train_uncond_denoiser.py

#!/usr/bin/env python3

from prefigure.prefigure import get_all_args, push_wandb_config
from contextlib import contextmanager
from copy import deepcopy
import math
from pathlib import Path

import sys
import torch
from torch import optim, nn
from torch.nn import functional as F
from torch.utils import data
from tqdm import trange
import pytorch_lightning as pl
from pytorch_lightning.utilities.distributed import rank_zero_only
from einops import rearrange

from diffusion.pqmf import CachedPQMF as PQMF
import torchaudio

import auraloss

import wandb

from aeiou.datasets import AudioDataset

from dataset.dataset import SampleDataset

from decoders.diffusion_decoder import DiffusionAttnUnet1D
from diffusion.model import ema_update
from viz.viz import embeddings_table, pca_point_cloud, audio_spectrogram_image, tokens_spectrogram_image


# Define the noise schedule and sampling loop
def get_alphas_sigmas(t):
    """Returns the scaling factors for the clean image (alpha) and for the
    noise (sigma), given a timestep."""
    return torch.cos(t * math.pi / 2), torch.sin(t * math.pi / 2)

def get_crash_schedule(t):
    sigma = torch.sin(t * math.pi / 2) ** 2
    alpha = (1 - sigma ** 2) ** 0.5
    return alpha_sigma_to_t(alpha, sigma)

def alpha_sigma_to_t(alpha, sigma):
    """Returns a timestep, given the scaling factors for the clean image and for
    the noise."""
    return torch.atan2(sigma, alpha) / math.pi * 2

@torch.no_grad()
def sample(model, x, steps, eta):
    """Draws samples from a model given starting noise."""
    ts = x.new_ones([x.shape[0]])

    # Create the noise schedule
    t = torch.linspace(1, 0, steps + 1)[:-1]

    t = get_crash_schedule(t)

    alphas, sigmas = get_alphas_sigmas(t)

    # The sampling loop
    for i in trange(steps):

        # Get the model output (v, the predicted velocity)
        with torch.cuda.amp.autocast():
            v = model(x, ts * t[i]).float()

        # Predict the noise and the denoised image
        pred = x * alphas[i] - v * sigmas[i]
        eps = x * sigmas[i] + v * alphas[i]

        # If we are not on the last timestep, compute the noisy image for the
        # next timestep.
        if i < steps - 1:
            # If eta > 0, adjust the scaling factor for the predicted noise
            # downward according to the amount of additional noise to add
            ddim_sigma = eta * (sigmas[i + 1]**2 / sigmas[i]**2).sqrt() * \
                (1 - alphas[i]**2 / alphas[i + 1]**2).sqrt()
            adjusted_sigma = (sigmas[i + 1]**2 - ddim_sigma**2).sqrt()

            # Recombine the predicted noise and predicted denoised image in the
            # correct proportions for the next step
            x = pred * alphas[i + 1] + eps * adjusted_sigma

            # Add the correct amount of fresh noise
            if eta:
                x += torch.randn_like(x) * ddim_sigma

    # If we are on the last timestep, output the denoised image
    return pred


@torch.no_grad()
def sample_pred(model, x, steps, **extra_args):
    """Draws samples from a model given starting noise."""
    ts = x.new_ones([x.shape[0]])

    # Create the noise schedule
    t = torch.linspace(1, 0, steps + 1)[:-1]

    alphas, sigmas = get_alphas_sigmas(t)

    # The sampling loop
    for i in trange(steps):

        # Get the model output (pred, the predicted denoised audio)
        with torch.cuda.amp.autocast():
            pred = model(x, ts * t[i], **extra_args).float()

        # Predict the noise and the denoised audio
        #pred = x * alphas[i] - v * sigmas[i]
        eps = x - pred

        # If we are not on the last timestep, compute the noisy audio for the
        # next timestep.
        if i < steps - 1:
            
            # Recombine the predicted noise and predicted denoised audio in the
            # correct proportions for the next step
            x = pred * alphas[i + 1] + eps * sigmas[i + 1]

    # If we are on the last timestep, output the denoised audio
    return pred

class DiffusionUncond(pl.LightningModule):
    def __init__(self, global_args):
        super().__init__()
        #self.diffusion = DiffusionAttnUnet1D(io_channels=2, pqmf_bands=global_args.pqmf_bands, n_attn_layers=4)

        self.diffusion = DiffusionAttnUnet1D(
            io_channels=2, 
            n_attn_layers=0, 
            depth=6,
            c_mults=[128, 128, 128, 256, 512, 512],
            learned_resample=True,
            strides=[2, 2, 4, 4, 4, 8],
            kernel_size=3
        )

        scales = [2048, 1024, 512, 256, 128]
        hop_sizes = []
        win_lengths = []
        overlap = 0.75
        for s in scales:
            hop_sizes.append(int(s * (1 - overlap)))
            win_lengths.append(s)
        self.sdstft = auraloss.freq.SumAndDifferenceSTFTLoss(fft_sizes=scales, hop_sizes=hop_sizes, win_lengths=win_lengths, sample_rate=global_args.sample_rate, perceptual_weighting=True)

        self.diffusion_ema = deepcopy(self.diffusion)
        self.rng = torch.quasirandom.SobolEngine(1, scramble=True, seed=global_args.seed)
        self.ema_decay = global_args.ema_decay
        
    def configure_optimizers(self):
        return optim.Adam([*self.diffusion.parameters()], lr=4e-5)
  
    def training_step(self, batch, batch_idx):
        reals = batch
        
       # Draw uniformly distributed continuous timesteps
        t = self.rng.draw(reals.shape[0])[:, 0].to(self.device)

        # Calculate the noise schedule parameters for those timesteps
        alphas, sigmas = get_alphas_sigmas(t)

        # Combine the ground truth images and the noise
        alphas = alphas[:, None, None]
        sigmas = sigmas[:, None, None]
        noise = torch.randn_like(reals)
        noised_reals = reals * alphas + noise * sigmas
        targets = reals # noise * alphas - reals * sigmas
        target_v = noise * alphas - reals * sigmas

        with torch.cuda.amp.autocast():
            denoised = self.diffusion(noised_reals, t)
            v = noise * alphas - denoised * sigmas
            mse_loss = F.mse_loss(v, target_v)

            stft_loss = self.sdstft(denoised, targets)
            loss = stft_loss + mse_loss

        log_dict = {
            'train/loss': loss.detach(),
            'train/stft_loss': stft_loss.detach(),
            'train/mse_loss': mse_loss.detach(),
        }

        self.log_dict(log_dict, prog_bar=True, on_step=True)
        return loss

    def on_before_zero_grad(self, *args, **kwargs):
        decay = 0.95 if self.current_epoch < 25 else self.ema_decay
        ema_update(self.diffusion, self.diffusion_ema, decay)

class ExceptionCallback(pl.Callback):
    def on_exception(self, trainer, module, err):
        print(f'{type(err).__name__}: {err}', file=sys.stderr)


class DemoCallback(pl.Callback):
    def __init__(self, global_args):
        super().__init__()
        self.demo_every = global_args.demo_every
        self.num_demos = global_args.num_demos
        self.demo_samples = global_args.sample_size
        self.demo_steps = global_args.demo_steps
        self.sample_rate = global_args.sample_rate
        

    @rank_zero_only
    @torch.no_grad()
    #def on_train_epoch_end(self, trainer, module):
    def on_train_batch_end(self, trainer, module, outputs, batch, batch_idx):        
        last_demo_step = -1
        if (trainer.global_step - 1) % self.demo_every != 0 or last_demo_step == trainer.global_step:
        #if trainer.current_epoch % self.demo_every != 0:
            return
        
        last_demo_step = trainer.global_step
        
        noise = torch.randn([self.num_demos, 2, self.demo_samples]).to(module.device)

        try:
            fakes = sample_pred(module.diffusion_ema, noise, self.demo_steps)

            # Put the demos together
            fakes = rearrange(fakes, 'b d n -> d (b n)')

            log_dict = {}
            
            filename = f'demo_{trainer.global_step:08}.wav'
            fakes = fakes.clamp(-1, 1).mul(32767).to(torch.int16).cpu()
            torchaudio.save(filename, fakes, self.sample_rate)


            log_dict[f'demo'] = wandb.Audio(filename,
                                                sample_rate=self.sample_rate,
                                                caption=f'Reconstructed')
        

            log_dict[f'demo_melspec_left'] = wandb.Image(audio_spectrogram_image(fakes))


            trainer.logger.experiment.log(log_dict, step=trainer.global_step)
        except Exception as e:
            print(f'{type(e).__name__}: {e}')

def main():

    args = get_all_args()

    args.latent_dim = 0

    #args.random_crop = False

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print('Using device:', device)
    torch.manual_seed(args.seed)

    train_set = AudioDataset(
        [args.training_dir],
        sample_rate=args.sample_rate,
        sample_size=args.sample_size,
        random_crop=args.random_crop,
        augs='Stereo(), PhaseFlipper()'
    )
    
    #train_set = SampleDataset([args.training_dir], args, keywords=["kick", "snare", "clap", "snap", "hat", "cymbal", "crash", "ride"])

    train_dl = data.DataLoader(train_set, args.batch_size, shuffle=True,
                               num_workers=args.num_workers, persistent_workers=True, pin_memory=True, drop_last=True)
    wandb_logger = pl.loggers.WandbLogger(project=args.name)
    exc_callback = ExceptionCallback()
    ckpt_callback = pl.callbacks.ModelCheckpoint(every_n_train_steps=args.checkpoint_every, save_top_k=-1)
    demo_callback = DemoCallback(args)
    diffusion_model = DiffusionUncond(args)
    wandb_logger.watch(diffusion_model)
    push_wandb_config(wandb_logger, args)

    diffusion_trainer = pl.Trainer(
        devices=args.num_gpus,
        accelerator="gpu",
        num_nodes = args.num_nodes,
        strategy='ddp',
        precision=16,
        accumulate_grad_batches=args.accum_batches, 
        callbacks=[ckpt_callback, demo_callback, exc_callback],
        logger=wandb_logger,
        log_every_n_steps=1,
        max_epochs=10000000,
    )

    diffusion_trainer.fit(diffusion_model, train_dl, ckpt_path=args.ckpt_path)

if __name__ == '__main__':
    main()