch_mult: tuple[int] = (1, 2, 4, 8),
+                 num_res_blocks: int,
+                 attn_layers: list[int] = [],
+                 down_layers: list[int] = [],
+                 kernel_size: int = 3,
+                 resamp_with_conv: bool = True,
+                 in_dim: int,
+                 embed_dim: int,
+                 clip_act: float = 256.0):
+        super().__init__()
+        self.ch = dim
+        self.num_layers = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.in_channels = in_dim
+        self.clip_act = clip_act
+        self.down_layers = [i + 1 for i in down_layers]  # each downlayer add one
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = dim * ch_mult[self.num_layers - 1]
+
+        # z to block_in
+        self.conv_in = MPConv1D(embed_dim, block_in, kernel_size=kernel_size)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
+        self.mid.attn_1 = AttnBlock1D(block_in)
+        self.mid.block_2 = ResnetBlock1D(in_dim=block_in, out_dim=block_in, use_norm=True)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_layers)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = dim * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock1D(in_dim=block_in, out_dim=block_out, use_norm=True))
+                block_in = block_out
+                if i_level in attn_layers:
+                    attn.append(AttnBlock1D(block_in))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level in self.down_layers:
+                up.upsample = Upsample1D(block_in, resamp_with_conv)
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.conv_out = MPConv1D(block_in, out_dim, kernel_size=kernel_size)
+        self.learnable_gain = nn.Parameter(torch.zeros([]))
+
+    def forward(self, z):
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        h = h.clamp(-self.clip_act, self.clip_act)
+
+        # upsampling
+        for i_level in reversed(range(self.num_layers)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+                h = h.clamp(-self.clip_act, self.clip_act)
+            if i_level in self.down_layers:
+                h = self.up[i_level].upsample(h)
+
+        h = nonlinearity(h)
+        h = self.conv_out(h, gain=(self.learnable_gain + 1))
+        return h
+
+
+def VAE_16k(**kwargs) -> VAE:
+    return VAE(data_dim=80, embed_dim=20, hidden_dim=384, **kwargs)
+
+
+def VAE_44k(**kwargs) -> VAE:
+    return VAE(data_dim=128, embed_dim=40, hidden_dim=512, **kwargs)
+
+
+def get_my_vae(name: str, **kwargs) -> VAE:
+    if name == '16k':
+        return VAE_16k(**kwargs)
+    if name == '44k':
+        return VAE_44k(**kwargs)
+    raise ValueError(f'Unknown model: {name}')
+
+
+if __name__ == '__main__':
+    network = get_my_vae('standard')
+
+    # print the number of parameters in terms of millions
+    num_params = sum(p.numel() for p in network.parameters()) / 1e6
+    print(f'Number of parameters: {num_params:.2f}M')

mmaudio/ext/autoencoder/vae_modules.py

@@ -0,0 +1,117 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from mmaudio.ext.autoencoder.edm2_utils import (MPConv1D, mp_silu, mp_sum, normalize)
+
+
+def nonlinearity(x):
+    # swish
+    return mp_silu(x)
+
+
+class ResnetBlock1D(nn.Module):
+
+    def __init__(self, *, in_dim, out_dim=None, conv_shortcut=False, kernel_size=3, use_norm=True):
+        super().__init__()
+        self.in_dim = in_dim
+        out_dim = in_dim if out_dim is None else out_dim
+        self.out_dim = out_dim
+        self.use_conv_shortcut = conv_shortcut
+        self.use_norm = use_norm
+
+        self.conv1 = MPConv1D(in_dim, out_dim, kernel_size=kernel_size)
+        self.conv2 = MPConv1D(out_dim, out_dim, kernel_size=kernel_size)
+        if self.in_dim != self.out_dim:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = MPConv1D(in_dim, out_dim, kernel_size=kernel_size)
+            else:
+                self.nin_shortcut = MPConv1D(in_dim, out_dim, kernel_size=1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+
+        # pixel norm
+        if self.use_norm:
+            x = normalize(x, dim=1)
+
+        h = x
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        h = nonlinearity(h)
+        h = self.conv2(h)
+
+        if self.in_dim != self.out_dim:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return mp_sum(x, h, t=0.3)
+
+
+class AttnBlock1D(nn.Module):
+
+    def __init__(self, in_channels, num_heads=1):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.num_heads = num_heads
+        self.qkv = MPConv1D(in_channels, in_channels * 3, kernel_size=1)
+        self.proj_out = MPConv1D(in_channels, in_channels, kernel_size=1)
+
+    def forward(self, x):
+        h = x
+        y = self.qkv(h)
+        y = y.reshape(y.shape[0], self.num_heads, -1, 3, y.shape[-1])
+        q, k, v = normalize(y, dim=2).unbind(3)
+
+        q = rearrange(q, 'b h c l -> b h l c')
+        k = rearrange(k, 'b h c l -> b h l c')
+        v = rearrange(v, 'b h c l -> b h l c')
+
+        h = F.scaled_dot_product_attention(q, k, v)
+        h = rearrange(h, 'b h l c -> b (h c) l')
+
+        h = self.proj_out(h)
+
+        return mp_sum(x, h, t=0.3)
+
+
+class Upsample1D(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = MPConv1D(in_channels, in_channels, kernel_size=3)
+
+    def forward(self, x):
+        x = F.interpolate(x, scale_factor=2.0, mode='nearest-exact')  # support 3D tensor(B,C,T)
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample1D(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv1 = MPConv1D(in_channels, in_channels, kernel_size=1)
+            self.conv2 = MPConv1D(in_channels, in_channels, kernel_size=1)
+
+    def forward(self, x):
+
+        if self.with_conv:
+            x = self.conv1(x)
+
+        x = F.avg_pool1d(x, kernel_size=2, stride=2)
+
+        if self.with_conv:
+            x = self.conv2(x)
+
+        return x

mmaudio/ext/bigvgan/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 NVIDIA CORPORATION.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software. 
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mmaudio/ext/bigvgan/__init__.py

@@ -0,0 +1 @@
+from .bigvgan import BigVGAN

mmaudio/ext/bigvgan/activations.py

@@ -0,0 +1,120 @@
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+
+
+class Snake(nn.Module):
+    '''
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale: # log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+        else: # linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        '''
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x
+
+
+class SnakeBeta(nn.Module):
+    '''
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    '''
+    def __init__(self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False):
+        '''
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha - trainable parameter that controls frequency
+            - beta - trainable parameter that controls magnitude
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+        '''
+        super(SnakeBeta, self).__init__()
+        self.in_features = in_features
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale: # log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+            self.beta = Parameter(torch.zeros(in_features) * alpha)
+        else: # linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+            self.beta = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        '''
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        '''
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1) # line up with x to [B, C, T]
+        beta = self.beta.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

mmaudio/ext/bigvgan/alias_free_torch/__init__.py

@@ -0,0 +1,6 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+from .filter import *
+from .resample import *
+from .act import *

mmaudio/ext/bigvgan/alias_free_torch/act.py

@@ -0,0 +1,28 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch.nn as nn
+from .resample import UpSample1d, DownSample1d
+
+
+class Activation1d(nn.Module):
+    def __init__(self,
+                 activation,
+                 up_ratio: int = 2,
+                 down_ratio: int = 2,
+                 up_kernel_size: int = 12,
+                 down_kernel_size: int = 12):
+        super().__init__()
+        self.up_ratio = up_ratio
+        self.down_ratio = down_ratio
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+    # x: [B,C,T]
+    def forward(self, x):
+        x = self.upsample(x)
+        x = self.act(x)
+        x = self.downsample(x)
+
+        return x

mmaudio/ext/bigvgan/alias_free_torch/filter.py

@@ -0,0 +1,95 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+if 'sinc' in dir(torch):
+    sinc = torch.sinc
+else:
+    # This code is adopted from adefossez's julius.core.sinc under the MIT License
+    # https://adefossez.github.io/julius/julius/core.html
+    #   LICENSE is in incl_licenses directory.
+    def sinc(x: torch.Tensor):
+        """
+        Implementation of sinc, i.e. sin(pi * x) / (pi * x)
+        __Warning__: Different to julius.sinc, the input is multiplied by `pi`!
+        """
+        return torch.where(x == 0,
+                           torch.tensor(1., device=x.device, dtype=x.dtype),
+                           torch.sin(math.pi * x) / math.pi / x)
+
+
+# This code is adopted from adefossez's julius.lowpass.LowPassFilters under the MIT License
+# https://adefossez.github.io/julius/julius/lowpass.html
+#   LICENSE is in incl_licenses directory.
+def kaiser_sinc_filter1d(cutoff, half_width, kernel_size): # return filter [1,1,kernel_size]
+    even = (kernel_size % 2 == 0)
+    half_size = kernel_size // 2
+
+    #For kaiser window
+    delta_f = 4 * half_width
+    A = 2.285 * (half_size - 1) * math.pi * delta_f + 7.95
+    if A > 50.:
+        beta = 0.1102 * (A - 8.7)
+    elif A >= 21.:
+        beta = 0.5842 * (A - 21)**0.4 + 0.07886 * (A - 21.)
+    else:
+        beta = 0.
+    window = torch.kaiser_window(kernel_size, beta=beta, periodic=False)
+
+    # ratio = 0.5/cutoff -> 2 * cutoff = 1 / ratio
+    if even:
+        time = (torch.arange(-half_size, half_size) + 0.5)
+    else:
+        time = torch.arange(kernel_size) - half_size
+    if cutoff == 0:
+        filter_ = torch.zeros_like(time)
+    else:
+        filter_ = 2 * cutoff * window * sinc(2 * cutoff * time)
+        # Normalize filter to have sum = 1, otherwise we will have a small leakage
+        # of the constant component in the input signal.
+        filter_ /= filter_.sum()
+        filter = filter_.view(1, 1, kernel_size)
+
+    return filter
+
+
+class LowPassFilter1d(nn.Module):
+    def __init__(self,
+                 cutoff=0.5,
+                 half_width=0.6,
+                 stride: int = 1,
+                 padding: bool = True,
+                 padding_mode: str = 'replicate',
+                 kernel_size: int = 12):
+        # kernel_size should be even number for stylegan3 setup,
+        # in this implementation, odd number is also possible.
+        super().__init__()
+        if cutoff < -0.:
+            raise ValueError("Minimum cutoff must be larger than zero.")
+        if cutoff > 0.5:
+            raise ValueError("A cutoff above 0.5 does not make sense.")
+        self.kernel_size = kernel_size
+        self.even = (kernel_size % 2 == 0)
+        self.pad_left = kernel_size // 2 - int(self.even)
+        self.pad_right = kernel_size // 2
+        self.stride = stride
+        self.padding = padding
+        self.padding_mode = padding_mode
+        filter = kaiser_sinc_filter1d(cutoff, half_width, kernel_size)
+        self.register_buffer("filter", filter)
+
+    #input [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        if self.padding:
+            x = F.pad(x, (self.pad_left, self.pad_right),
+                      mode=self.padding_mode)
+        out = F.conv1d(x, self.filter.expand(C, -1, -1),
+                       stride=self.stride, groups=C)
+
+        return out

mmaudio/ext/bigvgan/alias_free_torch/resample.py

@@ -0,0 +1,49 @@
+# Adapted from https://github.com/junjun3518/alias-free-torch under the Apache License 2.0
+#   LICENSE is in incl_licenses directory.
+
+import torch.nn as nn
+from torch.nn import functional as F
+from .filter import LowPassFilter1d
+from .filter import kaiser_sinc_filter1d
+
+
+class UpSample1d(nn.Module):
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.stride = ratio
+        self.pad = self.kernel_size // ratio - 1
+        self.pad_left = self.pad * self.stride + (self.kernel_size - self.stride) // 2
+        self.pad_right = self.pad * self.stride + (self.kernel_size - self.stride + 1) // 2
+        filter = kaiser_sinc_filter1d(cutoff=0.5 / ratio,
+                                      half_width=0.6 / ratio,
+                                      kernel_size=self.kernel_size)
+        self.register_buffer("filter", filter)
+
+    # x: [B, C, T]
+    def forward(self, x):
+        _, C, _ = x.shape
+
+        x = F.pad(x, (self.pad, self.pad), mode='replicate')
+        x = self.ratio * F.conv_transpose1d(
+            x, self.filter.expand(C, -1, -1), stride=self.stride, groups=C)
+        x = x[..., self.pad_left:-self.pad_right]
+
+        return x
+
+
+class DownSample1d(nn.Module):
+    def __init__(self, ratio=2, kernel_size=None):
+        super().__init__()
+        self.ratio = ratio
+        self.kernel_size = int(6 * ratio // 2) * 2 if kernel_size is None else kernel_size
+        self.lowpass = LowPassFilter1d(cutoff=0.5 / ratio,
+                                       half_width=0.6 / ratio,
+                                       stride=ratio,
+                                       kernel_size=self.kernel_size)
+
+    def forward(self, x):
+        xx = self.lowpass(x)
+
+        return xx

mmaudio/ext/bigvgan/bigvgan.py

@@ -0,0 +1,32 @@
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from omegaconf import OmegaConf
+
+from mmaudio.ext.bigvgan.models import BigVGANVocoder
+
+_bigvgan_vocoder_path = Path(__file__).parent / 'bigvgan_vocoder.yml'
+
+
+class BigVGAN(nn.Module):
+
+    def __init__(self, ckpt_path, config_path=_bigvgan_vocoder_path):
+        super().__init__()
+        vocoder_cfg = OmegaConf.load(config_path)
+        self.vocoder = BigVGANVocoder(vocoder_cfg).eval()
+        vocoder_ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=True)['generator']
+        self.vocoder.load_state_dict(vocoder_ckpt)
+
+        self.weight_norm_removed = False
+        self.remove_weight_norm()
+
+    @torch.inference_mode()
+    def forward(self, x):
+        assert self.weight_norm_removed, 'call remove_weight_norm() before inference'
+        return self.vocoder(x)
+
+    def remove_weight_norm(self):
+        self.vocoder.remove_weight_norm()
+        self.weight_norm_removed = True
+        return self

mmaudio/ext/bigvgan/bigvgan_vocoder.yml

@@ -0,0 +1,63 @@
+resblock: '1'
+num_gpus: 0
+batch_size: 64
+num_mels: 80
+learning_rate: 0.0001
+adam_b1: 0.8
+adam_b2: 0.99
+lr_decay: 0.999
+seed: 1234
+upsample_rates:
+- 4
+- 4
+- 2
+- 2
+- 2
+- 2
+upsample_kernel_sizes:
+- 8
+- 8
+- 4
+- 4
+- 4
+- 4
+upsample_initial_channel: 1536
+resblock_kernel_sizes:
+- 3
+- 7
+- 11
+resblock_dilation_sizes:
+- - 1
+  - 3
+  - 5
+- - 1
+  - 3
+  - 5
+- - 1
+  - 3
+  - 5
+activation: snakebeta
+snake_logscale: true
+resolutions:
+- - 1024
+  - 120
+  - 600
+- - 2048
+  - 240
+  - 1200
+- - 512
+  - 50
+  - 240
+mpd_reshapes:
+- 2
+- 3
+- 5
+- 7
+- 11
+use_spectral_norm: false
+discriminator_channel_mult: 1
+num_workers: 4
+dist_config:
+  dist_backend: nccl
+  dist_url: tcp://localhost:54341
+  world_size: 1

mmaudio/ext/bigvgan/env.py

@@ -0,0 +1,18 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import os
+import shutil
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+
+def build_env(config, config_name, path):
+    t_path = os.path.join(path, config_name)
+    if config != t_path:
+        os.makedirs(path, exist_ok=True)
+        shutil.copyfile(config, os.path.join(path, config_name))

mmaudio/ext/bigvgan/incl_licenses/LICENSE_1

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Jungil Kong
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mmaudio/ext/bigvgan/incl_licenses/LICENSE_2

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2020 Edward Dixon
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mmaudio/ext/bigvgan/incl_licenses/LICENSE_3

@@ -0,0 +1,201 @@
+                                Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

mmaudio/ext/bigvgan/incl_licenses/LICENSE_4

@@ -0,0 +1,29 @@
+BSD 3-Clause License
+
+Copyright (c) 2019, Seungwon Park 박승원
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+3. Neither the name of the copyright holder nor the names of its
+   contributors may be used to endorse or promote products derived from
+   this software without specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

mmaudio/ext/bigvgan/incl_licenses/LICENSE_5

@@ -0,0 +1,16 @@
+Copyright 2020 Alexandre Défossez
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
+associated documentation files (the "Software"), to deal in the Software without restriction,
+including without limitation the rights to use, copy, modify, merge, publish, distribute,
+sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or
+substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
+NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
+DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

mmaudio/ext/bigvgan/models.py

@@ -0,0 +1,255 @@
+# Copyright (c) 2022 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+import torch.nn as nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils.parametrizations import weight_norm
+from torch.nn.utils.parametrize import remove_parametrizations
+
+from mmaudio.ext.bigvgan import activations
+from mmaudio.ext.bigvgan.alias_free_torch import *
+from mmaudio.ext.bigvgan.utils import get_padding, init_weights
+
+LRELU_SLOPE = 0.1
+
+
+class AMPBlock1(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3, 5), activation=None):
+        super(AMPBlock1, self).__init__()
+        self.h = h
+
+        self.convs1 = nn.ModuleList([
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[0],
+                       padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[1],
+                       padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[2],
+                       padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=1,
+                       padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+
+        self.num_layers = len(self.convs1) + len(self.convs2)  # total number of conv layers
+
+        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        acts1, acts2 = self.activations[::2], self.activations[1::2]
+        for c1, c2, a1, a2 in zip(self.convs1, self.convs2, acts1, acts2):
+            xt = a1(x)
+            xt = c1(xt)
+            xt = a2(xt)
+            xt = c2(xt)
+            x = xt + x
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_parametrizations(l, 'weight')
+        for l in self.convs2:
+            remove_parametrizations(l, 'weight')
+
+
+class AMPBlock2(torch.nn.Module):
+
+    def __init__(self, h, channels, kernel_size=3, dilation=(1, 3), activation=None):
+        super(AMPBlock2, self).__init__()
+        self.h = h
+
+        self.convs = nn.ModuleList([
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[0],
+                       padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(
+                Conv1d(channels,
+                       channels,
+                       kernel_size,
+                       1,
+                       dilation=dilation[1],
+                       padding=get_padding(kernel_size, dilation[1])))
+        ])
+        self.convs.apply(init_weights)
+
+        self.num_layers = len(self.convs)  # total number of conv layers
+
+        if activation == 'snake':  # periodic nonlinearity with snake function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.Snake(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        elif activation == 'snakebeta':  # periodic nonlinearity with snakebeta function and anti-aliasing
+            self.activations = nn.ModuleList([
+                Activation1d(
+                    activation=activations.SnakeBeta(channels, alpha_logscale=h.snake_logscale))
+                for _ in range(self.num_layers)
+            ])
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+    def forward(self, x):
+        for c, a in zip(self.convs, self.activations):
+            xt = a(x)
+            xt = c(xt)
+            x = xt + x
+
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_parametrizations(l, 'weight')
+
+
+class BigVGANVocoder(torch.nn.Module):
+    # this is our main BigVGAN model. Applies anti-aliased periodic activation for resblocks.
+    def __init__(self, h):
+        super().__init__()
+        self.h = h
+
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+
+        # pre conv
+        self.conv_pre = weight_norm(Conv1d(h.num_mels, h.upsample_initial_channel, 7, 1, padding=3))
+
+        # define which AMPBlock to use. BigVGAN uses AMPBlock1 as default
+        resblock = AMPBlock1 if h.resblock == '1' else AMPBlock2
+
+        # transposed conv-based upsamplers. does not apply anti-aliasing
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                nn.ModuleList([
+                    weight_norm(
+                        ConvTranspose1d(h.upsample_initial_channel // (2**i),
+                                        h.upsample_initial_channel // (2**(i + 1)),
+                                        k,
+                                        u,
+                                        padding=(k - u) // 2))
+                ]))
+
+        # residual blocks using anti-aliased multi-periodicity composition modules (AMP)
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2**(i + 1))
+            for j, (k, d) in enumerate(zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)):
+                self.resblocks.append(resblock(h, ch, k, d, activation=h.activation))
+
+        # post conv
+        if h.activation == "snake":  # periodic nonlinearity with snake function and anti-aliasing
+            activation_post = activations.Snake(ch, alpha_logscale=h.snake_logscale)
+            self.activation_post = Activation1d(activation=activation_post)
+        elif h.activation == "snakebeta":  # periodic nonlinearity with snakebeta function and anti-aliasing
+            activation_post = activations.SnakeBeta(ch, alpha_logscale=h.snake_logscale)
+            self.activation_post = Activation1d(activation=activation_post)
+        else:
+            raise NotImplementedError(
+                "activation incorrectly specified. check the config file and look for 'activation'."
+            )
+
+        self.conv_post = weight_norm(Conv1d(ch, 1, 7, 1, padding=3))
+
+        # weight initialization
+        for i in range(len(self.ups)):
+            self.ups[i].apply(init_weights)
+        self.conv_post.apply(init_weights)
+
+    def forward(self, x):
+        # pre conv
+        x = self.conv_pre(x)
+
+        for i in range(self.num_upsamples):
+            # upsampling
+            for i_up in range(len(self.ups[i])):
+                x = self.ups[i][i_up](x)
+            # AMP blocks
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+
+        # post conv
+        x = self.activation_post(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            for l_i in l:
+                remove_parametrizations(l_i, 'weight')
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_parametrizations(self.conv_pre, 'weight')
+        remove_parametrizations(self.conv_post, 'weight')

mmaudio/ext/bigvgan/utils.py

@@ -0,0 +1,31 @@
+# Adapted from https://github.com/jik876/hifi-gan under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import os
+
+import torch
+from torch.nn.utils.parametrizations import weight_norm
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+def apply_weight_norm(m):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        weight_norm(m)
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def load_checkpoint(filepath, device):
+    assert os.path.isfile(filepath)
+    print("Loading '{}'".format(filepath))
+    checkpoint_dict = torch.load(filepath, map_location=device)
+    print("Complete.")
+    return checkpoint_dict

mmaudio/ext/bigvgan_v2/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 NVIDIA CORPORATION.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mmaudio/ext/bigvgan_v2/activations.py

@@ -0,0 +1,126 @@
+# Implementation adapted from https://github.com/EdwardDixon/snake under the MIT license.
+#   LICENSE is in incl_licenses directory.
+
+import torch
+from torch import nn, sin, pow
+from torch.nn import Parameter
+
+
+class Snake(nn.Module):
+    """
+    Implementation of a sine-based periodic activation function
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter
+    References:
+        - This activation function is from this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snake(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    """
+
+    def __init__(
+        self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False
+    ):
+        """
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha: trainable parameter
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            alpha will be trained along with the rest of your model.
+        """
+        super(Snake, self).__init__()
+        self.in_features = in_features
+
+        # Initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # Log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+        else:  # Linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        Snake ∶= x + 1/a * sin^2 (xa)
+        """
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # Line up with x to [B, C, T]
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+        x = x + (1.0 / (alpha + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x
+
+
+class SnakeBeta(nn.Module):
+    """
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+    """
+
+    def __init__(
+        self, in_features, alpha=1.0, alpha_trainable=True, alpha_logscale=False
+    ):
+        """
+        Initialization.
+        INPUT:
+            - in_features: shape of the input
+            - alpha - trainable parameter that controls frequency
+            - beta - trainable parameter that controls magnitude
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+        """
+        super(SnakeBeta, self).__init__()
+        self.in_features = in_features
+
+        # Initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # Log scale alphas initialized to zeros
+            self.alpha = Parameter(torch.zeros(in_features) * alpha)
+            self.beta = Parameter(torch.zeros(in_features) * alpha)
+        else:  # Linear scale alphas initialized to ones
+            self.alpha = Parameter(torch.ones(in_features) * alpha)
+            self.beta = Parameter(torch.ones(in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        """
+        alpha = self.alpha.unsqueeze(0).unsqueeze(-1)  # Line up with x to [B, C, T]
+        beta = self.beta.unsqueeze(0).unsqueeze(-1)
+        if self.alpha_logscale:
+            alpha = torch.exp(alpha)
+            beta = torch.exp(beta)
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * pow(sin(x * alpha), 2)
+
+        return x

mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/activation1d.py

@@ -0,0 +1,77 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+import torch
+import torch.nn as nn
+from alias_free_activation.torch.resample import UpSample1d, DownSample1d
+
+# load fused CUDA kernel: this enables importing anti_alias_activation_cuda
+from alias_free_activation.cuda import load
+
+anti_alias_activation_cuda = load.load()
+
+
+class FusedAntiAliasActivation(torch.autograd.Function):
+    """
+    Assumes filter size 12, replication padding on upsampling/downsampling, and logscale alpha/beta parameters as inputs.
+    The hyperparameters are hard-coded in the kernel to maximize speed.
+    NOTE: The fused kenrel is incorrect for Activation1d with different hyperparameters.
+    """
+
+    @staticmethod
+    def forward(ctx, inputs, up_ftr, down_ftr, alpha, beta):
+        activation_results = anti_alias_activation_cuda.forward(
+            inputs, up_ftr, down_ftr, alpha, beta
+        )
+
+        return activation_results
+
+    @staticmethod
+    def backward(ctx, output_grads):
+        raise NotImplementedError
+        return output_grads, None, None
+
+
+class Activation1d(nn.Module):
+    def __init__(
+        self,
+        activation,
+        up_ratio: int = 2,
+        down_ratio: int = 2,
+        up_kernel_size: int = 12,
+        down_kernel_size: int = 12,
+        fused: bool = True,
+    ):
+        super().__init__()
+        self.up_ratio = up_ratio
+        self.down_ratio = down_ratio
+        self.act = activation
+        self.upsample = UpSample1d(up_ratio, up_kernel_size)
+        self.downsample = DownSample1d(down_ratio, down_kernel_size)
+
+        self.fused = fused  # Whether to use fused CUDA kernel or not
+
+    def forward(self, x):
+        if not self.fused:
+            x = self.upsample(x)
+            x = self.act(x)
+            x = self.downsample(x)
+            return x
+        else:
+            if self.act.__class__.__name__ == "Snake":
+                beta = self.act.alpha.data  # Snake uses same params for alpha and beta
+            else:
+                beta = (
+                    self.act.beta.data
+                )  # Snakebeta uses different params for alpha and beta
+            alpha = self.act.alpha.data
+            if (
+                not self.act.alpha_logscale
+            ):  # Exp baked into cuda kernel, cancel it out with a log
+                alpha = torch.log(alpha)
+                beta = torch.log(beta)
+
+            x = FusedAntiAliasActivation.apply(
+                x, self.upsample.filter, self.downsample.lowpass.filter, alpha, beta
+            )
+            return x

mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation.cpp

@@ -0,0 +1,23 @@
+/* coding=utf-8
+ * Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ #include <torch/extension.h>
+
+extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta);
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+    m.def("forward", &fwd_cuda, "Anti-Alias Activation forward (CUDA)");
+}

mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/anti_alias_activation_cuda.cu

@@ -0,0 +1,246 @@
+/* coding=utf-8
+ * Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <ATen/ATen.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <cuda_fp16.h>
+#include <cuda_profiler_api.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/extension.h>
+#include "type_shim.h"
+#include <assert.h>
+#include <cfloat>
+#include <limits>
+#include <stdint.h>
+#include <c10/macros/Macros.h>
+
+namespace
+{
+    // Hard-coded hyperparameters
+    // WARP_SIZE and WARP_BATCH must match the return values batches_per_warp and
+    constexpr int ELEMENTS_PER_LDG_STG = 1; //(WARP_ITERATIONS < 4) ? 1 : 4;
+    constexpr int BUFFER_SIZE = 32;
+    constexpr int FILTER_SIZE = 12;
+    constexpr int HALF_FILTER_SIZE = 6;
+    constexpr int UPSAMPLE_REPLICATION_PAD = 5; // 5 on each side, matching torch impl
+    constexpr int DOWNSAMPLE_REPLICATION_PAD_LEFT = 5; // matching torch impl
+    constexpr int DOWNSAMPLE_REPLICATION_PAD_RIGHT = 6; // matching torch impl
+
+    template <typename input_t, typename output_t, typename acc_t>
+    __global__ void anti_alias_activation_forward(
+        output_t *dst,
+        const input_t *src,
+        const input_t *up_ftr,
+        const input_t *down_ftr,
+        const input_t *alpha,
+        const input_t *beta,
+        int batch_size,
+        int channels,
+        int seq_len)
+    {
+        // Up and downsample filters
+        input_t up_filter[FILTER_SIZE];
+        input_t down_filter[FILTER_SIZE];
+
+        // Load data from global memory including extra indices reserved for replication paddings
+        input_t elements[2 * FILTER_SIZE + 2 * BUFFER_SIZE + 2 * UPSAMPLE_REPLICATION_PAD] = {0};
+        input_t intermediates[2 * FILTER_SIZE + 2 * BUFFER_SIZE + DOWNSAMPLE_REPLICATION_PAD_LEFT + DOWNSAMPLE_REPLICATION_PAD_RIGHT] = {0};
+
+        // Output stores downsampled output before writing to dst
+        output_t output[BUFFER_SIZE];
+
+        // blockDim/threadIdx = (128, 1, 1)
+        // gridDim/blockIdx = (seq_blocks, channels, batches)
+        int block_offset = (blockIdx.x * 128 * BUFFER_SIZE + seq_len * (blockIdx.y + gridDim.y * blockIdx.z));
+        int local_offset = threadIdx.x * BUFFER_SIZE;
+        int seq_offset = blockIdx.x * 128 * BUFFER_SIZE + local_offset;
+
+        // intermediate have double the seq_len
+        int intermediate_local_offset = threadIdx.x * BUFFER_SIZE * 2;
+        int intermediate_seq_offset = blockIdx.x * 128 * BUFFER_SIZE * 2 + intermediate_local_offset;
+
+        // Get values needed for replication padding before moving pointer
+        const input_t *right_most_pntr = src + (seq_len * (blockIdx.y + gridDim.y * blockIdx.z));
+        input_t seq_left_most_value = right_most_pntr[0];
+        input_t seq_right_most_value = right_most_pntr[seq_len - 1];
+
+        // Move src and dst pointers
+        src += block_offset + local_offset;
+        dst += block_offset + local_offset;
+
+        // Alpha and beta values for snake activatons. Applies exp by default
+        alpha = alpha + blockIdx.y;
+        input_t alpha_val = expf(alpha[0]);
+        beta = beta + blockIdx.y;
+        input_t beta_val = expf(beta[0]);
+
+        #pragma unroll
+        for (int it = 0; it < FILTER_SIZE; it += 1)
+        {
+            up_filter[it] = up_ftr[it];
+            down_filter[it] = down_ftr[it];
+        }
+
+        // Apply replication padding for upsampling, matching torch impl
+        #pragma unroll
+        for (int it = -HALF_FILTER_SIZE; it < BUFFER_SIZE + HALF_FILTER_SIZE; it += 1)
+        {
+            int element_index = seq_offset + it; // index for element
+            if ((element_index < 0) && (element_index >= -UPSAMPLE_REPLICATION_PAD))
+            {
+                elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_left_most_value;
+            }
+            if ((element_index >= seq_len) && (element_index < seq_len + UPSAMPLE_REPLICATION_PAD))
+            {
+                elements[2 * (HALF_FILTER_SIZE + it)] = 2 * seq_right_most_value;
+            }
+            if ((element_index >= 0) && (element_index < seq_len))
+            {
+                elements[2 * (HALF_FILTER_SIZE + it)] = 2 * src[it];
+            }
+        }
+
+        // Apply upsampling strided convolution and write to intermediates. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT for replication padding of the downsampilng conv later
+        #pragma unroll
+        for (int it = 0; it < (2 * BUFFER_SIZE + 2 * FILTER_SIZE); it += 1)
+        {
+            input_t acc = 0.0;
+            int element_index = intermediate_seq_offset + it; // index for intermediate
+            #pragma unroll
+            for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1)
+            {
+                if ((element_index + f_idx) >= 0)
+                {
+                    acc += up_filter[f_idx] * elements[it + f_idx];
+                }
+            }
+            intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] = acc;
+        }
+
+        // Apply activation function. It reserves DOWNSAMPLE_REPLICATION_PAD_LEFT and DOWNSAMPLE_REPLICATION_PAD_RIGHT for replication padding of the downsampilng conv later
+        double no_div_by_zero = 0.000000001;
+        #pragma unroll
+        for (int it = 0; it < 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it += 1)
+        {
+            intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] += (1.0 / (beta_val + no_div_by_zero)) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val) * sinf(intermediates[it + DOWNSAMPLE_REPLICATION_PAD_LEFT] * alpha_val);
+        }
+
+        // Apply replication padding before downsampling conv from intermediates
+        #pragma unroll
+        for (int it = 0; it < DOWNSAMPLE_REPLICATION_PAD_LEFT; it += 1)
+        {
+            intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT];
+        }
+        #pragma unroll
+        for (int it = DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE; it < DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE + DOWNSAMPLE_REPLICATION_PAD_RIGHT; it += 1)
+        {
+            intermediates[it] = intermediates[DOWNSAMPLE_REPLICATION_PAD_LEFT + 2 * BUFFER_SIZE + 2 * FILTER_SIZE - 1];
+        }
+
+        // Apply downsample strided convolution (assuming stride=2) from intermediates
+        #pragma unroll
+        for (int it = 0; it < BUFFER_SIZE; it += 1)
+        {
+            input_t acc = 0.0;
+            #pragma unroll
+            for (int f_idx = 0; f_idx < FILTER_SIZE; f_idx += 1)
+            {
+                // Add constant DOWNSAMPLE_REPLICATION_PAD_RIGHT to match torch implementation
+                acc += down_filter[f_idx] * intermediates[it * 2 + f_idx + DOWNSAMPLE_REPLICATION_PAD_RIGHT];
+            }
+            output[it] = acc;
+        }
+
+        // Write output to dst
+        #pragma unroll
+        for (int it = 0;  it < BUFFER_SIZE;  it += ELEMENTS_PER_LDG_STG)
+        {
+            int element_index = seq_offset + it;
+            if (element_index < seq_len)
+            {
+                dst[it] = output[it];
+            }
+        }
+
+    }
+
+    template <typename input_t, typename output_t, typename acc_t>
+    void dispatch_anti_alias_activation_forward(
+        output_t *dst,
+        const input_t *src,
+        const input_t *up_ftr,
+        const input_t *down_ftr,
+        const input_t *alpha,
+        const input_t *beta,
+        int batch_size,
+        int channels,
+        int seq_len)
+    {
+        if (seq_len == 0)
+        {
+            return;
+        }
+        else
+        {
+            // Use 128 threads per block to maximimize gpu utilization
+            constexpr int threads_per_block = 128;
+            constexpr int seq_len_per_block = 4096;
+            int blocks_per_seq_len = (seq_len + seq_len_per_block - 1) / seq_len_per_block;
+            dim3 blocks(blocks_per_seq_len, channels, batch_size);
+            dim3 threads(threads_per_block, 1, 1);
+
+            anti_alias_activation_forward<input_t, output_t, acc_t>
+                <<<blocks, threads, 0, at::cuda::getCurrentCUDAStream()>>>(dst, src, up_ftr, down_ftr, alpha, beta, batch_size, channels, seq_len);
+        }
+    }
+}
+
+extern "C" torch::Tensor fwd_cuda(torch::Tensor const &input, torch::Tensor const &up_filter, torch::Tensor const &down_filter, torch::Tensor const &alpha, torch::Tensor const &beta)
+{
+    // Input is a 3d tensor with dimensions [batches, channels, seq_len]
+    const int batches = input.size(0);
+    const int channels = input.size(1);
+    const int seq_len = input.size(2);
+
+    // Output
+    auto act_options = input.options().requires_grad(false);
+
+    torch::Tensor anti_alias_activation_results =
+        torch::empty({batches, channels, seq_len}, act_options);
+
+    void *input_ptr = static_cast<void *>(input.data_ptr());
+    void *up_filter_ptr = static_cast<void *>(up_filter.data_ptr());
+    void *down_filter_ptr = static_cast<void *>(down_filter.data_ptr());
+    void *alpha_ptr = static_cast<void *>(alpha.data_ptr());
+    void *beta_ptr = static_cast<void *>(beta.data_ptr());
+    void *anti_alias_activation_results_ptr = static_cast<void *>(anti_alias_activation_results.data_ptr());
+
+    DISPATCH_FLOAT_HALF_AND_BFLOAT(
+        input.scalar_type(),
+        "dispatch anti alias activation_forward",
+        dispatch_anti_alias_activation_forward<scalar_t, scalar_t, float>(
+            reinterpret_cast<scalar_t *>(anti_alias_activation_results_ptr),
+            reinterpret_cast<const scalar_t *>(input_ptr),
+            reinterpret_cast<const scalar_t *>(up_filter_ptr),
+            reinterpret_cast<const scalar_t *>(down_filter_ptr),
+            reinterpret_cast<const scalar_t *>(alpha_ptr),
+            reinterpret_cast<const scalar_t *>(beta_ptr),
+            batches,
+            channels,
+            seq_len););
+    return anti_alias_activation_results;
+}

mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/compat.h

@@ -0,0 +1,29 @@
+/* coding=utf-8
+ * Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/*This code is copied fron NVIDIA apex:
+ *     https://github.com/NVIDIA/apex
+ *     with minor changes. */
+
+#ifndef TORCH_CHECK
+#define TORCH_CHECK AT_CHECK
+#endif
+
+#ifdef VERSION_GE_1_3
+#define DATA_PTR data_ptr
+#else
+#define DATA_PTR data
+#endif

mmaudio/ext/bigvgan_v2/alias_free_activation/cuda/load.py

@@ -0,0 +1,86 @@
+# Copyright (c) 2024 NVIDIA CORPORATION.
+#   Licensed under the MIT license.
+
+import os
+import pathlib
+import subprocess
+
+from torch.utils import cpp_extension
+
+"""
+Setting this param to a list has a problem of generating different compilation commands (with diferent order of architectures) and leading to recompilation of fused kernels. 
+Set it to empty stringo avoid recompilation and assign arch flags explicity in extra_cuda_cflags below
+"""
+os.environ["TORCH_CUDA_ARCH_LIST"] = ""
+
+
+def load():
+    # Check if cuda 11 is installed for compute capability 8.0
+    cc_flag = []
+    _, bare_metal_major, _ = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
+    if int(bare_metal_major) >= 11:
+        cc_flag.append("-gencode")
+        cc_flag.append("arch=compute_80,code=sm_80")
+
+    # Build path
+    srcpath = pathlib.Path(__file__).parent.absolute()
+    buildpath = srcpath / "build"
+    _create_build_dir(buildpath)
+
+    # Helper function to build the kernels.
+    def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
+        return cpp_extension.load(
+            name=name,
+            sources=sources,
+            build_directory=buildpath,
+            extra_cflags=[
+                "-O3",
+            ],
+            extra_cuda_cflags=[
+                "-O3",
+                "-gencode",
+                "arch=compute_70,code=sm_70",
+                "--use_fast_math",
+            ]
+            + extra_cuda_flags
+            + cc_flag,
+            verbose=True,
+        )
+
+    extra_cuda_flags = [
+        "-U__CUDA_NO_HALF_OPERATORS__",
+        "-U__CUDA_NO_HALF_CONVERSIONS__",
+        "--expt-relaxed-constexpr",
+        "--expt-extended-lambda",
+    ]
+
+    sources = [
+        srcpath / "anti_alias_activation.cpp",
+        srcpath / "anti_alias_activation_cuda.cu",
+    ]
+    anti_alias_activation_cuda = _cpp_extention_load_helper(
+        "anti_alias_activation_cuda", sources, extra_cuda_flags
+    )
+
+    return anti_alias_activation_cuda
+
+
+def _get_cuda_bare_metal_version(cuda_dir):
+    raw_output = subprocess.check_output(
+        [cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True
+    )
+    output = raw_output.split()
+    release_idx = output.index("release") + 1
+    release = output[release_idx].split(".")
+    bare_metal_major = release[0]
+    bare_metal_minor = release[1][0]
+
+    return raw_output, bare_metal_major, bare_metal_minor
+
+
+def _create_build_dir(buildpath):
+    try:
+        os.mkdir(buildpath)
+    except OSError:
+        if not os.path.isdir(buildpath):
+            print(f"Creation of the build directory {buildpath} failed")