⚗️ | More architectural changes

AudioUtils.py

@@ -1,18 +1,41 @@
 import torch
 import torch.nn.functional as F
 
-def stereo_tensor_to_mono(waveform):
-    if waveform.shape[0] > 1:
-        # Average across channels
-        mono_waveform = torch.mean(waveform, dim=0, keepdim=True)
-    else:
-        # Already mono
-        mono_waveform = waveform
-    return mono_waveform
 
-def stretch_tensor(tensor, target_length):
+def stereo_tensor_to_mono(waveform: torch.Tensor) -> torch.Tensor:
-    scale_factor = target_length / tensor.size(1)
+    mono_tensor = torch.mean(waveform, dim=0, keepdim=True)
+    return mono_tensor
 
-    tensor = F.interpolate(tensor, scale_factor=scale_factor, mode='linear', align_corners=False)
 
-    return tensor
+def pad_tensor(audio_tensor: torch.Tensor, target_length: int = 512) -> torch.Tensor:
+    padding_amount = target_length - audio_tensor.size(-1)
+    if padding_amount <= 0:
+        return audio_tensor
+
+    padded_audio_tensor = F.pad(audio_tensor, (0, padding_amount))
+
+    return padded_audio_tensor
+
+
+def split_audio(audio_tensor: torch.Tensor, chunk_size: int = 512, pad_last_tensor: bool = False) -> list[torch.Tensor]:
+    chunks = list(torch.split(audio_tensor, chunk_size, dim=1))
+
+    if pad_last_tensor:
+        last_chunk = chunks[-1]
+
+        if last_chunk.size(-1) < chunk_size:
+                    chunks[-1] = pad_tensor(last_chunk, chunk_size)
+
+    return chunks
+
+
+def reconstruct_audio(chunks: list[torch.Tensor]) -> torch.Tensor:
+    reconstructed_tensor = torch.cat(chunks, dim=-1)
+    return reconstructed_tensor
+
+
+def normalize(audio_tensor: torch.Tensor, eps: float = 1e-8) -> torch.Tensor:
+    max_val = torch.max(torch.abs(audio_tensor))
+    if max_val < eps:
+        return audio_tensor
+    return audio_tensor / max_val

README.md

@@ -18,6 +18,7 @@ SISU (Super Ingenious Sound Upscaler) is a project that uses GANs (Generative Ad
 1. **Set Up**:
    - Make sure you have Python installed (version 3.8 or higher).
    - Install needed packages: `pip install -r requirements.txt`
+   - Install current version of PyTorch (CUDA/ROCm/What ever your device supports)
 
 2. **Prepare Audio Data**:
    - Put your audio files in the `dataset/good` folder.

app.py

@@ -0,0 +1,128 @@
+import argparse
+
+import torch
+import torchaudio
+import torchcodec
+import tqdm
+from accelerate import Accelerator
+
+import AudioUtils
+from generator import SISUGenerator
+
+# Init script argument parser
+parser = argparse.ArgumentParser(description="Training script")
+parser.add_argument("--device", type=str, default="cpu", help="Select device")
+parser.add_argument("--model", type=str, help="Model to use for upscaling")
+parser.add_argument(
+    "--clip_length",
+    type=int,
+    default=8000,
+    help="Internal clip length, leave unspecified if unsure",
+)
+parser.add_argument(
+    "--sample_rate", type=int, default=44100, help="Output clip sample rate"
+)
+parser.add_argument(
+    "--bitrate",
+    type=int,
+    default=192000,
+    help="Output clip bitrate",
+)
+parser.add_argument("-i", "--input", type=str, help="Input audio file")
+parser.add_argument("-o", "--output", type=str, help="Output audio file")
+
+args = parser.parse_args()
+
+if args.sample_rate < 8000:
+    print(
+        "Sample rate cannot be lower than 8000! (44100 is recommended for base models)"
+    )
+    exit()
+
+# ---------------------------
+# Init accelerator
+# ---------------------------
+
+accelerator = Accelerator(mixed_precision="bf16")
+
+# ---------------------------
+# Models
+# ---------------------------
+generator = SISUGenerator()
+
+accelerator.print("🔨 | Compiling models...")
+
+generator = torch.compile(generator)
+
+accelerator.print("✅ | Compiling done!")
+
+# ---------------------------
+# Prepare accelerator
+# ---------------------------
+
+generator = accelerator.prepare(generator)
+
+# ---------------------------
+# Checkpoint helpers
+# ---------------------------
+
+models_dir = args.model
+clip_length = args.clip_length
+input_audio = args.input
+output_audio = args.output
+
+if models_dir:
+    ckpt = torch.load(models_dir)
+
+    accelerator.unwrap_model(generator).load_state_dict(ckpt["G"])
+
+    accelerator.print("💾 | Loaded model!")
+else:
+    print(
+        "Generator model (--model) isn't specified. Do you have the trained model? If not, you need to train it OR acquire it from somewhere (DON'T ASK ME, YET!)"
+    )
+
+
+def start():
+    # To Mono!
+    decoder = torchcodec.decoders.AudioDecoder(input_audio)
+
+    decoded_samples = decoder.get_all_samples()
+    audio = decoded_samples.data
+    original_sample_rate = decoded_samples.sample_rate
+
+    # Support for multichannel audio
+    # audio = AudioUtils.stereo_tensor_to_mono(audio)
+    audio = AudioUtils.normalize(audio)
+
+    resample_transform = torchaudio.transforms.Resample(
+        original_sample_rate, args.sample_rate
+    )
+
+    audio = resample_transform(audio)
+
+    splitted_audio = AudioUtils.split_audio(audio, clip_length)
+    splitted_audio_on_device = [t.view(1, t.shape[0], t.shape[-1]).to(accelerator.device) for t in splitted_audio]
+    processed_audio = []
+    with torch.no_grad():
+        for clip in tqdm.tqdm(splitted_audio_on_device, desc="Processing..."):
+            channels = []
+            for audio_channel in torch.split(clip, 1, dim=1):
+                output_piece = generator(audio_channel)
+                channels.append(output_piece.detach().cpu())
+            output_clip = torch.cat(channels, dim=1)
+            processed_audio.append(output_clip)
+
+    reconstructed_audio = AudioUtils.reconstruct_audio(processed_audio)
+    reconstructed_audio = reconstructed_audio.squeeze(0)
+    print(f"🔊 | Saving {output_audio}!")
+    torchaudio.save_with_torchcodec(
+        uri=output_audio,
+        src=reconstructed_audio,
+        sample_rate=args.sample_rate,
+        channels_first=True,
+        compression=args.bitrate,
+    )
+
+
+start()

data.py

@@ -1,53 +1,71 @@
-from torch.utils.data import Dataset
-import torch.nn.functional as F
-import torch
-import torchaudio
 import os
 import random
-import torchaudio.transforms as T
+
+import torch
+import torchaudio
+import torchcodec.decoders as decoders
+import tqdm
+from torch.utils.data import Dataset
+
 import AudioUtils
 
-class AudioDataset(Dataset):
-    audio_sample_rates = [11025]
-    MAX_LENGTH = 88200  # Define your desired maximum length here
 
-    def __init__(self, input_dir, device):
+class AudioDataset(Dataset):
-        self.input_files = [os.path.join(root, f) for root, _, files in os.walk(input_dir) for f in files if f.endswith('.wav')]
+    audio_sample_rates = [8000, 11025, 12000, 16000, 22050, 24000, 32000, 44100]
-        self.device = device
+
+    def __init__(self, input_dir, clip_length: int = 512, normalize: bool = True):
+        self.clip_length = clip_length
+        self.normalize = normalize
+
+        input_files = [
+            os.path.join(input_dir, f)
+            for f in os.listdir(input_dir)
+            if os.path.isfile(os.path.join(input_dir, f))
+            and f.lower().endswith((".wav", ".mp3", ".flac"))
+        ]
+
+        data = []
+        for audio_clip in tqdm.tqdm(
+            input_files, desc=f"Processing {len(input_files)} audio file(s)"
+        ):
+            decoder = decoders.AudioDecoder(audio_clip)
+            decoded_samples = decoder.get_all_samples()
+
+            audio = decoded_samples.data.float()
+            original_sample_rate = decoded_samples.sample_rate
+
+            if normalize:
+                audio = AudioUtils.normalize(audio)
+
+            splitted_high_quality_audio = AudioUtils.split_audio(audio, clip_length, True)
+
+            if not splitted_high_quality_audio:
+                continue
+
+            for splitted_audio_clip in splitted_high_quality_audio:
+                for audio_clip in torch.split(splitted_audio_clip, 1):
+                    data.append((audio_clip, original_sample_rate))
+
+        self.audio_data = data
 
     def __len__(self):
-        return len(self.input_files)
+        return len(self.audio_data)
 
     def __getitem__(self, idx):
-        # Load high-quality audio
+        audio_clip = self.audio_data[idx]
-        high_quality_audio, original_sample_rate = torchaudio.load(self.input_files[idx], normalize=True)
-
-        # 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_transform_high = torchaudio.transforms.Resample(mangled_sample_rate, original_sample_rate)
+        resample_transform_low = torchaudio.transforms.Resample(
-        low_quality_audio = resample_transform_high(low_quality_audio)
+            audio_clip[1], mangled_sample_rate
+        )
 
-        high_quality_audio = AudioUtils.stereo_tensor_to_mono(high_quality_audio)
+        resample_transform_high = torchaudio.transforms.Resample(
-        low_quality_audio = AudioUtils.stereo_tensor_to_mono(low_quality_audio)
+            mangled_sample_rate, audio_clip[1]
+        )
 
-        # Pad or truncate high-quality audio
+        low_audio_clip = resample_transform_high(resample_transform_low(audio_clip[0]))
-        if high_quality_audio.shape[1] < self.MAX_LENGTH:
+        if audio_clip[0].shape[1] < low_audio_clip.shape[1]:
-            padding = self.MAX_LENGTH - high_quality_audio.shape[1]
+            low_audio_clip = low_audio_clip[:, :audio_clip[0].shape[1]]
-            high_quality_audio = F.pad(high_quality_audio, (0, padding))
+        elif audio_clip[0].shape[1] > low_audio_clip.shape[1]:
-        elif high_quality_audio.shape[1] > self.MAX_LENGTH:
+            low_audio_clip = AudioUtils.pad_tensor(low_audio_clip, self.clip_length)
-            high_quality_audio = high_quality_audio[:, :self.MAX_LENGTH]
+        return ((audio_clip[0], low_audio_clip), (audio_clip[1], mangled_sample_rate))
-
-        # 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

@@ -1,18 +1,31 @@
-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):
+
-    padding = (kernel_size // 2) * dilation
+def discriminator_block(
-    conv_layer = nn.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, dilation=dilation, padding=padding)
+    in_channels,
-    if spectral_norm:
+    out_channels,
-        conv_layer = utils.spectral_norm(conv_layer)
+    kernel_size=15,
-    return nn.Sequential(
+    stride=1,
-        conv_layer,
+    dilation=1
-        nn.LeakyReLU(0.2, inplace=True),
+):
-        nn.BatchNorm1d(out_channels)
+    padding = dilation * (kernel_size - 1) // 2
+
+    conv_layer = nn.Conv1d(
+        in_channels,
+        out_channels,
+        kernel_size=kernel_size,
+        stride=stride,
+        dilation=dilation,
+        padding=padding
     )
 
+    conv_layer = utils.spectral_norm(conv_layer)
+    leaky_relu = nn.LeakyReLU(0.2)
+
+    return nn.Sequential(conv_layer, leaky_relu)
+
+
 class AttentionBlock(nn.Module):
     def __init__(self, channels):
         super(AttentionBlock, self).__init__()
@@ -20,36 +33,38 @@ class AttentionBlock(nn.Module):
             nn.Conv1d(channels, channels // 4, kernel_size=1),
             nn.ReLU(),
             nn.Conv1d(channels // 4, channels, kernel_size=1),
-            nn.Sigmoid()
+            nn.Sigmoid(),
         )
 
     def forward(self, x):
         attention_weights = self.attention(x)
-        return x * attention_weights
+        return x + (x * attention_weights)
+
 
 class SISUDiscriminator(nn.Module):
-    def __init__(self, layers=4): #Increased base layer count
+    def __init__(self, layers=8):
         super(SISUDiscriminator, self).__init__()
-        self.model = nn.Sequential(
+        self.discriminator_blocks = nn.Sequential(
-            discriminator_block(1, layers, kernel_size=7, stride=4), #Aggressive downsampling
+            # 1 -> 32
-            discriminator_block(layers, layers * 2, kernel_size=5, stride=2),
+            discriminator_block(2, layers),
-            discriminator_block(layers * 2, layers * 4, kernel_size=5, dilation=2),
+            AttentionBlock(layers),
-            discriminator_block(layers * 4, layers * 8, kernel_size=5, dilation=4),
+            # 32 -> 64
-            AttentionBlock(layers * 8), #Added attention
+            discriminator_block(layers, layers * 2, dilation=2),
-            discriminator_block(layers * 8, layers * 16, kernel_size=5, dilation=8),
+            # 64 -> 128
-            discriminator_block(layers * 16, layers * 16, kernel_size=3, dilation=1),
+            discriminator_block(layers * 2, layers * 4, dilation=4),
-            discriminator_block(layers * 16, layers * 8, kernel_size=3, dilation=2),
+            AttentionBlock(layers * 4),
-            discriminator_block(layers * 8, layers * 4, kernel_size=3, dilation=1),
+            # 128 -> 256
-            discriminator_block(layers * 4, layers * 2, kernel_size=3, stride=1),
+            discriminator_block(layers * 4, layers * 8, stride=4),
-            discriminator_block(layers * 2, layers, kernel_size=3, stride=1),
+            # 256 -> 512
-            discriminator_block(layers, 1, kernel_size=3, stride=1, spectral_norm=False) #last layer no spectral norm.
+            # discriminator_block(layers * 8, layers * 16, stride=4)
         )
-        self.global_avg_pool = nn.AdaptiveAvgPool1d(1)
+
-        self.sigmoid = nn.Sigmoid()
+        self.final_conv = nn.Conv1d(layers * 8, 1, kernel_size=3, padding=1)
+
+        self.avg_pool = nn.AdaptiveAvgPool1d(1)
 
     def forward(self, x):
-        x = self.model(x)
+        x = self.discriminator_blocks(x)
-        x = self.global_avg_pool(x)
+        x = self.final_conv(x)
-        x = x.view(-1, 1)
+        x = self.avg_pool(x)
-        x = self.sigmoid(x)
+        return x.squeeze(2)
-        return x

generator.py

@@ -1,30 +1,46 @@
+import torch
 import torch.nn as nn
 
-def conv_block(in_channels, out_channels, kernel_size=3, dilation=1):
+
+def GeneratorBlock(in_channels, out_channels, kernel_size=3, stride=1, dilation=1):
+    padding = (kernel_size - 1) // 2 * dilation
     return nn.Sequential(
-        nn.Conv1d(in_channels, out_channels, kernel_size=kernel_size, dilation=dilation, padding=(kernel_size // 2) * dilation),
+        nn.Conv1d(
-        nn.BatchNorm1d(out_channels),
+            in_channels,
-        nn.PReLU()
+            out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            dilation=dilation,
+            padding=padding
+        ),
+        nn.InstanceNorm1d(out_channels),
+        nn.PReLU(num_parameters=1, init=0.1),
     )
 
+
 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.ReLU(inplace=True),
             nn.Conv1d(channels // 4, channels, kernel_size=1),
-            nn.Sigmoid()
+            nn.Sigmoid(),
         )
 
     def forward(self, x):
         attention_weights = self.attention(x)
-        return x * attention_weights
+        return x + (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.conv_layers = nn.Sequential(
+            *[GeneratorBlock(channels, channels) for _ in range(num_convs)]
+        )
+
         self.attention = AttentionBlock(channels)
 
     def forward(self, x):
@@ -33,20 +49,74 @@ class ResidualInResidualBlock(nn.Module):
         x = self.attention(x)
         return x + residual
 
+def UpsampleBlock(in_channels, out_channels):
+    return nn.Sequential(
+        nn.ConvTranspose1d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=4,
+            stride=2,
+            padding=1
+        ),
+        nn.InstanceNorm1d(out_channels),
+        nn.PReLU(num_parameters=1, init=0.1)
+    )
+
 class SISUGenerator(nn.Module):
-    def __init__(self, layer=4, num_rirb=4): #increased base layer and rirb amounts
+    def __init__(self, channels=32, num_rirb=1):
         super(SISUGenerator, self).__init__()
-        self.conv1 = nn.Sequential(
+
-            nn.Conv1d(1, layer, kernel_size=7, padding=3),
+        self.first_conv = GeneratorBlock(1, channels)
-            nn.BatchNorm1d(layer),
+
-            nn.PReLU(),
+        self.downsample = GeneratorBlock(channels, channels * 2, stride=2)
+        self.downsample_attn = AttentionBlock(channels * 2)
+        self.downsample_2 = GeneratorBlock(channels * 2, channels * 4, stride=2)
+        self.downsample_2_attn = AttentionBlock(channels * 4)
+
+        self.rirb = ResidualInResidualBlock(channels * 4)
+        # self.rirb = nn.Sequential(
+        #     *[ResidualInResidualBlock(channels * 4) for _ in range(num_rirb)]
+        # )
+
+        self.upsample = UpsampleBlock(channels * 4, channels * 2)
+        self.upsample_attn = AttentionBlock(channels * 2)
+        self.compress_1 = GeneratorBlock(channels * 4, channels * 2)
+
+        self.upsample_2 = UpsampleBlock(channels * 2, channels)
+        self.upsample_2_attn = AttentionBlock(channels)
+        self.compress_2 = GeneratorBlock(channels * 2, channels)
+
+        self.final_conv = nn.Sequential(
+            nn.Conv1d(channels, 1, kernel_size=7, padding=3),
+            nn.Tanh()
         )
-        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
+        residual_input = x
-        x = self.conv1(x)
+        x1 = self.first_conv(x)
-        x = self.rir_blocks(x)
+
-        x = self.final_layer(x)
+        x2 = self.downsample(x1)
-        return x + residual
+        x2 = self.downsample_attn(x2)
+
+        x3 = self.downsample_2(x2)
+        x3 = self.downsample_2_attn(x3)
+
+        x_rirb = self.rirb(x3)
+
+        up1 = self.upsample(x_rirb)
+        up1 = self.upsample_attn(up1)
+
+        cat1 = torch.cat((up1, x2), dim=1)
+        comp1 = self.compress_1(cat1)
+
+        up2 = self.upsample_2(comp1)
+        up2 = self.upsample_2_attn(up2)
+
+        cat2 = torch.cat((up2, x1), dim=1)
+        comp2 = self.compress_2(cat2)
+
+        learned_residual = self.final_conv(comp2)
+
+        output = residual_input + learned_residual
+        return output

requirements.txt

@@ -1,14 +0,0 @@
-filelock==3.16.1
-fsspec==2024.10.0
-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
-setuptools==70.2.0
-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

@@ -1,189 +1,254 @@
+import argparse
+import datetime
+import os
+
 import torch
 import torch.nn as nn
 import torch.optim as optim
-
-import torch.nn.functional as F
-import torchaudio
 import tqdm
+from accelerate import Accelerator
+from torch.utils.data import DataLoader, DistributedSampler
 
-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
+from generator import SISUGenerator
+from utils.TrainingTools import discriminator_train, generator_train
 
-import torchaudio.transforms as T
+# ---------------------------
-
+# Argument parsing
-# Init script argument parser
+# ---------------------------
-parser = argparse.ArgumentParser(description="Training script")
+parser = argparse.ArgumentParser(description="Training script (safer defaults)")
-parser.add_argument("--generator", type=str, default=None,
+parser.add_argument("--resume", action="store_true", help="Resume training")
-                    help="Path to the generator model file")
+parser.add_argument(
-parser.add_argument("--discriminator", type=str, default=None,
+    "--epochs", type=int, default=5000, help="Number of training epochs"
-                    help="Path to the discriminator model file")
+)
-parser.add_argument("--device", type=str, default="cpu",  help="Select device")
+parser.add_argument("--batch_size", type=int, default=8, help="Batch size")
-parser.add_argument("--epoch", type=int, default=0, help="Current epoch for model versioning")
+parser.add_argument("--num_workers", type=int, default=2, help="DataLoader num_workers")
-parser.add_argument("--verbose", action="store_true",  help="Increase output verbosity")
+parser.add_argument("--debug", action="store_true", help="Print debug logs")
-
+parser.add_argument(
+    "--no_pin_memory", action="store_true", help="Disable pin_memory even on CUDA"
+)
 args = parser.parse_args()
 
-device = torch.device(args.device if torch.cuda.is_available() else "cpu")
+# ---------------------------
-print(f"Using device: {device}")
+# Init accelerator
+# ---------------------------
 
-mfcc_transform = T.MFCC(
+accelerator = Accelerator(mixed_precision="bf16")
-    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)
+# Models
-    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()
-
-    # 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=128, shuffle=True)
-
-# Initialize models and move them to device
 generator = SISUGenerator()
 discriminator = SISUDiscriminator()
 
-epoch: int = args.epoch
+accelerator.print("🔨 | Compiling models...")
 
-generator = generator.to(device)
+generator = torch.compile(generator)
-discriminator = discriminator.to(device)
+discriminator = torch.compile(discriminator)
 
-if args.generator is not None:
+accelerator.print("✅ | Compiling done!")
-    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()
+# Dataset / DataLoader
-criterion_d = nn.BCELoss()
+# ---------------------------
+accelerator.print("📊 | Fetching dataset...")
+dataset = AudioDataset("./dataset", 8192)
 
-# Optimizers
+sampler = DistributedSampler(dataset) if accelerator.num_processes > 1 else None
-optimizer_g = optim.Adam(generator.parameters(), lr=0.0001, betas=(0.5, 0.999))
+pin_memory = torch.cuda.is_available() and not args.no_pin_memory
-optimizer_d = optim.Adam(discriminator.parameters(), lr=0.0001, betas=(0.5, 0.999))
 
-# Scheduler
+train_loader = DataLoader(
-scheduler_g = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_g, mode='min', factor=0.5, patience=5)
+    dataset,
-scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min', factor=0.5, patience=5)
+    sampler=sampler,
+    batch_size=args.batch_size,
+    shuffle=(sampler is None),
+    num_workers=args.num_workers,
+    pin_memory=pin_memory,
+    persistent_workers=pin_memory,
+)
 
-models_dir = "models"
+if not train_loader or not train_loader.batch_size or train_loader.batch_size == 0:
+    accelerator.print("🪹 | There is no data to train with! Exiting...")
+    exit()
+
+loader_batch_size = train_loader.batch_size
+
+accelerator.print("✅ | Dataset fetched!")
+
+# ---------------------------
+# Losses / Optimizers / Scalers
+# ---------------------------
+
+optimizer_g = optim.AdamW(
+    generator.parameters(), lr=0.0003, betas=(0.5, 0.999), weight_decay=0.0001
+)
+optimizer_d = optim.AdamW(
+    discriminator.parameters(), lr=0.0003, betas=(0.5, 0.999), weight_decay=0.0001
+)
+
+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
+)
+
+criterion_d = nn.MSELoss()
+
+# ---------------------------
+# Prepare accelerator
+# ---------------------------
+
+generator, discriminator, optimizer_g, optimizer_d, train_loader = accelerator.prepare(
+    generator, discriminator, optimizer_g, optimizer_d, train_loader
+)
+
+# ---------------------------
+# Checkpoint helpers
+# ---------------------------
+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
+def save_ckpt(path, epoch):
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        accelerator.save(
+            {
+                "epoch": epoch,
+                "G": accelerator.unwrap_model(generator).state_dict(),
+                "D": accelerator.unwrap_model(discriminator).state_dict(),
+                "optG": optimizer_g.state_dict(),
+                "optD": optimizer_d.state_dict(),
+                "schedG": scheduler_g.state_dict(),
+                "schedD": scheduler_d.state_dict(),
+            },
+            path,
+        )
 
-        # ========= 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 =========
+start_epoch = 0
-            batch_size = high_quality_clip[0].size(0)
+if args.resume:
-            real_labels = torch.ones(batch_size, 1).to(device)
+    ckpt_path = os.path.join(models_dir, "last.pt")
-            fake_labels = torch.zeros(batch_size, 1).to(device)
+    ckpt = torch.load(ckpt_path)
 
-            # ========= DISCRIMINATOR =========
+    accelerator.unwrap_model(generator).load_state_dict(ckpt["G"])
-            discriminator.train()
+    accelerator.unwrap_model(discriminator).load_state_dict(ckpt["D"])
-            d_loss = discriminator_train(high_quality_sample, low_quality_sample, real_labels, fake_labels)
+    optimizer_g.load_state_dict(ckpt["optG"])
+    optimizer_d.load_state_dict(ckpt["optD"])
+    scheduler_g.load_state_dict(ckpt["schedG"])
+    scheduler_d.load_state_dict(ckpt["schedD"])
 
-            # ========= GENERATOR =========
+    start_epoch = ckpt.get("epoch", 1)
-            generator.train()
+    accelerator.print(f"🔁 | Resumed from epoch {start_epoch}!")
-            generator_output, combined_loss, adversarial_loss, mfcc_l = generator_train(low_quality_sample, high_quality_sample, real_labels)
 
-            if debug:
+real_buf = torch.full((loader_batch_size, 1), 1, device=accelerator.device, dtype=torch.float32)
-                print(d_loss, combined_loss, adversarial_loss, mfcc_l)
+fake_buf = torch.zeros((loader_batch_size, 1), device=accelerator.device, dtype=torch.float32)
-            scheduler_d.step(d_loss)
-            scheduler_g.step(combined_loss)
 
-            # ========= SAVE LATEST AUDIO =========
+accelerator.print("🏋️ | Started training...")
-            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
+try:
+    for epoch in range(start_epoch, args.epochs):
+        generator.train()
+        discriminator.train()
 
-        if generator_epoch % 10 == 0:
+        discriminator_time = 0
-            print(f"Saved epoch {new_epoch}!")
+        generator_time = 0
-            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:
+        running_d, running_g, steps = 0.0, 0.0, 0
-            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")
+        progress_bar = tqdm.tqdm(train_loader, desc=f"Epoch {epoch} | D {discriminator_time}μs | G {generator_time}μs")
-    torch.save(generator, "models/epoch-5000-generator.pt")
-    print("Training complete!")
 
-start_training()
+        for i, (
+            (high_quality, low_quality),
+            (high_sample_rate, low_sample_rate),
+        ) in enumerate(progress_bar):
+            batch_size = high_quality.size(0)
+
+            real_labels = real_buf[:batch_size].to(accelerator.device)
+            fake_labels = fake_buf[:batch_size].to(accelerator.device)
+
+            with accelerator.autocast():
+                generator_output = generator(low_quality)
+
+            # --- Discriminator ---
+            d_time = datetime.datetime.now()
+            optimizer_d.zero_grad(set_to_none=True)
+            with accelerator.autocast():
+                d_loss = discriminator_train(
+                    high_quality,
+                    low_quality.detach(),
+                    real_labels,
+                    fake_labels,
+                    discriminator,
+                    criterion_d,
+                    generator_output.detach()
+                )
+
+            accelerator.backward(d_loss)
+            optimizer_d.step()
+            discriminator_time = (datetime.datetime.now() - d_time).microseconds
+
+            # --- Generator ---
+            g_time = datetime.datetime.now()
+            optimizer_g.zero_grad(set_to_none=True)
+            with accelerator.autocast():
+                g_total, g_adv = generator_train(
+                    low_quality,
+                    high_quality,
+                    real_labels,
+                    generator,
+                    discriminator,
+                    criterion_d,
+                    generator_output
+                )
+
+            accelerator.backward(g_total)
+            torch.nn.utils.clip_grad_norm_(generator.parameters(), 1)
+            optimizer_g.step()
+            generator_time = (datetime.datetime.now() - g_time).microseconds
+
+            d_val = accelerator.gather(d_loss.detach()).mean()
+            g_val = accelerator.gather(g_total.detach()).mean()
+
+            if torch.isfinite(d_val):
+                running_d += d_val.item()
+            else:
+                accelerator.print(
+                    f"🫥 | NaN in discriminator loss at step {i}, skipping update."
+                )
+
+            if torch.isfinite(g_val):
+                running_g += g_val.item()
+            else:
+                accelerator.print(
+                    f"🫥 | NaN in generator loss at step {i}, skipping update."
+                )
+
+            steps += 1
+            progress_bar.set_description(f"Epoch {epoch} | D {discriminator_time}μs | G {generator_time}μs")
+
+        # epoch averages & schedulers
+        if steps == 0:
+            accelerator.print("🪹 | No steps in epoch (empty dataloader?). Exiting.")
+            break
+
+        mean_d = running_d / steps
+        mean_g = running_g / steps
+
+        scheduler_d.step(mean_d)
+        scheduler_g.step(mean_g)
+
+        save_ckpt(os.path.join(models_dir, "last.pt"), epoch)
+        accelerator.print(f"🤝 | Epoch {epoch} done | D {mean_d:.4f} | G {mean_g:.4f}")
+
+except Exception:
+    try:
+        save_ckpt(os.path.join(models_dir, "crash_last.pt"), epoch)
+        accelerator.print(f"💾 | Saved crash checkpoint for epoch {epoch}")
+    except Exception as e:
+        accelerator.print("😬 | Failed saving crash checkpoint:", e)
+    raise
+
+accelerator.print("🏁 | Training finished.")

utils/MultiResolutionSTFTLoss.py

@@ -0,0 +1,68 @@
+from typing import Dict, List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchaudio.transforms as T
+
+
+class MultiResolutionSTFTLoss(nn.Module):
+    def __init__(
+        self,
+        fft_sizes: List[int] = [512, 1024, 2048, 4096, 8192],
+        hop_sizes: List[int] = [64, 128, 256, 512, 1024],
+        win_lengths: List[int] = [256, 512, 1024, 2048, 4096],
+        eps: float = 1e-7,
+        center: bool = True
+    ):
+        super().__init__()
+
+        self.eps = eps
+        self.n_resolutions = len(fft_sizes)
+
+        self.stft_transforms = nn.ModuleList()
+        for i, (n_fft, hop_len, win_len) in enumerate(zip(fft_sizes, hop_sizes, win_lengths)):
+            stft = T.Spectrogram(
+                n_fft=n_fft,
+                hop_length=hop_len,
+                win_length=win_len,
+                window_fn=torch.hann_window,
+                power=None,
+                center=center,
+                pad_mode="reflect",
+                normalized=False,
+            )
+            self.stft_transforms.append(stft)
+
+    def forward(
+        self, y_true: torch.Tensor, y_pred: torch.Tensor
+    ) -> Dict[str, torch.Tensor]:
+        if y_true.dim() == 3 and y_true.size(1) == 1:
+            y_true = y_true.squeeze(1)
+        if y_pred.dim() == 3 and y_pred.size(1) == 1:
+            y_pred = y_pred.squeeze(1)
+
+        sc_loss = 0.0
+        mag_loss = 0.0
+
+        for stft in self.stft_transforms:
+            stft.window = stft.window.to(y_true.device)
+            stft_true = stft(y_true)
+            stft_pred = stft(y_pred)
+
+            stft_mag_true = torch.abs(stft_true)
+            stft_mag_pred = torch.abs(stft_pred)
+
+            norm_true = torch.linalg.norm(stft_mag_true, dim=(-2, -1))
+            norm_diff = torch.linalg.norm(stft_mag_true - stft_mag_pred, dim=(-2, -1))
+            sc_loss += torch.mean(norm_diff / (norm_true + self.eps))
+
+            log_mag_pred = torch.log(stft_mag_pred + self.eps)
+            log_mag_true = torch.log(stft_mag_true + self.eps)
+            mag_loss += F.l1_loss(log_mag_pred, log_mag_true)
+
+        sc_loss /= self.n_resolutions
+        mag_loss /= self.n_resolutions
+        total_loss = sc_loss + mag_loss
+
+        return {"total": total_loss, "sc": sc_loss, "mag": mag_loss}

utils/TrainingTools.py

@@ -0,0 +1,58 @@
+import torch
+
+from utils.MultiResolutionSTFTLoss import MultiResolutionSTFTLoss
+
+# stft_loss_fn = MultiResolutionSTFTLoss(
+#     fft_sizes=[512, 1024, 2048, 4096],
+#     hop_sizes=[128, 256, 512, 1024],
+#     win_lengths=[512, 1024, 2048, 4096]
+# )
+stft_loss_fn = MultiResolutionSTFTLoss(
+    fft_sizes=[512, 1024, 2048],
+    hop_sizes=[64, 128, 256],
+    win_lengths=[256, 512, 1024]
+)
+
+def signal_mae(input_one: torch.Tensor, input_two: torch.Tensor) -> torch.Tensor:
+    absolute_difference = torch.abs(input_one - input_two)
+    return torch.mean(absolute_difference)
+
+
+def discriminator_train(
+    high_quality,
+    low_quality,
+    high_labels,
+    low_labels,
+    discriminator,
+    criterion,
+    generator_output
+):
+
+    real_pair = torch.cat((low_quality, high_quality), dim=1)
+    decision_real = discriminator(real_pair)
+    d_loss_real = criterion(decision_real, high_labels)
+
+    fake_pair = torch.cat((low_quality, generator_output), dim=1)
+    decision_fake = discriminator(fake_pair)
+    d_loss_fake = criterion(decision_fake, low_labels)
+
+    d_loss = (d_loss_real + d_loss_fake) / 2.0
+    return d_loss
+
+
+def generator_train(
+    low_quality, high_quality, real_labels, generator, discriminator, adv_criterion, generator_output):
+
+    fake_pair = torch.cat((low_quality, generator_output), dim=1)
+
+    discriminator_decision = discriminator(fake_pair)
+    adversarial_loss = adv_criterion(discriminator_decision, real_labels)
+
+    mae_loss = signal_mae(generator_output, high_quality)
+    stft_loss = stft_loss_fn(high_quality, generator_output)["total"]
+
+    lambda_mae = 10.0
+    lambda_stft = 2.5
+    lambda_adv = 2.5
+    combined_loss = (lambda_mae * mae_loss) + (lambda_stft * stft_loss) + (lambda_adv * adversarial_loss)
+    return combined_loss, adversarial_loss