models/frn.py

import os

import librosa
import pytorch_lightning as pl
import soundfile as sf
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchmetrics.audio.pesq import PerceptualEvaluationSpeechQuality as PESQ
from torchmetrics.audio.stoi import ShortTimeObjectiveIntelligibility as STOI

from PLCMOS.plc_mos import PLCMOSEstimator
from config import CONFIG
from loss import Loss
from models.blocks import Encoder, Predictor
from utils.utils import visualize, LSD

plcmos = PLCMOSEstimator()


class PLCModel(pl.LightningModule):
    def __init__(self, train_dataset=None, val_dataset=None, window_size=960, enc_layers=4, enc_in_dim=384, enc_dim=768,
                 pred_dim=512, pred_layers=1, pred_ckpt_path='lightning_logs/predictor/checkpoints/predictor.ckpt'):
        super(PLCModel, self).__init__()
        self.window_size = window_size
        self.hop_size = window_size // 2
        self.learning_rate = CONFIG.TRAIN.lr
        self.hparams.batch_size = CONFIG.TRAIN.batch_size

        self.enc_layers = enc_layers
        self.enc_in_dim = enc_in_dim
        self.enc_dim = enc_dim
        self.pred_dim = pred_dim
        self.pred_layers = pred_layers
        self.train_dataset = train_dataset
        self.val_dataset = val_dataset
        self.stoi = STOI(48000)
        self.pesq = PESQ(16000, 'wb')

        if pred_ckpt_path is not None:
            self.predictor = Predictor.load_from_checkpoint(pred_ckpt_path)
        else:
            self.predictor = Predictor(window_size=self.window_size, lstm_dim=self.pred_dim,
                                       lstm_layers=self.pred_layers)
        self.joiner = nn.Sequential(
            nn.Conv2d(3, 48, kernel_size=(9, 1), stride=1, padding=(4, 0), padding_mode='reflect',
                      groups=3),
            nn.LeakyReLU(0.2),
            nn.Conv2d(48, 2, kernel_size=1, stride=1, padding=0, groups=2),
        )

        self.encoder = Encoder(in_dim=self.window_size, dim=self.enc_in_dim, depth=self.enc_layers,
                               mlp_dim=self.enc_dim)

        self.loss = Loss()
        self.window = torch.sqrt(torch.hann_window(self.window_size))
        self.save_hyperparameters('window_size', 'enc_layers', 'enc_in_dim', 'enc_dim', 'pred_dim', 'pred_layers')

    def forward(self, x):
        """
        Input: real-imaginary; shape (B, F, T, 2); F = hop_size + 1
        Output: real-imaginary
        """

        B, C, F, T = x.shape

        x = x.permute(3, 0, 1, 2).unsqueeze(-1)
        prev_mag = torch.zeros((B, 1, F, 1), device=x.device)
        predictor_state = torch.zeros((2, self.predictor.lstm_layers, B, self.predictor.lstm_dim), device=x.device)
        mlp_state = torch.zeros((self.encoder.depth, 2, 1, B, self.encoder.dim), device=x.device)
        result = []
        for step in x:
            feat, mlp_state = self.encoder(step, mlp_state)
            prev_mag, predictor_state = self.predictor(prev_mag, predictor_state)
            feat = torch.cat((feat, prev_mag), 1)
            feat = self.joiner(feat)
            feat = feat + step
            result.append(feat)
            prev_mag = torch.linalg.norm(feat, dim=1, ord=1, keepdims=True)  # compute magnitude
        output = torch.cat(result, -1)
        return output

    def forward_onnx(self, x, prev_mag, predictor_state=None, mlp_state=None):
        prev_mag, predictor_state = self.predictor(prev_mag, predictor_state)
        feat, mlp_state = self.encoder(x, mlp_state)

        feat = torch.cat((feat, prev_mag), 1)
        feat = self.joiner(feat)
        prev_mag = torch.linalg.norm(feat, dim=1, ord=1, keepdims=True)
        feat = feat + x
        return feat, prev_mag, predictor_state, mlp_state

    def train_dataloader(self):
        return DataLoader(self.train_dataset, shuffle=False, batch_size=self.hparams.batch_size,
                          num_workers=CONFIG.TRAIN.workers, persistent_workers=True)

    def val_dataloader(self):
        return DataLoader(self.val_dataset, shuffle=False, batch_size=self.hparams.batch_size,
                          num_workers=CONFIG.TRAIN.workers, persistent_workers=True)

    def training_step(self, batch, batch_idx):
        x_in, y = batch
        f_0 = x_in[:, :, 0:1, :]
        x = x_in[:, :, 1:, :]

        x = self(x)
        x = torch.cat([f_0, x], dim=2)

        loss = self.loss(x, y)
        self.log('train_loss', loss, logger=True)
        return loss

    def validation_step(self, val_batch, batch_idx):
        x, y = val_batch
        f_0 = x[:, :, 0:1, :]
        x_in = x[:, :, 1:, :]

        pred = self(x_in)
        pred = torch.cat([f_0, pred], dim=2)

        loss = self.loss(pred, y)
        self.window = self.window.to(pred.device)
        pred = torch.view_as_complex(pred.permute(0, 2, 3, 1).contiguous())
        pred = torch.istft(pred, self.window_size, self.hop_size, window=self.window)
        y = torch.view_as_complex(y.permute(0, 2, 3, 1).contiguous())
        y = torch.istft(y, self.window_size, self.hop_size, window=self.window)

        self.log('val_loss', loss, on_step=False, on_epoch=True, logger=True, prog_bar=True, sync_dist=True)

        if batch_idx == 0:
            i = torch.randint(0, x.shape[0], (1,)).item()
            x = torch.view_as_complex(x.permute(0, 2, 3, 1).contiguous())
            x = torch.istft(x[i], self.window_size, self.hop_size, window=self.window)

            self.trainer.logger.log_spectrogram(y[i], x, pred[i], self.current_epoch)
            self.trainer.logger.log_audio(y[i], x, pred[i], self.current_epoch)

    def test_step(self, test_batch, batch_idx):
        inp, tar, inp_wav, tar_wav = test_batch
        inp_wav = inp_wav.squeeze()
        tar_wav = tar_wav.squeeze()
        f_0 = inp[:, :, 0:1, :]
        x = inp[:, :, 1:, :]
        pred = self(x)
        pred = torch.cat([f_0, pred], dim=2)
        pred = torch.istft(pred.squeeze(0).permute(1, 2, 0), self.window_size, self.hop_size,
                           window=self.window.to(pred.device))
        stoi = self.stoi(pred, tar_wav)

        tar_wav = tar_wav.cpu().numpy()
        inp_wav = inp_wav.cpu().numpy()
        pred = pred.detach().cpu().numpy()
        lsd, _ = LSD(tar_wav, pred)

        if batch_idx in [3, 5, 7]:
            sample_path = os.path.join(CONFIG.LOG.sample_path)
            path = os.path.join(sample_path, 'sample_' + str(batch_idx))
            visualize(tar_wav, inp_wav, pred, path)
            sf.write(os.path.join(path, 'enhanced_output.wav'), pred, samplerate=CONFIG.DATA.sr, subtype='PCM_16')
            sf.write(os.path.join(path, 'lossy_input.wav'), inp_wav, samplerate=CONFIG.DATA.sr, subtype='PCM_16')
            sf.write(os.path.join(path, 'target.wav'), tar_wav, samplerate=CONFIG.DATA.sr, subtype='PCM_16')
        if CONFIG.DATA.sr != 16000:
            pred = librosa.resample(pred, orig_sr=48000, target_sr=16000)
            tar_wav = librosa.resample(tar_wav, orig_sr=48000, target_sr=16000, res_type='kaiser_fast')
        ret = plcmos.run(pred, tar_wav)
        pesq = self.pesq(torch.tensor(pred), torch.tensor(tar_wav))
        metrics = {
            "Intrusive": ret[0],
            "Non-intrusive": ret[1],
            'LSD': lsd,
            'STOI': stoi,
            'PESQ': pesq,
        }
        self.log_dict(metrics)
        return metrics

    def predict_step(self, batch, batch_idx: int, dataloader_idx: int = 0):
        f_0 = batch[:, :, 0:1, :]
        x = batch[:, :, 1:, :]
        pred = self(x)
        pred = torch.cat([f_0, pred], dim=2)
        pred = torch.istft(pred.squeeze(0).permute(1, 2, 0), self.window_size, self.hop_size,
                           window=self.window.to(pred.device))
        return pred

    def configure_optimizers(self):
        optimizer = torch.optim.Adam(self.parameters(), lr=self.learning_rate)
        lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=CONFIG.TRAIN.patience,
                                                                  factor=CONFIG.TRAIN.factor, verbose=True)

        scheduler = {
            'scheduler': lr_scheduler,
            'reduce_on_plateau': True,
            'monitor': 'val_loss'
        }
        return [optimizer], [scheduler]


class OnnxWrapper(pl.LightningModule):
    def __init__(self, model, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.model = model
        batch_size = 1
        pred_states = torch.zeros((2, 1, batch_size, model.predictor.lstm_dim))
        mlp_states = torch.zeros((model.encoder.depth, 2, 1, batch_size, model.encoder.dim))
        mag = torch.zeros((batch_size, 1, model.hop_size, 1))
        x = torch.randn(batch_size, model.hop_size + 1, 2)
        self.sample = (x, mag, pred_states, mlp_states)
        self.input_names = ['input', 'mag_in_cached_', 'pred_state_in_cached_', 'mlp_state_in_cached_']
        self.output_names = ['output', 'mag_out_cached_', 'pred_state_out_cached_', 'mlp_state_out_cached_']

    def forward(self, x, prev_mag, predictor_state=None, mlp_state=None):
        x = x.permute(0, 2, 1).unsqueeze(-1)
        f_0 = x[:, :, 0:1, :]
        x = x[:, :, 1:, :]

        output, prev_mag, predictor_state, mlp_state = self.model.forward_onnx(x, prev_mag, predictor_state, mlp_state)
        output = torch.cat([f_0, output], dim=2)
        output = output.squeeze(-1).permute(0, 2, 1)
        return output, prev_mag, predictor_state, mlp_state