⚗️ | Experimenting...

data.py

@@ -4,50 +4,49 @@ import torch
 import torchaudio
 import os
 import random
-import torchaudio.transforms as T
-import AudioUtils
+from AudioUtils import stereo_tensor_to_mono, stretch_tensor
 
 class AudioDataset(Dataset):
     audio_sample_rates = [11025]
-    MAX_LENGTH = 44100  # Define your desired maximum length here
 
-    def __init__(self, input_dir, device):
-        self.input_files = [os.path.join(root, f) for root, _, files in os.walk(input_dir) for f in files if f.endswith('.wav')]
-        self.device = device
+    def __init__(self, input_dir):
+        self.input_files = [
+            os.path.join(root, f)
+            for root, _, files in os.walk(input_dir)
+            for f in files if f.endswith('.wav')
+        ]
 
     def __len__(self):
         return len(self.input_files)
 
     def __getitem__(self, idx):
         # Load high-quality audio
-        high_quality_audio, original_sample_rate = torchaudio.load(self.input_files[idx], normalize=True)
+        high_quality_path = self.input_files[idx]
+        high_quality_audio, original_sample_rate = torchaudio.load(high_quality_path)
+        high_quality_audio = stereo_tensor_to_mono(high_quality_audio)
 
         # Generate low-quality audio with random downsampling
         mangled_sample_rate = random.choice(self.audio_sample_rates)
-        resample_transform_low = torchaudio.transforms.Resample(original_sample_rate, mangled_sample_rate)
-        low_quality_audio = resample_transform_low(high_quality_audio)
+        resample_low = torchaudio.transforms.Resample(original_sample_rate, mangled_sample_rate)
+        low_quality_audio = resample_low(high_quality_audio)
 
-        resample_transform_high = torchaudio.transforms.Resample(mangled_sample_rate, original_sample_rate)
-        low_quality_audio = resample_transform_high(low_quality_audio)
+        resample_high = torchaudio.transforms.Resample(mangled_sample_rate, original_sample_rate)
+        low_quality_audio = resample_high(low_quality_audio)
 
-        high_quality_audio = AudioUtils.stereo_tensor_to_mono(high_quality_audio)
-        low_quality_audio = AudioUtils.stereo_tensor_to_mono(low_quality_audio)
-
-        # Pad or truncate high-quality audio
-        if high_quality_audio.shape[1] < self.MAX_LENGTH:
-            padding = self.MAX_LENGTH - high_quality_audio.shape[1]
-            high_quality_audio = F.pad(high_quality_audio, (0, padding))
-        elif high_quality_audio.shape[1] > self.MAX_LENGTH:
-            high_quality_audio = high_quality_audio[:, :self.MAX_LENGTH]
-
-        # Pad or truncate low-quality audio
-        if low_quality_audio.shape[1] < self.MAX_LENGTH:
-            padding = self.MAX_LENGTH - low_quality_audio.shape[1]
-            low_quality_audio = F.pad(low_quality_audio, (0, padding))
-        elif low_quality_audio.shape[1] > self.MAX_LENGTH:
-            low_quality_audio = low_quality_audio[:, :self.MAX_LENGTH]
-
-        high_quality_audio = high_quality_audio.to(self.device)
-        low_quality_audio = low_quality_audio.to(self.device)
+        # Pad or truncate to match a fixed length
+        target_length = 44100  # Adjust this based on your data
+        high_quality_audio = self.pad_or_truncate(high_quality_audio, target_length)
+        low_quality_audio = self.pad_or_truncate(low_quality_audio, target_length)
 
         return (high_quality_audio, original_sample_rate), (low_quality_audio, mangled_sample_rate)
+
+    def pad_or_truncate(self, tensor, target_length):
+        current_length = tensor.size(1)
+        if current_length < target_length:
+            # Pad with zeros
+            padding = target_length - current_length
+            tensor = F.pad(tensor, (0, padding))
+        else:
+            # Truncate to target length
+            tensor = tensor[:, :target_length]
+        return tensor

discriminator.py

@@ -2,57 +2,37 @@ import torch
 import torch.nn as nn
 import torch.nn.utils as utils
 
-def discriminator_block(in_channels, out_channels, kernel_size=3, stride=1, dilation=1, spectral_norm=True):
+def discriminator_block(in_channels, out_channels, kernel_size=3, stride=1, dilation=1):
     padding = (kernel_size // 2) * dilation
-    conv_layer = nn.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding)
-    if spectral_norm:
-        conv_layer = utils.spectral_norm(conv_layer)
     return nn.Sequential(
-        conv_layer,
-        nn.LeakyReLU(0.2, inplace=True),
-        nn.BatchNorm1d(out_channels)
+        utils.spectral_norm(
+            nn.Conv1d(in_channels, out_channels,
+                kernel_size=kernel_size,
+                stride=stride,
+                dilation=dilation,
+                padding=padding
+            )
+        ),
+        nn.BatchNorm1d(out_channels),
+        nn.LeakyReLU(0.2, inplace=True)
     )
 
-class AttentionBlock(nn.Module):
-    def __init__(self, channels):
-        super(AttentionBlock, self).__init__()
-        self.attention = nn.Sequential(
-            nn.Conv1d(channels, channels // 4, kernel_size=1),
-            nn.ReLU(),
-            nn.Conv1d(channels // 4, channels, kernel_size=1),
-            nn.Sigmoid()
-        )
-
-    def forward(self, x):
-        attention_weights = self.attention(x)
-        return x * attention_weights
-
 class SISUDiscriminator(nn.Module):
-    def __init__(self, layers=4): #Increased base layer count
+    def __init__(self):
         super(SISUDiscriminator, self).__init__()
+        layers = 4
         self.model = nn.Sequential(
-            discriminator_block(1, layers, kernel_size=3, stride=1), #Aggressive downsampling
-            discriminator_block(layers, layers * 2, kernel_size=5, stride=2),
-            discriminator_block(layers * 2, layers * 4, kernel_size=5, dilation=4),
-
-            #AttentionBlock(layers * 4), #Added attention
-
-            #discriminator_block(layers * 4, layers * 8, kernel_size=5, dilation=4),
-            #AttentionBlock(layers * 8), #Added attention
-            #discriminator_block(layers * 8, layers * 16, kernel_size=5, dilation=8),
-            #discriminator_block(layers * 16, layers * 16, kernel_size=3, dilation=1),
-            #discriminator_block(layers * 16, layers * 8, kernel_size=3, dilation=2),
-            #discriminator_block(layers * 8, layers * 4, kernel_size=3, dilation=1),
-            discriminator_block(layers * 4, layers * 2, kernel_size=5, stride=2),
-            discriminator_block(layers * 2, layers, kernel_size=3, stride=1),
-            discriminator_block(layers, 1, kernel_size=3, stride=1, spectral_norm=False) #last layer no spectral norm.
+            discriminator_block(1, layers, kernel_size=7, stride=2, dilation=1),
+            discriminator_block(layers, layers * 2, kernel_size=5, stride=2, dilation=1),
+            discriminator_block(layers * 2, layers * 4, kernel_size=3, dilation=4),
+            discriminator_block(layers * 4, layers * 4, kernel_size=5, dilation=8),
+            discriminator_block(layers * 4, layers * 2, kernel_size=3, dilation=16),
+            discriminator_block(layers * 2, layers, kernel_size=5, dilation=2),
+            discriminator_block(layers, 1, kernel_size=3, stride=1)
         )
         self.global_avg_pool = nn.AdaptiveAvgPool1d(1)
-        self.sigmoid = nn.Sigmoid()
 
     def forward(self, x):
         x = self.model(x)
         x = self.global_avg_pool(x)
-        x = x.view(-1, 1)
-        x = self.sigmoid(x)
-        return x
+        return x.view(-1, 1)

generator.py

@@ -1,52 +1,41 @@
 import torch.nn as nn
 
-def conv_block(in_channels, out_channels, kernel_size=3, dilation=1):
+def conv_residual_block(in_channels, out_channels, kernel_size=3, dilation=1):
+    padding = (kernel_size // 2) * dilation
     return nn.Sequential(
-        nn.Conv1d(in_channels, out_channels, kernel_size=kernel_size, dilation=dilation, padding=(kernel_size // 2) * dilation),
+        nn.Conv1d(in_channels, out_channels, kernel_size, dilation=dilation, padding=padding),
         nn.BatchNorm1d(out_channels),
-        nn.PReLU()
+        nn.PReLU(),
+        nn.Conv1d(out_channels, out_channels, kernel_size, dilation=dilation, padding=padding),
+        nn.BatchNorm1d(out_channels)
     )
 
-class AttentionBlock(nn.Module):
-    def __init__(self, channels):
-        super(AttentionBlock, self).__init__()
-        self.attention = nn.Sequential(
-            nn.Conv1d(channels, channels // 4, kernel_size=1),
-            nn.ReLU(),
-            nn.Conv1d(channels // 4, channels, kernel_size=1),
-            nn.Sigmoid()
-        )
-
-    def forward(self, x):
-        attention_weights = self.attention(x)
-        return x * attention_weights
-
-class ResidualInResidualBlock(nn.Module):
-    def __init__(self, channels, num_convs=3):
-        super(ResidualInResidualBlock, self).__init__()
-        self.conv_layers = nn.Sequential(*[conv_block(channels, channels) for _ in range(num_convs)])
-        self.attention = AttentionBlock(channels)
-
-    def forward(self, x):
-        residual = x
-        x = self.conv_layers(x)
-        x = self.attention(x)
-        return x + residual
-
 class SISUGenerator(nn.Module):
-    def __init__(self, layer=4, num_rirb=4): #increased base layer and rirb amounts
+    def __init__(self):
         super(SISUGenerator, self).__init__()
+        layers = 4
         self.conv1 = nn.Sequential(
-            nn.Conv1d(1, layer, kernel_size=7, padding=3),
-            nn.BatchNorm1d(layer),
-            nn.PReLU(),
+            nn.Conv1d(1, layers, kernel_size=7, padding=3),
+            nn.BatchNorm1d(layers),
+            nn.PReLU()
+        )
+
+        self.conv_blocks = nn.Sequential(
+            conv_residual_block(layers, layers, kernel_size=3, dilation=1),
+            conv_residual_block(layers, layers * 2, kernel_size=5, dilation=2),
+            conv_residual_block(layers * 2, layers * 4, kernel_size=3, dilation=16),
+            conv_residual_block(layers * 4, layers * 2, kernel_size=5, dilation=8),
+            conv_residual_block(layers * 2, layers, kernel_size=5, dilation=2),
+            conv_residual_block(layers, layers, kernel_size=3, dilation=1)
+        )
+
+        self.final_layer = nn.Sequential(
+            nn.Conv1d(layers, 1, kernel_size=3, padding=1)
         )
-        self.rir_blocks = nn.Sequential(*[ResidualInResidualBlock(layer) for _ in range(num_rirb)])
-        self.final_layer = nn.Conv1d(layer, 1, kernel_size=3, padding=1)
 
     def forward(self, x):
         residual = x
         x = self.conv1(x)
-        x = self.rir_blocks(x)
+        x = self.conv_blocks(x) + x  # Adding residual connection after blocks
         x = self.final_layer(x)
         return x + residual

requirements.txt

@@ -4,11 +4,11 @@ Jinja2==3.1.4
 MarkupSafe==2.1.5
 mpmath==1.3.0
 networkx==3.4.2
-numpy==2.2.3
-pytorch-triton-rocm==3.2.0+git4b3bb1f8
+numpy==2.2.1
+pytorch-triton-rocm==3.2.0+git0d4682f0
 setuptools==70.2.0
-sympy==1.13.3
-torch==2.7.0.dev20250226+rocm6.3
-torchaudio==2.6.0.dev20250226+rocm6.3
+sympy==1.13.1
+torch==2.6.0.dev20241222+rocm6.2.4
+torchaudio==2.6.0.dev20241222+rocm6.2.4
 tqdm==4.67.1
 typing_extensions==4.12.2

training.py

@@ -10,8 +10,6 @@ import argparse
 
 import math
 
-import os
-
 from torch.utils.data import random_split
 from torch.utils.data import DataLoader
 
@@ -20,37 +18,8 @@ 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("--debug", action="store_true",  help="Print debug logs")
-
-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 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()
@@ -74,49 +43,56 @@ def discriminator_train(high_quality, low_quality, real_labels, fake_labels):
 
     return d_loss
 
-def generator_train(low_quality, high_quality, real_labels):
+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])
-
-    #mfcc_l = gpu_mfcc_loss(high_quality[0], generator_output)
-
     discriminator_decision = discriminator(generator_output)
-    adversarial_loss = criterion_g(discriminator_decision, real_labels)
+    g_loss = criterion_g(discriminator_decision, real_labels)
 
-    #combined_loss = adversarial_loss + 0.5 * mfcc_l
-
-    adversarial_loss.backward()
+    g_loss.backward()
     optimizer_g.step()
+    return generator_output
 
-    #return (generator_output, combined_loss, adversarial_loss, mfcc_l)
-    return (generator_output, adversarial_loss)
+def first(objects):
+    if len(objects) >= 1:
+        return objects[0]
+    return objects
 
-debug = args.debug
+# 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, device)
+dataset = AudioDataset(dataset_dir)
 
 # ========= SINGLE =========
 
-train_data_loader = DataLoader(dataset, batch_size=256, shuffle=True)
+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
+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)
 
-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()
@@ -129,9 +105,6 @@ optimizer_d = optim.Adam(discriminator.parameters(), lr=0.0001, betas=(0.5, 0.99
 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):
@@ -142,10 +115,10 @@ def start_training():
         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 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], high_quality_clip[1])
-            low_quality_sample = (low_quality_clip[0], low_quality_clip[1])
+            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)
@@ -154,39 +127,38 @@ def start_training():
 
             # ========= DISCRIMINATOR =========
             discriminator.train()
-            d_loss = discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
+            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)
-            generator_output, adversarial_loss = generator_train(low_quality_sample, high_quality_sample, real_labels)
-
-            if debug:
-                print(d_loss, adversarial_loss)
-            scheduler_d.step(d_loss)
-            scheduler_g.step(adversarial_loss)
+            generator_output = generator_train(low_quality_sample, real_labels)
 
             # ========= 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])
+            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)
 
-        new_epoch = generator_epoch+epoch
+            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 {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])
+            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])
 
-        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.state_dict(), f"models/current-epoch-discriminator.pt")
+        torch.save(generator.state_dict(), f"models/current-epoch-generator.pt")
 
-    torch.save(discriminator, "models/epoch-5000-discriminator.pt")
-    torch.save(generator, "models/epoch-5000-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()