NikkeDoy/SISU
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

⚗️ | Experimenting still...

Browse Source
This commit is contained in:
NikkeDoy
2024-12-25 00:09:57 +02:00
parent 1000692f32
commit eca71ff5ea
6 changed files with 167 additions and 149 deletions
Download Patch File Download Diff File Expand all files Collapse all files

155
training.py
View File

@ -6,66 +6,73 @@ 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
# 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 perceptual_loss(y_true, y_pred):
return torch.mean((y_true - y_pred) ** 2)
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):
def discriminator_train(high_quality, low_quality, real_labels, fake_labels):
optimizer_d.zero_grad()
discriminator_decision_from_real = discriminator(high_quality)
# TODO: Experiment with criterions HERE!
# Forward pass for real samples
discriminator_decision_from_real = discriminator(high_quality[0])
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!
integer_scale = math.ceil(high_quality[1]/low_quality[1])
# Forward pass for fake samples (from generator output)
generator_output = generator(low_quality[0], integer_scale)
resample_transform = torchaudio.transforms.Resample(low_quality[1] * integer_scale, high_quality[1]).to(device)
resampled = resample_transform(generator_output.detach())
discriminator_decision_from_fake = discriminator(resampled)
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
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):
def generator_train(low_quality, real_labels, target_sample_rate=44100):
optimizer_g.zero_grad()
generator_output = generator(low_quality, scale)
discriminator_decision = discriminator(generator_output)
# TODO: Fix this shit
scale = math.ceil(target_sample_rate/low_quality[1])
# Forward pass for fake samples (from generator output)
generator_output = generator(low_quality[0], scale)
resample_transform = torchaudio.transforms.Resample(low_quality[1] * scale, target_sample_rate).to(device)
resampled = resample_transform(generator_output)
discriminator_decision = discriminator(resampled)
g_loss = criterion_g(discriminator_decision, real_labels)
g_loss.backward()
optimizer_g.step()
return generator_output
return resampled
# 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")
@ -73,28 +80,38 @@ print(f"Using device: {device}")
# Initialize dataset and dataloader
dataset_dir = './dataset/good'
dataset = AudioDataset(dataset_dir, target_duration=2.0)
dataset = AudioDataset(dataset_dir)
dataset_size = len(dataset)
train_size = int(dataset_size * .9)
val_size = int(dataset_size-train_size)
# ========= MULTIPLE =========
train_dataset, val_dataset = random_split(dataset, [train_size, val_size])
# dataset_size = len(dataset)
# train_size = int(dataset_size * .9)
# val_size = int(dataset_size-train_size)
train_data_loader = DataLoader(train_dataset, batch_size=1, shuffle=True)
val_data_loader = DataLoader(val_dataset, batch_size=1, shuffle=True)
#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)
# ========= SINGLE =========
train_data_loader = DataLoader(dataset, batch_size=1, 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.L1Loss()
criterion_g_mel = MelSpectrogramLoss().to(device)
criterion_d = nn.BCEWithLogitsLoss()
criterion_d = nn.BCELoss()
# Optimizers
optimizer_g = optim.Adam(generator.parameters(), lr=0.0001, betas=(0.5, 0.999))
@ -109,39 +126,40 @@ def start_training():
# Training loop
# ========= DISCRIMINATOR PRE-TRAINING =========
discriminator_epochs = 1
for discriminator_epoch in range(discriminator_epochs):
# 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)
# # ========= 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]
# 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)
# # ========= 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)
# # ========= 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")
# torch.save(discriminator.state_dict(), "models/discriminator-single-shot-pre-train.pt")
generator_epochs = 500
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}"):
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]
# 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)
@ -150,21 +168,20 @@ def start_training():
# ========= DISCRIMINATOR =========
discriminator.train()
for _ in range(3):
discriminator_train(high_quality_sample, low_quality_sample, scale, real_labels, fake_labels)
discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
# ========= GENERATOR =========
generator.train()
generator_output = generator_train(low_quality_sample, scale, real_labels)
generator_output = generator_train(low_quality_sample, real_labels, high_quality_sample[1])
# ========= 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)
#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}!")