:albemic: | Experimenting with other model layouts.

data.py

@@ -4,22 +4,20 @@ import torch
 import torchaudio
 import os
 import random
-
 import torchaudio.transforms as T
 import AudioUtils
 
 class AudioDataset(Dataset):
-    #audio_sample_rates = [8000, 11025, 16000, 22050]
     audio_sample_rates = [11025]
+    MAX_LENGTH = 88200  # Define your desired maximum length here
 
-    def __init__(self, input_dir):
+    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 __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)
@@ -32,4 +30,24 @@ class AudioDataset(Dataset):
         resample_transform_high = torchaudio.transforms.Resample(mangled_sample_rate, original_sample_rate)
         low_quality_audio = resample_transform_high(low_quality_audio)
 
-        return (AudioUtils.stereo_tensor_to_mono(high_quality_audio), original_sample_rate), (AudioUtils.stereo_tensor_to_mono(low_quality_audio), mangled_sample_rate)
+        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)
+
+        return (high_quality_audio, original_sample_rate), (low_quality_audio, mangled_sample_rate)

discriminator.py

@@ -2,36 +2,54 @@ 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):
+def discriminator_block(in_channels, out_channels, kernel_size=3, stride=1, dilation=1, spectral_norm=True):
     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(
-        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) # Changed activation to LeakyReLU
+        conv_layer,
+        nn.LeakyReLU(0.2, inplace=True),
+        nn.BatchNorm1d(out_channels)
     )
 
-class SISUDiscriminator(nn.Module):
-    def __init__(self):
-        super(SISUDiscriminator, self).__init__()
-        layers = 4 # Increased base layer count
-        self.model = nn.Sequential(
-            # Initial Convolution
-            discriminator_block(1, layers, kernel_size=7, stride=2, dilation=1), # Downsample
-
-            # Core Discriminator Blocks with varied kernels and dilations
-            discriminator_block(layers, layers * 2, kernel_size=5, stride=2, dilation=1), # Downsample
-            discriminator_block(layers * 2, layers * 4, kernel_size=5, dilation=4),
-            discriminator_block(layers * 4, layers * 4, kernel_size=5, dilation=16),
-            discriminator_block(layers * 4, layers * 2, kernel_size=3, dilation=8),
-            discriminator_block(layers * 2, layers, kernel_size=3, dilation=1),
-            # Final Convolution
-            discriminator_block(layers, 1, kernel_size=3, stride=1),
+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()
         )
-        self.global_avg_pool = nn.AdaptiveAvgPool1d(1)
 
     def forward(self, x):
-      # Gaussian noise is not necessary here for discriminator as it is already implicit in the training process
+        attention_weights = self.attention(x)
+        return x * attention_weights
+
+class SISUDiscriminator(nn.Module):
+    def __init__(self, layers=4): #Increased base layer count
+        super(SISUDiscriminator, self).__init__()
+        self.model = nn.Sequential(
+            discriminator_block(1, layers, kernel_size=7, stride=4), #Aggressive downsampling
+            discriminator_block(layers, layers * 2, kernel_size=5, stride=2),
+            discriminator_block(layers * 2, layers * 4, kernel_size=5, dilation=2),
+            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=3, stride=1),
+            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.
+        )
+        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

generator.py

@@ -7,30 +7,46 @@ def conv_block(in_channels, out_channels, kernel_size=3, dilation=1):
         nn.PReLU()
     )
 
+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):
+    def __init__(self, layer=4, num_rirb=4): #increased base layer and rirb amounts
         super(SISUGenerator, self).__init__()
-        layer = 4 # Increased base layer count
         self.conv1 = nn.Sequential(
             nn.Conv1d(1, layer, kernel_size=7, padding=3),
             nn.BatchNorm1d(layer),
             nn.PReLU(),
         )
-        self.conv_blocks = nn.Sequential(
-            conv_block(layer, layer, kernel_size=3, dilation=1),      # Local details
-            conv_block(layer, layer*2, kernel_size=5, dilation=2),    # Local Context
-            conv_block(layer*2, layer*2, kernel_size=3, dilation=16),   # Longer range dependencies
-            conv_block(layer*2, layer*2, kernel_size=5, dilation=8),    # Wider context
-            conv_block(layer*2, layer, kernel_size=5, dilation=2),    # Local Context
-            conv_block(layer, layer, kernel_size=3, dilation=1),      # Local details
-        )
-        self.final_layer = nn.Sequential(
-            nn.Conv1d(layer, 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.conv_blocks(x)
+        x = self.rir_blocks(x)
         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==2.2.1
-pytorch-triton-rocm==3.2.0+git0d4682f0
+numpy==2.2.3
+pytorch-triton-rocm==3.2.0+git4b3bb1f8
 setuptools==70.2.0
-sympy==1.13.1
-torch==2.6.0.dev20241222+rocm6.2.4
-torchaudio==2.6.0.dev20241222+rocm6.2.4
+sympy==1.13.3
+torch==2.7.0.dev20250226+rocm6.3
+torchaudio==2.6.0.dev20250226+rocm6.3
 tqdm==4.67.1
 typing_extensions==4.12.2

training.py

@@ -10,6 +10,8 @@ import argparse
 
 import math
 
+import os
+
 from torch.utils.data import random_split
 from torch.utils.data import DataLoader
 
@@ -18,8 +20,36 @@ 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)
+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,  # Adjust to your sample rate
+    n_mfcc=20,
+    melkwargs={'n_fft': 2048, 'hop_length': 512} # adjust n_fft and hop_length to your needs.
+).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]
+    return torch.mean((mfccs_true - mfccs_pred)**2)
 
 def discriminator_train(high_quality, low_quality, real_labels, fake_labels):
     optimizer_d.zero_grad()
@@ -43,62 +73,48 @@ def discriminator_train(high_quality, low_quality, real_labels, fake_labels):
 
     return d_loss
 
-def generator_train(low_quality, real_labels):
+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)
-    g_loss = criterion_g(discriminator_decision, real_labels)
+    adversarial_loss = criterion_g(discriminator_decision, real_labels)
 
-    g_loss.backward()
+    combined_loss = adversarial_loss + 0.5 * mfcc_l
+
+    combined_loss.backward()
     optimizer_g.step()
-    return generator_output
 
-# 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")
+    return (generator_output, combined_loss, adversarial_loss, mfcc_l)
 
-args = parser.parse_args()
-
-# Check for CUDA availability
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-print(f"Using device: {device}")
+debug = args.verbose
 
 # Initialize dataset and dataloader
 dataset_dir = './dataset/good'
-dataset = AudioDataset(dataset_dir)
-
-# ========= MULTIPLE =========
-
-# 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)
+dataset = AudioDataset(dataset_dir, device)
 
 # ========= SINGLE =========
 
-train_data_loader = DataLoader(dataset, batch_size=16, shuffle=True)
+train_data_loader = DataLoader(dataset, batch_size=128, 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))
+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()
@@ -111,32 +127,10 @@ 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():
-
-    # 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 = 5000
     for generator_epoch in range(generator_epochs):
         low_quality_audio = (torch.empty((1)), 1)
@@ -146,10 +140,10 @@ def start_training():
         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 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].to(device), high_quality_clip[1])
-            low_quality_sample = (low_quality_clip[0].to(device), low_quality_clip[1])
+            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)
@@ -158,32 +152,38 @@ def start_training():
 
             # ========= DISCRIMINATOR =========
             discriminator.train()
-            discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
+            d_loss = discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
 
             # ========= GENERATOR =========
             generator.train()
-            generator_output = generator_train(low_quality_sample, real_labels)
+            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
-            low_quality_audio = low_quality_clip
-            ai_enhanced_audio = (generator_output, high_quality_clip[1])
+            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])
 
-        #metric = snr(high_quality_audio[0].to(device), ai_enhanced_audio[0])
-        #print(f"Generator metric {metric}!")
-        #scheduler_g.step(metric)
+        new_epoch = generator_epoch+epoch
 
         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])
+            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])
 
-        torch.save(discriminator.state_dict(), f"models/current-epoch-discriminator.pt")
-        torch.save(generator.state_dict(), f"models/current-epoch-generator.pt")
+        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(), "models/epoch-5000-discriminator.pt")
-    torch.save(generator.state_dict(), "models/epoch-5000-generator.pt")
+    torch.save(discriminator, "models/epoch-5000-discriminator.pt")
+    torch.save(generator, "models/epoch-5000-generator.pt")
     print("Training complete!")
 
 start_training()