184 lines
7.0 KiB
Python
184 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
|
|
|
|
from torch.utils.data import random_split
|
|
from torch.utils.data import DataLoader
|
|
|
|
from data import AudioDataset
|
|
from generator import SISUGenerator
|
|
from discriminator import SISUDiscriminator
|
|
|
|
# Mel Spectrogram Loss
|
|
class MelSpectrogramLoss(nn.Module):
|
|
def __init__(self, sample_rate=44100, n_fft=2048, hop_length=512, n_mels=128):
|
|
super(MelSpectrogramLoss, self).__init__()
|
|
self.mel_transform = torchaudio.transforms.MelSpectrogram(
|
|
sample_rate=sample_rate,
|
|
n_fft=n_fft,
|
|
hop_length=hop_length,
|
|
n_mels=n_mels
|
|
).to(device) # Move to device
|
|
|
|
def forward(self, y_pred, y_true):
|
|
mel_pred = self.mel_transform(y_pred)
|
|
mel_true = self.mel_transform(y_true)
|
|
return F.l1_loss(mel_pred, mel_true)
|
|
|
|
def snr(y_true, y_pred):
|
|
noise = y_true - y_pred
|
|
signal_power = torch.mean(y_true ** 2)
|
|
noise_power = torch.mean(noise ** 2)
|
|
snr_db = 10 * torch.log10(signal_power / noise_power)
|
|
return snr_db
|
|
|
|
def discriminator_train(high_quality, low_quality, scale, real_labels, fake_labels):
|
|
optimizer_d.zero_grad()
|
|
|
|
discriminator_decision_from_real = discriminator(high_quality)
|
|
# TODO: Experiment with criterions HERE!
|
|
d_loss_real = criterion_d(discriminator_decision_from_real, real_labels)
|
|
|
|
generator_output = generator(low_quality, scale)
|
|
discriminator_decision_from_fake = discriminator(generator_output.detach())
|
|
# TODO: Experiment with criterions HERE!
|
|
d_loss_fake = criterion_d(discriminator_decision_from_fake, fake_labels)
|
|
|
|
d_loss = (d_loss_real + d_loss_fake) / 2.0
|
|
|
|
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, scale, real_labels):
|
|
optimizer_g.zero_grad()
|
|
|
|
generator_output = generator(low_quality, scale)
|
|
discriminator_decision = discriminator(generator_output)
|
|
# TODO: Fix this shit
|
|
g_loss = criterion_g(discriminator_decision, real_labels)
|
|
|
|
g_loss.backward()
|
|
optimizer_g.step()
|
|
return generator_output
|
|
|
|
# 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, target_duration=2.0)
|
|
|
|
dataset_size = len(dataset)
|
|
train_size = int(dataset_size * .9)
|
|
val_size = int(dataset_size-train_size)
|
|
|
|
train_dataset, val_dataset = random_split(dataset, [train_size, val_size])
|
|
|
|
train_data_loader = DataLoader(train_dataset, batch_size=1, shuffle=True)
|
|
val_data_loader = DataLoader(val_dataset, batch_size=1, shuffle=True)
|
|
|
|
# Initialize models and move them to device
|
|
generator = SISUGenerator()
|
|
discriminator = SISUDiscriminator()
|
|
|
|
generator = generator.to(device)
|
|
discriminator = discriminator.to(device)
|
|
|
|
# Loss
|
|
criterion_g = nn.L1Loss()
|
|
criterion_g_mel = MelSpectrogramLoss().to(device)
|
|
criterion_d = nn.BCEWithLogitsLoss()
|
|
|
|
# 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():
|
|
|
|
# Training loop
|
|
|
|
# ========= DISCRIMINATOR PRE-TRAINING =========
|
|
discriminator_epochs = 1
|
|
for discriminator_epoch in range(discriminator_epochs):
|
|
|
|
# ========= TRAINING =========
|
|
for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Epoch {discriminator_epoch+1}/{discriminator_epochs}"):
|
|
high_quality_sample = high_quality_clip[0].to(device)
|
|
low_quality_sample = low_quality_clip[0].to(device)
|
|
|
|
scale = high_quality_clip[0].shape[2]/low_quality_clip[0].shape[2]
|
|
|
|
# ========= LABELS =========
|
|
batch_size = high_quality_sample.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, scale, real_labels, fake_labels)
|
|
|
|
torch.save(discriminator.state_dict(), "models/discriminator-single-shot-pre-train.pt")
|
|
|
|
generator_epochs = 500
|
|
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)
|
|
|
|
# ========= TRAINING =========
|
|
for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Epoch {generator_epoch+1}/{generator_epochs}"):
|
|
high_quality_sample = high_quality_clip[0].to(device)
|
|
low_quality_sample = low_quality_clip[0].to(device)
|
|
|
|
scale = high_quality_clip[0].shape[2]/low_quality_clip[0].shape[2]
|
|
|
|
# ========= 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()
|
|
for _ in range(3):
|
|
discriminator_train(high_quality_sample, low_quality_sample, scale, real_labels, fake_labels)
|
|
|
|
# ========= GENERATOR =========
|
|
generator.train()
|
|
generator_output = generator_train(low_quality_sample, scale, real_labels)
|
|
|
|
# ========= SAVE LATEST AUDIO =========
|
|
high_quality_audio = high_quality_clip
|
|
low_quality_audio = low_quality_clip
|
|
ai_enhanced_audio = (generator_output, high_quality_clip[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(), low_quality_audio[1])
|
|
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])
|
|
|
|
if generator_epoch % 50 == 0:
|
|
torch.save(discriminator.state_dict(), f"models/epoch-{generator_epoch}-discriminator.pt")
|
|
torch.save(generator.state_dict(), f"models/epoch-{generator_epoch}-generator.pt")
|
|
|
|
torch.save(discriminator.state_dict(), "models/epoch-500-discriminator.pt")
|
|
torch.save(generator.state_dict(), "models/epoch-500-generator.pt")
|
|
print("Training complete!")
|
|
|
|
start_training()
|