import torch import torch.nn as nn import torch.optim as optim import torch.nn.functional as F import torchaudio import tqdm import argparse import math from torch.utils.data import random_split from torch.utils.data import DataLoader import AudioUtils from data import AudioDataset from generator import SISUGenerator from discriminator import SISUDiscriminator def perceptual_loss(y_true, y_pred): return torch.mean((y_true - y_pred) ** 2) def discriminator_train(high_quality, low_quality, real_labels, fake_labels): optimizer_d.zero_grad() # Forward pass for real samples discriminator_decision_from_real = discriminator(high_quality[0]) d_loss_real = criterion_d(discriminator_decision_from_real, real_labels) # Forward pass for fake samples (from generator output) generator_output = generator(low_quality[0]) discriminator_decision_from_fake = discriminator(generator_output.detach()) d_loss_fake = criterion_d(discriminator_decision_from_fake, fake_labels) # Combine real and fake losses d_loss = (d_loss_real + d_loss_fake) / 2.0 # Backward pass and optimization d_loss.backward() nn.utils.clip_grad_norm_(discriminator.parameters(), max_norm=1.0) # Gradient Clipping optimizer_d.step() return d_loss def generator_train(low_quality, real_labels): optimizer_g.zero_grad() # Forward pass for fake samples (from generator output) generator_output = generator(low_quality[0]) discriminator_decision = discriminator(generator_output) g_loss = criterion_g(discriminator_decision, real_labels) g_loss.backward() optimizer_g.step() return generator_output def first(objects): if len(objects) >= 1: return objects[0] return objects # Init script argument parser parser = argparse.ArgumentParser(description="Training script") parser.add_argument("--generator", type=str, default=None, help="Path to the generator model file") parser.add_argument("--discriminator", type=str, default=None, help="Path to the discriminator model file") args = parser.parse_args() # Check for CUDA availability device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"Using device: {device}") # Initialize dataset and dataloader dataset_dir = './dataset/good' dataset = AudioDataset(dataset_dir) # ========= SINGLE ========= train_data_loader = DataLoader(dataset, batch_size=16, shuffle=True) # Initialize models and move them to device generator = SISUGenerator() discriminator = SISUDiscriminator() if args.generator is not None: generator.load_state_dict(torch.load(args.generator, weights_only=True)) if args.discriminator is not None: discriminator.load_state_dict(torch.load(args.discriminator, weights_only=True)) generator = generator.to(device) discriminator = discriminator.to(device) # Loss criterion_g = nn.MSELoss() criterion_d = nn.BCELoss() # Optimizers optimizer_g = optim.Adam(generator.parameters(), lr=0.0001, betas=(0.5, 0.999)) optimizer_d = optim.Adam(discriminator.parameters(), lr=0.0001, betas=(0.5, 0.999)) # Scheduler scheduler_g = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_g, mode='min', factor=0.5, patience=5) scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min', factor=0.5, patience=5) def start_training(): generator_epochs = 5000 for generator_epoch in range(generator_epochs): low_quality_audio = (torch.empty((1)), 1) high_quality_audio = (torch.empty((1)), 1) ai_enhanced_audio = (torch.empty((1)), 1) times_correct = 0 # ========= TRAINING ========= for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Epoch {generator_epoch+1}/{generator_epochs}"): # for high_quality_clip, low_quality_clip in train_data_loader: high_quality_sample = (high_quality_clip[0].to(device), high_quality_clip[1]) low_quality_sample = (low_quality_clip[0].to(device), low_quality_clip[1]) # ========= LABELS ========= batch_size = high_quality_clip[0].size(0) real_labels = torch.ones(batch_size, 1).to(device) fake_labels = torch.zeros(batch_size, 1).to(device) # ========= DISCRIMINATOR ========= discriminator.train() discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels) # ========= GENERATOR ========= generator.train() generator_output = generator_train(low_quality_sample, real_labels) # ========= SAVE LATEST AUDIO ========= high_quality_audio = (first(high_quality_clip[0]), high_quality_clip[1][0]) low_quality_audio = (first(low_quality_clip[0]), low_quality_clip[1][0]) ai_enhanced_audio = (first(generator_output[0]), high_quality_clip[1][0]) print(high_quality_audio) print(f"Saved epoch {generator_epoch}!") torchaudio.save(f"./output/epoch-{generator_epoch}-audio-crap.wav", low_quality_audio[0][0].cpu(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again. torchaudio.save(f"./output/epoch-{generator_epoch}-audio-ai.wav", ai_enhanced_audio[0][0].cpu(), ai_enhanced_audio[1]) torchaudio.save(f"./output/epoch-{generator_epoch}-audio-orig.wav", high_quality_audio[0][0].cpu(), high_quality_audio[1]) #metric = snr(high_quality_audio[0].to(device), ai_enhanced_audio[0]) #print(f"Generator metric {metric}!") #scheduler_g.step(metric) if generator_epoch % 10 == 0: print(f"Saved epoch {generator_epoch}!") torchaudio.save(f"./output/epoch-{generator_epoch}-audio-crap.wav", low_quality_audio[0][0].cpu(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again. torchaudio.save(f"./output/epoch-{generator_epoch}-audio-ai.wav", ai_enhanced_audio[0][0].cpu(), ai_enhanced_audio[1]) torchaudio.save(f"./output/epoch-{generator_epoch}-audio-orig.wav", high_quality_audio[0][0].cpu(), high_quality_audio[1]) torch.save(discriminator.state_dict(), f"models/current-epoch-discriminator.pt") torch.save(generator.state_dict(), f"models/current-epoch-generator.pt") torch.save(discriminator.state_dict(), "models/epoch-5000-discriminator.pt") torch.save(generator.state_dict(), "models/epoch-5000-generator.pt") print("Training complete!") start_training()