Merge pull request 'Merge new-arch, because it has proven to give the best results' (#1) from new-arch into main

Reviewed-on: #1

AudioUtils.py

@@ -1,97 +1,18 @@
 import torch
 import torch.nn.functional as F
 
-
-def stereo_tensor_to_mono(waveform: torch.Tensor) -> torch.Tensor:
-    """
-    Convert stereo (C, N) to mono (1, N). Ensures a channel dimension.
-    """
-    if waveform.dim() == 1:
-        waveform = waveform.unsqueeze(0)  # (N,) -> (1, N)
-
+def stereo_tensor_to_mono(waveform):
     if waveform.shape[0] > 1:
-        mono_waveform = torch.mean(waveform, dim=0, keepdim=True)  # (1, N)
+        # 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):
+    scale_factor = target_length / tensor.size(1)
 
-def stretch_tensor(tensor: torch.Tensor, target_length: int) -> torch.Tensor:
-    """
-    Stretch audio along time dimension to target_length.
-    Input assumed (1, N). Returns (1, target_length).
-    """
-    if tensor.dim() == 1:
-        tensor = tensor.unsqueeze(0)  # ensure (1, N)
+    tensor = F.interpolate(tensor, scale_factor=scale_factor, mode='linear', align_corners=False)
 
-    tensor = tensor.unsqueeze(0)  # (1, 1, N) for interpolate
-    stretched = F.interpolate(
-        tensor, size=target_length, mode="linear", align_corners=False
-    )
-    return stretched.squeeze(0)  # back to (1, target_length)
-
-
-def pad_tensor(audio_tensor: torch.Tensor, target_length: int = 128) -> torch.Tensor:
-    """
-    Pad to fixed length. Input assumed (1, N). Returns (1, target_length).
-    """
-    if audio_tensor.dim() == 1:
-        audio_tensor = audio_tensor.unsqueeze(0)
-
-    current_length = audio_tensor.shape[-1]
-    if current_length < target_length:
-        padding_needed = target_length - current_length
-        padding_tuple = (0, padding_needed)
-        padded_audio_tensor = F.pad(
-            audio_tensor, padding_tuple, mode="constant", value=0
-        )
-    else:
-        padded_audio_tensor = audio_tensor[..., :target_length]  # crop if too long
-
-    return padded_audio_tensor
-
-
-def split_audio(
-    audio_tensor: torch.Tensor, chunk_size: int = 128
-) -> list[torch.Tensor]:
-    """
-    Split into chunks of (1, chunk_size).
-    """
-    if not isinstance(chunk_size, int) or chunk_size <= 0:
-        raise ValueError("chunk_size must be a positive integer.")
-
-    if audio_tensor.dim() == 1:
-        audio_tensor = audio_tensor.unsqueeze(0)
-
-    num_samples = audio_tensor.shape[-1]
-    if num_samples == 0:
-        return []
-
-    chunks = list(torch.split(audio_tensor, chunk_size, dim=-1))
-    return chunks
-
-
-def reconstruct_audio(chunks: list[torch.Tensor]) -> torch.Tensor:
-    """
-    Reconstruct audio from chunks. Returns (1, N).
-    """
-    if not chunks:
-        return torch.empty(1, 0)
-
-    chunks = [c if c.dim() == 2 else c.unsqueeze(0) for c in chunks]
-    try:
-        reconstructed_tensor = torch.cat(chunks, dim=-1)
-    except RuntimeError as e:
-        raise RuntimeError(
-            f"Failed to concatenate audio chunks. Ensure chunks have compatible shapes "
-            f"for concatenation along dim -1. Original error: {e}"
-        )
-
-    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  # silence, skip normalization
-    return audio_tensor / max_val
+    return tensor

app.py

@@ -1,97 +0,0 @@
-import argparse
-
-import torch
-import torchaudio
-import torchcodec
-import tqdm
-
-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=16384,
-    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()
-
-device = torch.device(args.device if torch.cuda.is_available() else "cpu")
-print(f"Using device: {device}")
-
-generator = SISUGenerator().to(device)
-
-generator = torch.compile(generator)
-
-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, map_location=device)
-    generator.load_state_dict(ckpt["G"])
-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
-
-    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.to(device) for t in splitted_audio]
-    processed_audio = []
-
-    for clip in tqdm.tqdm(splitted_audio_on_device, desc="Processing..."):
-        processed_audio.append(generator(clip))
-
-    reconstructed_audio = AudioUtils.reconstruct_audio(processed_audio)
-    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

@@ -1,79 +1,53 @@
+from torch.utils.data import Dataset
+import torch.nn.functional as F
+import torch
+import torchaudio
 import os
 import random
-
-import torchaudio
-import torchcodec.decoders as decoders
-import tqdm
-from torch.utils.data import Dataset
-
+import torchaudio.transforms as T
 import AudioUtils
 
-
 class AudioDataset(Dataset):
     audio_sample_rates = [11025]
+    MAX_LENGTH = 44100  # Define your desired maximum length here
 
-    def __init__(self, input_dir, clip_length: int = 8000, 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()  # ensure float32
-            original_sample_rate = decoded_samples.sample_rate
-
-            audio = AudioUtils.stereo_tensor_to_mono(audio)
-            if normalize:
-                audio = AudioUtils.normalize(audio)
-
-            mangled_sample_rate = random.choice(self.audio_sample_rates)
-            resample_transform_low = torchaudio.transforms.Resample(
-                original_sample_rate, mangled_sample_rate
-            )
-            resample_transform_high = torchaudio.transforms.Resample(
-                mangled_sample_rate, original_sample_rate
-            )
-
-            low_audio = resample_transform_high(resample_transform_low(audio))
-
-            splitted_high_quality_audio = AudioUtils.split_audio(audio, clip_length)
-            splitted_low_quality_audio = AudioUtils.split_audio(low_audio, clip_length)
-
-            if not splitted_high_quality_audio or not splitted_low_quality_audio:
-                continue  # skip empty or invalid clips
-
-            splitted_high_quality_audio[-1] = AudioUtils.pad_tensor(
-                splitted_high_quality_audio[-1], clip_length
-            )
-            splitted_low_quality_audio[-1] = AudioUtils.pad_tensor(
-                splitted_low_quality_audio[-1], clip_length
-            )
-
-            for high_quality_data, low_quality_data in zip(
-                splitted_high_quality_audio, splitted_low_quality_audio
-            ):
-                data.append(
-                    (
-                        (high_quality_data, low_quality_data),
-                        (original_sample_rate, mangled_sample_rate),
-                    )
-                )
-
-        self.audio_data = data
+    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.audio_data)
+        return len(self.input_files)
 
     def __getitem__(self, idx):
-        return self.audio_data[idx]
+        # Load high-quality audio
+        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)
+        low_quality_audio = resample_transform_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)
+
+        return (high_quality_audio, original_sample_rate), (low_quality_audio, mangled_sample_rate)

discriminator.py

@@ -1,16 +1,8 @@
+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,
-    use_instance_norm=True,
-):
+def discriminator_block(in_channels, out_channels, kernel_size=3, stride=1, dilation=1, spectral_norm=True, use_instance_norm=True):
     padding = (kernel_size // 2) * dilation
     conv_layer = nn.Conv1d(
         in_channels,
@@ -18,7 +10,7 @@ def discriminator_block(
         kernel_size=kernel_size,
         stride=stride,
         dilation=dilation,
-        padding=padding,
+        padding=padding
     )
 
     if spectral_norm:
@@ -32,7 +24,6 @@ def discriminator_block(
 
     return nn.Sequential(*layers)
 
-
 class AttentionBlock(nn.Module):
     def __init__(self, channels):
         super(AttentionBlock, self).__init__()
@@ -40,30 +31,27 @@ class AttentionBlock(nn.Module):
             nn.Conv1d(channels, channels // 4, kernel_size=1),
             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
 
-
 class SISUDiscriminator(nn.Module):
-    def __init__(self, layers=32):
+    def __init__(self, base_channels=16):
         super(SISUDiscriminator, self).__init__()
+        layers = base_channels
         self.model = nn.Sequential(
-            discriminator_block(1, layers, kernel_size=7, stride=1),
-            discriminator_block(layers, layers * 2, kernel_size=5, stride=2),
-            discriminator_block(layers * 2, layers * 4, kernel_size=5, dilation=2),
+            discriminator_block(1, layers, kernel_size=7, stride=1, spectral_norm=True, use_instance_norm=False),
+            discriminator_block(layers, layers * 2, kernel_size=5, stride=2, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 2, layers * 4, kernel_size=5, stride=1, dilation=2, spectral_norm=True, use_instance_norm=True),
             AttentionBlock(layers * 4),
-            discriminator_block(layers * 4, layers * 8, kernel_size=5, dilation=4),
-            discriminator_block(layers * 8, layers * 2, kernel_size=5, stride=2),
-            discriminator_block(
-                layers * 2,
-                1,
-                spectral_norm=False,
-                use_instance_norm=False,
-            ),
+            discriminator_block(layers * 4, layers * 8, kernel_size=5, stride=1, dilation=4, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 8, layers * 4, kernel_size=5, stride=2, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 4, layers * 2, kernel_size=3, stride=1, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers * 2, layers, kernel_size=3, stride=1, spectral_norm=True, use_instance_norm=True),
+            discriminator_block(layers, 1, kernel_size=3, stride=1, spectral_norm=False, use_instance_norm=False)
         )
 
         self.global_avg_pool = nn.AdaptiveAvgPool1d(1)

file_utils.py

@@ -0,0 +1,28 @@
+import json
+
+filepath = "my_data.json"
+
+def write_data(filepath, data):
+  try:
+    with open(filepath, 'w') as f:
+      json.dump(data, f, indent=4)  # Use indent for pretty formatting
+    print(f"Data written to '{filepath}'")
+  except Exception as e:
+    print(f"Error writing to file: {e}")
+
+
+def read_data(filepath):
+  try:
+    with open(filepath, 'r') as f:
+      data = json.load(f)
+    print(f"Data read from '{filepath}'")
+    return data
+  except FileNotFoundError:
+    print(f"File not found: {filepath}")
+    return None
+  except json.JSONDecodeError:
+    print(f"Error decoding JSON from file: {filepath}")
+    return None
+  except Exception as e:
+    print(f"Error reading from file: {e}")
+    return None

generator.py

@@ -1,7 +1,6 @@
 import torch
 import torch.nn as nn
 
-
 def conv_block(in_channels, out_channels, kernel_size=3, dilation=1):
     return nn.Sequential(
         nn.Conv1d(
@@ -9,32 +8,29 @@ def conv_block(in_channels, out_channels, kernel_size=3, dilation=1):
             out_channels,
             kernel_size=kernel_size,
             dilation=dilation,
-            padding=(kernel_size // 2) * dilation,
+            padding=(kernel_size // 2) * dilation
         ),
         nn.InstanceNorm1d(out_channels),
-        nn.PReLU(),
+        nn.PReLU()
     )
 
-
 class AttentionBlock(nn.Module):
     """
     Simple Channel Attention Block. Learns to weight channels based on their importance.
     """
-
     def __init__(self, channels):
         super(AttentionBlock, self).__init__()
         self.attention = nn.Sequential(
             nn.Conv1d(channels, channels // 4, kernel_size=1),
             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
 
-
 class ResidualInResidualBlock(nn.Module):
     def __init__(self, channels, num_convs=3):
         super(ResidualInResidualBlock, self).__init__()
@@ -51,9 +47,8 @@ class ResidualInResidualBlock(nn.Module):
         x = self.attention(x)
         return x + residual
 
-
 class SISUGenerator(nn.Module):
-    def __init__(self, channels=16, num_rirb=4, alpha=1):
+    def __init__(self, channels=16, num_rirb=4, alpha=1.0):
         super(SISUGenerator, self).__init__()
         self.alpha = alpha
 
@@ -67,9 +62,7 @@ class SISUGenerator(nn.Module):
             *[ResidualInResidualBlock(channels) for _ in range(num_rirb)]
         )
 
-        self.final_layer = nn.Sequential(
-            nn.Conv1d(channels, 1, kernel_size=3, padding=1), nn.Tanh()
-        )
+        self.final_layer = nn.Conv1d(channels, 1, kernel_size=3, padding=1)
 
     def forward(self, x):
         residual_input = x
@@ -78,4 +71,4 @@ class SISUGenerator(nn.Module):
         learned_residual = self.final_layer(x_rirb_out)
         output = residual_input + self.alpha * learned_residual
 
-        return torch.tanh(output)
+        return output

requirements.txt

@@ -0,0 +1,12 @@
+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
+pillow==11.0.0
+setuptools==70.2.0
+sympy==1.13.3
+tqdm==4.67.1
+typing_extensions==4.12.2

training.py

@@ -1,245 +1,194 @@
-import argparse
-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 discriminator import SISUDiscriminator
 from generator import SISUGenerator
-from utils.TrainingTools import discriminator_train, generator_train
+from discriminator import SISUDiscriminator
 
-# ---------------------------
-# Argument parsing
-# ---------------------------
-parser = argparse.ArgumentParser(description="Training script (safer defaults)")
-parser.add_argument("--resume", action="store_true", help="Resume training")
-parser.add_argument(
-    "--epochs", type=int, default=5000, help="Number of training epochs"
-)
-parser.add_argument("--batch_size", type=int, default=8, help="Batch size")
-parser.add_argument("--num_workers", type=int, default=2, help="DataLoader num_workers")
+from training_utils import discriminator_train, generator_train
+import file_utils as Data
+
+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")
-parser.add_argument(
-    "--no_pin_memory", action="store_true", help="Disable pin_memory even on CUDA"
-)
+parser.add_argument("--continue_training", action="store_true", help="Continue training using temp_generator and temp_discriminator models")
+
 args = parser.parse_args()
 
-# ---------------------------
-# Init accelerator
-# ---------------------------
+device = torch.device(args.device if torch.cuda.is_available() else "cpu")
+print(f"Using device: {device}")
 
-accelerator = Accelerator(mixed_precision="bf16")
+# Parameters
+sample_rate = 44100
+n_fft = 2048
+hop_length = 256
+win_length = n_fft
+n_mels = 128
+n_mfcc = 20 # If using MFCC
+
+mfcc_transform = T.MFCC(
+    sample_rate,
+    n_mfcc,
+    melkwargs = {'n_fft': n_fft, 'hop_length': hop_length}
+).to(device)
+
+mel_transform = T.MelSpectrogram(
+    sample_rate=sample_rate, n_fft=n_fft, hop_length=hop_length,
+    win_length=win_length, n_mels=n_mels, power=1.0 # Magnitude Mel
+).to(device)
+
+stft_transform = T.Spectrogram(
+    n_fft=n_fft, win_length=win_length, hop_length=hop_length
+).to(device)
+
+debug = args.debug
+
+# Initialize dataset and dataloader
+dataset_dir = './dataset/good'
+dataset = AudioDataset(dataset_dir, device)
+models_dir = "models"
+os.makedirs(models_dir, exist_ok=True)
+audio_output_dir = "output"
+os.makedirs(audio_output_dir, exist_ok=True)
+
+# ========= SINGLE =========
+
+train_data_loader = DataLoader(dataset, batch_size=64, shuffle=True)
+
+
+# ========= MODELS =========
 
-# ---------------------------
-# Models
-# ---------------------------
 generator = SISUGenerator()
 discriminator = SISUDiscriminator()
 
-accelerator.print("🔨 | Compiling models...")
+epoch: int = args.epoch
+epoch_from_file = Data.read_data(f"{models_dir}/epoch_data.json")
 
-generator = torch.compile(generator)
-discriminator = torch.compile(discriminator)
+if args.continue_training:
+    generator.load_state_dict(torch.load(f"{models_dir}/temp_generator.pt", map_location=device, weights_only=True))
+    discriminator.load_state_dict(torch.load(f"{models_dir}/temp_generator.pt", map_location=device, weights_only=True))
+    epoch = epoch_from_file["epoch"] + 1
+else:
+    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))
 
-accelerator.print("✅ | Compiling done!")
-
-# ---------------------------
-# Dataset / DataLoader
-# ---------------------------
-accelerator.print("📊 | Fetching dataset...")
-dataset = AudioDataset("./dataset")
-
-sampler = DistributedSampler(dataset) if accelerator.num_processes > 1 else None
-pin_memory = torch.cuda.is_available() and not args.no_pin_memory
-
-train_loader = DataLoader(
-    dataset,
-    sampler=sampler,
-    batch_size=args.batch_size,
-    shuffle=(sampler is None),
-    num_workers=args.num_workers,
-    pin_memory=pin_memory,
-    persistent_workers=pin_memory,
-)
-
-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
-)
+generator = generator.to(device)
+discriminator = discriminator.to(device)
 
+# Loss
 criterion_g = nn.BCEWithLogitsLoss()
-criterion_d = nn.MSELoss()
+criterion_d = nn.BCEWithLogitsLoss()
 
-# ---------------------------
-# Prepare accelerator
-# ---------------------------
+# 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))
 
-generator, discriminator, optimizer_g, optimizer_d, train_loader = accelerator.prepare(
-    generator, discriminator, optimizer_g, optimizer_d, train_loader
-)
+# 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)
 
-# ---------------------------
-# 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 =========
+            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)
 
-start_epoch = 0
-if args.resume:
-    ckpt_path = os.path.join(models_dir, "last.pt")
-    ckpt = torch.load(ckpt_path)
-
-    accelerator.unwrap_model(generator).load_state_dict(ckpt["G"])
-    accelerator.unwrap_model(discriminator).load_state_dict(ckpt["D"])
-    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"])
-
-    start_epoch = ckpt.get("epoch", 1)
-    accelerator.print(f"🔁 | Resumed from epoch {start_epoch}!")
-
-real_buf = torch.full(
-    (loader_batch_size, 1), 1, device=accelerator.device, dtype=torch.float32
-)
-fake_buf = torch.zeros(
-    (loader_batch_size, 1), device=accelerator.device, dtype=torch.float32
-)
-
-accelerator.print("🏋️ | Started training...")
-
-try:
-    for epoch in range(start_epoch, args.epochs):
-        generator.train()
+            # ========= DISCRIMINATOR =========
             discriminator.train()
-
-        running_d, running_g, steps = 0.0, 0.0, 0
-
-        for i, (
-            (high_quality, low_quality),
-            (high_sample_rate, low_sample_rate),
-        ) in enumerate(tqdm.tqdm(train_loader, desc=f"Epoch {epoch}")):
-            batch_size = high_quality.size(0)
-
-            real_labels = real_buf[:batch_size].to(accelerator.device)
-            fake_labels = fake_buf[:batch_size].to(accelerator.device)
-
-            # --- Discriminator ---
-            optimizer_d.zero_grad(set_to_none=True)
-            with accelerator.autocast():
             d_loss = discriminator_train(
-                    high_quality,
-                    low_quality,
+                high_quality_sample,
+                low_quality_sample,
                 real_labels,
                 fake_labels,
                 discriminator,
                 generator,
                 criterion_d,
+                optimizer_d
             )
 
-            accelerator.backward(d_loss)
-            torch.nn.utils.clip_grad_norm_(discriminator.parameters(), 1)
-            optimizer_d.step()
-
-            # --- Generator ---
-            optimizer_g.zero_grad(set_to_none=True)
-            with accelerator.autocast():
-                g_total, g_adv = generator_train(
-                    low_quality,
-                    high_quality,
+            # ========= GENERATOR =========
+            generator.train()
+            generator_output, combined_loss, adversarial_loss, mel_l1_tensor, log_stft_l1_tensor, mfcc_l_tensor = generator_train(
+                low_quality_sample,
+                high_quality_sample,
                 real_labels,
                 generator,
                 discriminator,
                 criterion_d,
+                optimizer_g,
+                device,
+                mel_transform,
+                stft_transform,
+                mfcc_transform
             )
 
-            accelerator.backward(g_total)
-            torch.nn.utils.clip_grad_norm_(generator.parameters(), 1)
-            optimizer_g.step()
+            if debug:
+                print(f"D_LOSS: {d_loss.item():.4f}, COMBINED_LOSS: {combined_loss.item():.4f}, ADVERSARIAL_LOSS: {adversarial_loss.item():.4f}, MEL_L1_LOSS: {mel_l1_tensor.item():.4f}, LOG_STFT_L1_LOSS: {log_stft_l1_tensor.item():.4f}, MFCC_LOSS: {mfcc_l_tensor.item():.4f}")
+            scheduler_d.step(d_loss.detach())
+            scheduler_g.step(adversarial_loss.detach())
 
-            d_val = accelerator.gather(d_loss.detach()).mean()
-            g_val = accelerator.gather(g_total.detach()).mean()
+            # ========= 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])
 
-            if torch.isfinite(d_val):
-                running_d += d_val.item()
-            else:
-                accelerator.print(
-                    f"🫥 | NaN in discriminator loss at step {i}, skipping update."
-                )
+        new_epoch = generator_epoch+epoch
 
-            if torch.isfinite(g_val):
-                running_g += g_val.item()
-            else:
-                accelerator.print(
-                    f"🫥 | NaN in generator loss at step {i}, skipping update."
-                )
+        if generator_epoch % 25 == 0:
+            print(f"Saved epoch {new_epoch}!")
+            torchaudio.save(f"{audio_output_dir}/epoch-{new_epoch}-audio-crap.wav", low_quality_audio[0].cpu().detach(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again.
+            torchaudio.save(f"{audio_output_dir}/epoch-{new_epoch}-audio-ai.wav", ai_enhanced_audio[0].cpu().detach(), ai_enhanced_audio[1])
+            torchaudio.save(f"{audio_output_dir}/epoch-{new_epoch}-audio-orig.wav", high_quality_audio[0].cpu().detach(), high_quality_audio[1])
 
-            steps += 1
+        #if debug:
+        #    print(generator.state_dict().keys())
+        #    print(discriminator.state_dict().keys())
+        torch.save(discriminator.state_dict(), f"{models_dir}/temp_discriminator.pt")
+        torch.save(generator.state_dict(), f"{models_dir}/temp_generator.pt")
+        Data.write_data(f"{models_dir}/epoch_data.json", {"epoch": new_epoch})
 
-        # 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
+    torch.save(discriminator, "models/epoch-5000-discriminator.pt")
+    torch.save(generator, "models/epoch-5000-generator.pt")
+    print("Training complete!")
 
-        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.")
+start_training()

training_utils.py

@@ -0,0 +1,144 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+
+import torchaudio
+import torchaudio.transforms as T
+
+def gpu_mfcc_loss(mfcc_transform, 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 mel_spectrogram_l1_loss(mel_transform: T.MelSpectrogram, y_true: torch.Tensor, y_pred: torch.Tensor) -> torch.Tensor:
+    mel_spec_true = mel_transform(y_true)
+    mel_spec_pred = mel_transform(y_pred)
+
+    # Ensure same time dimension length (due to potential framing differences)
+    min_len = min(mel_spec_true.shape[-1], mel_spec_pred.shape[-1])
+    mel_spec_true = mel_spec_true[..., :min_len]
+    mel_spec_pred = mel_spec_pred[..., :min_len]
+
+    # L1 Loss (Mean Absolute Error)
+    loss = torch.mean(torch.abs(mel_spec_true - mel_spec_pred))
+    return loss
+
+def mel_spectrogram_l2_loss(mel_transform: T.MelSpectrogram, y_true: torch.Tensor, y_pred: torch.Tensor) -> torch.Tensor:
+    mel_spec_true = mel_transform(y_true)
+    mel_spec_pred = mel_transform(y_pred)
+
+    min_len = min(mel_spec_true.shape[-1], mel_spec_pred.shape[-1])
+    mel_spec_true = mel_spec_true[..., :min_len]
+    mel_spec_pred = mel_spec_pred[..., :min_len]
+
+    loss = torch.mean((mel_spec_true - mel_spec_pred)**2)
+    return loss
+
+def log_stft_magnitude_loss(stft_transform: T.Spectrogram, y_true: torch.Tensor, y_pred: torch.Tensor, eps: float = 1e-7) -> torch.Tensor:
+    stft_mag_true = stft_transform(y_true)
+    stft_mag_pred = stft_transform(y_pred)
+
+    min_len = min(stft_mag_true.shape[-1], stft_mag_pred.shape[-1])
+    stft_mag_true = stft_mag_true[..., :min_len]
+    stft_mag_pred = stft_mag_pred[..., :min_len]
+
+    loss = torch.mean(torch.abs(torch.log(stft_mag_true + eps) - torch.log(stft_mag_pred + eps)))
+    return loss
+
+def spectral_convergence_loss(stft_transform: T.Spectrogram, y_true: torch.Tensor, y_pred: torch.Tensor, eps: float = 1e-7) -> torch.Tensor:
+    stft_mag_true = stft_transform(y_true)
+    stft_mag_pred = stft_transform(y_pred)
+
+    min_len = min(stft_mag_true.shape[-1], stft_mag_pred.shape[-1])
+    stft_mag_true = stft_mag_true[..., :min_len]
+    stft_mag_pred = stft_mag_pred[..., :min_len]
+
+    norm_true = torch.linalg.norm(stft_mag_true, ord='fro', dim=(-2, -1))
+    norm_diff = torch.linalg.norm(stft_mag_true - stft_mag_pred, ord='fro', dim=(-2, -1))
+
+    loss = torch.mean(norm_diff / (norm_true + eps))
+    return loss
+
+def discriminator_train(high_quality, low_quality, real_labels, fake_labels, discriminator, generator, criterion, optimizer):
+    optimizer.zero_grad()
+
+    # Forward pass for real samples
+    discriminator_decision_from_real = discriminator(high_quality[0])
+    d_loss_real = criterion(discriminator_decision_from_real, real_labels)
+
+    with torch.no_grad():
+        generator_output = generator(low_quality[0])
+    discriminator_decision_from_fake = discriminator(generator_output)
+    d_loss_fake = criterion(discriminator_decision_from_fake, fake_labels.expand_as(discriminator_decision_from_fake))
+
+    d_loss = (d_loss_real + d_loss_fake) / 2.0
+
+    d_loss.backward()
+    # Optional: Gradient Clipping (can be helpful)
+    # nn.utils.clip_grad_norm_(discriminator.parameters(), max_norm=1.0)  # Gradient Clipping
+    optimizer.step()
+
+    return d_loss
+
+def generator_train(
+    low_quality,
+    high_quality,
+    real_labels,
+    generator,
+    discriminator,
+    adv_criterion,
+    g_optimizer,
+    device,
+    mel_transform: T.MelSpectrogram,
+    stft_transform: T.Spectrogram,
+    mfcc_transform: T.MFCC,
+    lambda_adv: float = 1.0,
+    lambda_mel_l1: float = 10.0,
+    lambda_log_stft: float = 1.0,
+    lambda_mfcc: float = 1.0
+):
+    g_optimizer.zero_grad()
+
+    generator_output = generator(low_quality[0])
+
+    discriminator_decision = discriminator(generator_output)
+    adversarial_loss = adv_criterion(discriminator_decision, real_labels.expand_as(discriminator_decision))
+
+    mel_l1 = 0.0
+    log_stft_l1 = 0.0
+    mfcc_l = 0.0
+
+    # Calculate Mel L1 Loss if weight is positive
+    if lambda_mel_l1 > 0:
+        mel_l1 = mel_spectrogram_l1_loss(mel_transform, high_quality[0], generator_output)
+
+    # Calculate Log STFT L1 Loss if weight is positive
+    if lambda_log_stft > 0:
+        log_stft_l1 = log_stft_magnitude_loss(stft_transform, high_quality[0], generator_output)
+
+    # Calculate MFCC Loss if weight is positive
+    if lambda_mfcc > 0:
+        mfcc_l = gpu_mfcc_loss(mfcc_transform, high_quality[0], generator_output)
+
+    mel_l1_tensor = torch.tensor(mel_l1, device=device) if isinstance(mel_l1, float) else mel_l1
+    log_stft_l1_tensor = torch.tensor(log_stft_l1, device=device) if isinstance(log_stft_l1, float) else log_stft_l1
+    mfcc_l_tensor = torch.tensor(mfcc_l, device=device) if isinstance(mfcc_l, float) else mfcc_l
+
+    combined_loss = (lambda_adv * adversarial_loss) + \
+                    (lambda_mel_l1 * mel_l1_tensor) + \
+                    (lambda_log_stft * log_stft_l1_tensor) + \
+                    (lambda_mfcc * mfcc_l_tensor)
+
+    combined_loss.backward()
+    # Optional: Gradient Clipping
+    # nn.utils.clip_grad_norm_(generator.parameters(), max_norm=1.0)
+    g_optimizer.step()
+
+    # 6. Return values for logging
+    return generator_output, combined_loss, adversarial_loss, mel_l1_tensor, log_stft_l1_tensor, mfcc_l_tensor

utils/MultiResolutionSTFTLoss.py

@@ -1,87 +0,0 @@
-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):
-    """
-    Multi-resolution STFT loss.
-    Combines spectral convergence loss and log-magnitude loss
-    across multiple STFT resolutions.
-    """
-
-    def __init__(
-        self,
-        fft_sizes: List[int] = [1024, 2048, 512],
-        hop_sizes: List[int] = [120, 240, 50],
-        win_lengths: List[int] = [600, 1200, 240],
-        eps: float = 1e-7,
-    ):
-        super().__init__()
-
-        self.eps = eps
-        self.n_resolutions = len(fft_sizes)
-
-        self.stft_transforms = nn.ModuleList()
-        for n_fft, hop_len, win_len in zip(fft_sizes, hop_sizes, win_lengths):
-            window = torch.hann_window(win_len)
-            stft = T.Spectrogram(
-                n_fft=n_fft,
-                hop_length=hop_len,
-                win_length=win_len,
-                window_fn=lambda _: window,
-                power=None,  # Keep complex output
-                center=True,
-                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]:
-        """
-        Args:
-            y_true: (B, T) or (B, 1, T) waveform
-            y_pred: (B, T) or (B, 1, T) waveform
-        """
-        # Ensure correct shape (B, T)
-        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 = stft.to(y_pred.device)
-
-            # Complex STFTs: (B, F, T, 2)
-            stft_true = stft(y_true)
-            stft_pred = stft(y_pred)
-
-            # Magnitudes
-            stft_mag_true = torch.abs(stft_true)
-            stft_mag_pred = torch.abs(stft_pred)
-
-            # --- Spectral Convergence Loss ---
-            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 STFT Magnitude Loss ---
-            mag_loss += F.l1_loss(
-                torch.log(stft_mag_pred + self.eps),
-                torch.log(stft_mag_true + self.eps),
-            )
-
-        # Average across resolutions
-        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

@@ -1,60 +0,0 @@
-import torch
-
-# In case if needed again...
-# from utils.MultiResolutionSTFTLoss import MultiResolutionSTFTLoss
-#
-# stft_loss_fn = MultiResolutionSTFTLoss(
-#     fft_sizes=[1024, 2048, 512], hop_sizes=[120, 240, 50], win_lengths=[600, 1200, 240]
-# )
-
-
-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,
-    generator,
-    criterion,
-):
-    decision_high = discriminator(high_quality)
-    d_loss_high = criterion(decision_high, high_labels)
-    # print(f"Is this real?: {discriminator_decision_from_real} | {d_loss_real}")
-
-    decision_low = discriminator(low_quality)
-    d_loss_low = criterion(decision_low, low_labels)
-    # print(f"Is this real?: {discriminator_decision_from_fake} | {d_loss_fake}")
-
-    with torch.no_grad():
-        generator_quality = generator(low_quality)
-    decision_gen = discriminator(generator_quality)
-    d_loss_gen = criterion(decision_gen, low_labels)
-
-    noise = torch.rand_like(high_quality) * 0.08
-    decision_noise = discriminator(high_quality + noise)
-    d_loss_noise = criterion(decision_noise, low_labels)
-
-    d_loss = (d_loss_high + d_loss_low + d_loss_gen + d_loss_noise) / 4.0
-
-    return d_loss
-
-
-def generator_train(
-    low_quality, high_quality, real_labels, generator, discriminator, adv_criterion
-):
-    generator_output = generator(low_quality)
-
-    discriminator_decision = discriminator(generator_output)
-    adversarial_loss = adv_criterion(discriminator_decision, real_labels)
-
-    # Signal similarity
-    similarity_loss = signal_mae(generator_output, high_quality)
-
-    combined_loss = adversarial_loss + (similarity_loss * 100)
-
-    return combined_loss, adversarial_loss