191 lines
7.0 KiB
Python
191 lines
7.0 KiB
Python
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
|
|
|
|
import os
|
|
|
|
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
|
|
|
|
import torchaudio.transforms as T
|
|
|
|
# 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")
|
|
parser.add_argument("--device", type=str, default="cpu", help="Select device")
|
|
parser.add_argument("--epoch", type=int, default=0, help="Current epoch for model versioning")
|
|
parser.add_argument("--verbose", action="store_true", help="Increase output verbosity")
|
|
|
|
args = parser.parse_args()
|
|
|
|
device = torch.device(args.device if torch.cuda.is_available() else "cpu")
|
|
print(f"Using device: {device}")
|
|
|
|
mfcc_transform = T.MFCC(
|
|
sample_rate=44100,
|
|
n_mfcc=20,
|
|
melkwargs={'n_fft': 2048, 'hop_length': 256}
|
|
).to(device)
|
|
|
|
def gpu_mfcc_loss(y_true, y_pred):
|
|
mfccs_true = mfcc_transform(y_true)
|
|
mfccs_pred = mfcc_transform(y_pred)
|
|
min_len = min(mfccs_true.shape[2], mfccs_pred.shape[2])
|
|
mfccs_true = mfccs_true[:, :, :min_len]
|
|
mfccs_pred = mfccs_pred[:, :, :min_len]
|
|
loss = torch.mean((mfccs_true - mfccs_pred)**2)
|
|
return loss
|
|
|
|
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, high_quality, real_labels):
|
|
optimizer_g.zero_grad()
|
|
|
|
# Forward pass for fake samples (from generator output)
|
|
generator_output = generator(low_quality[0])
|
|
|
|
mfcc_l = gpu_mfcc_loss(high_quality[0], generator_output)
|
|
|
|
discriminator_decision = discriminator(generator_output)
|
|
adversarial_loss = criterion_g(discriminator_decision, real_labels)
|
|
|
|
combined_loss = adversarial_loss + 0.5 * mfcc_l
|
|
|
|
combined_loss.backward()
|
|
optimizer_g.step()
|
|
|
|
return (generator_output, combined_loss, adversarial_loss, mfcc_l)
|
|
|
|
debug = args.verbose
|
|
|
|
# Initialize dataset and dataloader
|
|
dataset_dir = './dataset/good'
|
|
dataset = AudioDataset(dataset_dir, device)
|
|
|
|
# ========= SINGLE =========
|
|
|
|
train_data_loader = DataLoader(dataset, batch_size=16, shuffle=True)
|
|
|
|
# Initialize models and move them to device
|
|
generator = SISUGenerator()
|
|
discriminator = SISUDiscriminator()
|
|
|
|
epoch: int = args.epoch
|
|
|
|
generator = generator.to(device)
|
|
discriminator = discriminator.to(device)
|
|
|
|
if args.generator is not None:
|
|
generator.load_state_dict(torch.load(args.generator, map_location=device, weights_only=True))
|
|
if args.discriminator is not None:
|
|
discriminator.load_state_dict(torch.load(args.discriminator, map_location=device, weights_only=True))
|
|
|
|
# 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)
|
|
|
|
models_dir = "models"
|
|
os.makedirs(models_dir, exist_ok=True)
|
|
|
|
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"Training epoch {generator_epoch+1}/{generator_epochs}, Current epoch {epoch+1}"):
|
|
# for high_quality_clip, low_quality_clip in train_data_loader:
|
|
high_quality_sample = (high_quality_clip[0], high_quality_clip[1])
|
|
low_quality_sample = (low_quality_clip[0], 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()
|
|
d_loss = discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
|
|
|
|
# ========= GENERATOR =========
|
|
generator.train()
|
|
generator_output, combined_loss, adversarial_loss, mfcc_l = generator_train(low_quality_sample, high_quality_sample, real_labels)
|
|
|
|
if debug:
|
|
print(d_loss, combined_loss, adversarial_loss, mfcc_l)
|
|
scheduler_d.step(d_loss)
|
|
#scheduler_g.step(combined_loss)
|
|
|
|
# ========= SAVE LATEST AUDIO =========
|
|
high_quality_audio = (high_quality_clip[0][0], high_quality_clip[1][0])
|
|
low_quality_audio = (low_quality_clip[0][0], low_quality_clip[1][0])
|
|
ai_enhanced_audio = (generator_output[0], high_quality_clip[1][0])
|
|
|
|
new_epoch = generator_epoch+epoch
|
|
|
|
if generator_epoch % 10 == 0:
|
|
print(f"Saved epoch {new_epoch}!")
|
|
torchaudio.save(f"./output/epoch-{new_epoch}-audio-crap.wav", low_quality_audio[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-{new_epoch}-audio-ai.wav", ai_enhanced_audio[0].cpu(), ai_enhanced_audio[1])
|
|
torchaudio.save(f"./output/epoch-{new_epoch}-audio-orig.wav", high_quality_audio[0].cpu(), high_quality_audio[1])
|
|
|
|
if debug:
|
|
print(generator.state_dict().keys())
|
|
print(discriminator.state_dict().keys())
|
|
torch.save(discriminator.state_dict(), f"{models_dir}/discriminator_epoch_{new_epoch}.pt")
|
|
torch.save(generator.state_dict(), f"{models_dir}/generator_epoch_{new_epoch}.pt")
|
|
|
|
torch.save(discriminator, "models/epoch-5000-discriminator.pt")
|
|
torch.save(generator, "models/epoch-5000-generator.pt")
|
|
print("Training complete!")
|
|
|
|
start_training()
|