🐛 | Fix loading wrong model.

AudioUtils.py

@@ -16,3 +16,56 @@ def stretch_tensor(tensor, target_length):
     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 = 128):
+    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
+
+    return padded_audio_tensor
+
+def split_audio(audio_tensor: torch.Tensor, chunk_size: int = 128) -> list[torch.Tensor]:
+    if not isinstance(chunk_size, int) or chunk_size <= 0:
+        raise ValueError("chunk_size must be a positive integer.")
+
+    # Handle scalar tensor edge case if necessary
+    if audio_tensor.dim() == 0:
+        return [audio_tensor] if audio_tensor.numel() > 0 else []
+
+    # Identify the dimension to split (usually the last one, representing time/samples)
+    split_dim = -1
+    num_samples = audio_tensor.shape[split_dim]
+
+    if num_samples == 0:
+        return [] # Return empty list if the dimension to split is empty
+
+    # Use torch.split to divide the tensor into chunks
+    # It handles the last chunk being potentially smaller automatically.
+    chunks = list(torch.split(audio_tensor, chunk_size, dim=split_dim))
+
+    return chunks
+
+def reconstruct_audio(chunks: list[torch.Tensor]) -> torch.Tensor:
+    if not chunks:
+        return torch.empty(0)
+
+    if len(chunks) == 1 and chunks[0].dim() == 0:
+        return chunks[0]
+
+    concat_dim = -1
+
+    try:
+        reconstructed_tensor = torch.cat(chunks, dim=concat_dim)
+    except RuntimeError as e:
+        raise RuntimeError(
+            f"Failed to concatenate audio chunks. Ensure chunks have compatible shapes "
+            f"for concatenation along dimension {concat_dim}. Original error: {e}"
+        )
+
+    return reconstructed_tensor

app.py

@@ -0,0 +1,60 @@
+import torch
+import torch.nn as nn
+import torch.optim as optim
+
+import torch.nn.functional as F
+import torchaudio
+import tqdm
+
+import argparse
+import math
+import os
+
+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=1024, help="Internal clip length, leave unspecified if unsure")
+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()
+
+device = torch.device(args.device if torch.cuda.is_available() else "cpu")
+print(f"Using device: {device}")
+
+generator = SISUGenerator()
+
+models_dir = args.model
+clip_length = args.clip_length
+input_audio = args.input
+output_audio = args.output
+
+if models_dir:
+    generator.load_state_dict(torch.load(f"{models_dir}", map_location=device, weights_only=True))
+else:
+    print(f"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!)")
+
+generator = generator.to(device)
+
+def start():
+    # To Mono!
+    audio, original_sample_rate = torchaudio.load(input_audio, normalize=True)
+    audio = AudioUtils.stereo_tensor_to_mono(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(output_audio, reconstructed_audio.cpu().detach(), original_sample_rate)
+
+start()

data.py

@@ -5,49 +5,42 @@ import torchaudio
 import os
 import random
 import torchaudio.transforms as T
+import tqdm
 import AudioUtils
 
 class AudioDataset(Dataset):
     audio_sample_rates = [11025]
-    MAX_LENGTH = 44100  # Define your desired maximum length here
 
-    def __init__(self, input_dir, device):
+    def __init__(self, input_dir, device, clip_length = 1024):
-        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
+        input_files = [os.path.join(root, f) for root, _, files in os.walk(input_dir) for f in files if f.endswith('.wav') or f.endswith('.mp3') or f.endswith('.flac')]
+
+        data = []
+        for audio_clip in tqdm.tqdm(input_files, desc=f"Processing {len(input_files)} audio file(s)"):
+            audio, original_sample_rate = torchaudio.load(audio_clip, normalize=True)
+            audio = AudioUtils.stereo_tensor_to_mono(audio)
+
+            # Generate low-quality audio with random downsampling
+            mangled_sample_rate = random.choice(self.audio_sample_rates)
+            resample_transform_low = torchaudio.transforms.Resample(original_sample_rate, mangled_sample_rate)
+            resample_transform_high = torchaudio.transforms.Resample(mangled_sample_rate, original_sample_rate)
+
+            low_audio = resample_transform_low(audio)
+            low_audio = resample_transform_high(low_audio)
+
+            splitted_high_quality_audio = AudioUtils.split_audio(audio, clip_length)
+            splitted_high_quality_audio[-1] = AudioUtils.pad_tensor(splitted_high_quality_audio[-1], clip_length)
+
+            splitted_low_quality_audio = AudioUtils.split_audio(low_audio, clip_length)
+            splitted_low_quality_audio[-1] = AudioUtils.pad_tensor(splitted_low_quality_audio[-1], clip_length)
+
+            for high_quality_sample, low_quality_sample in zip(splitted_high_quality_audio, splitted_low_quality_audio):
+                data.append(((high_quality_sample, low_quality_sample), (original_sample_rate, mangled_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
+        return 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)
-        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)

file_utils.py

@@ -2,20 +2,22 @@ import json
 
 filepath = "my_data.json"
 
-def write_data(filepath, data):
+def write_data(filepath, data, debug=False):
   try:
     with open(filepath, 'w') as f:
       json.dump(data, f, indent=4)  # Use indent for pretty formatting
-    print(f"Data written to '{filepath}'")
+    if debug:
+        print(f"Data written to '{filepath}'")
   except Exception as e:
     print(f"Error writing to file: {e}")
 
 
-def read_data(filepath):
+def read_data(filepath, debug=False):
   try:
     with open(filepath, 'r') as f:
       data = json.load(f)
-    print(f"Data read from '{filepath}'")
+    if debug:
+        print(f"Data read from '{filepath}'")
     return data
   except FileNotFoundError:
     print(f"File not found: {filepath}")

training.py

@@ -43,40 +43,51 @@ print(f"Using device: {device}")
 
 # Parameters
 sample_rate = 44100
-n_fft = 2048
+n_fft = 1024
-hop_length = 256
 win_length = n_fft
-n_mels = 128
+hop_length = n_fft // 4
-n_mfcc = 20 # If using MFCC
+n_mels = 40
+n_mfcc = 13
 
 mfcc_transform = T.MFCC(
-    sample_rate,
+    sample_rate=sample_rate,
-    n_mfcc,
+    n_mfcc=n_mfcc,
-    melkwargs = {'n_fft': n_fft, 'hop_length': hop_length}
+    melkwargs={
+        'n_fft': n_fft,
+        'hop_length': hop_length,
+        'win_length': win_length,
+        'n_mels': n_mels,
+        'power': 1.0,
+    }
 ).to(device)
 
 mel_transform = T.MelSpectrogram(
-    sample_rate=sample_rate, n_fft=n_fft, hop_length=hop_length,
+    sample_rate=sample_rate,
-    win_length=win_length, n_mels=n_mels, power=1.0 # Magnitude Mel
+    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
+    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"
+models_dir = "./models"
 os.makedirs(models_dir, exist_ok=True)
-audio_output_dir = "output"
+audio_output_dir = "./output"
 os.makedirs(audio_output_dir, exist_ok=True)
 
 # ========= SINGLE =========
 
-train_data_loader = DataLoader(dataset, batch_size=64, shuffle=True)
+train_data_loader = DataLoader(dataset, batch_size=2048, shuffle=True, num_workers=24)
 
 
 # ========= MODELS =========
@@ -85,17 +96,18 @@ generator = SISUGenerator()
 discriminator = SISUDiscriminator()
 
 epoch: int = args.epoch
-epoch_from_file = Data.read_data(f"{models_dir}/epoch_data.json")
 
 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:
+    elif args.discriminator is not None:
         discriminator.load_state_dict(torch.load(args.discriminator, map_location=device, weights_only=True))
+    else:
+        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_discriminator.pt", map_location=device, weights_only=True))
+
+        epoch_from_file = Data.read_data(f"{models_dir}/epoch_data.json")
+        epoch = epoch_from_file["epoch"] + 1
 
 generator = generator.to(device)
 discriminator = discriminator.to(device)
@@ -115,28 +127,35 @@ scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min'
 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)
-        high_quality_audio = (torch.empty((1)), 1)
+        low_quality_audio = ([torch.empty((1))], 1)
-        ai_enhanced_audio = (torch.empty((1)), 1)
+        ai_enhanced_audio = ([torch.empty((1))], 1)
 
         times_correct = 0
 
         # ========= TRAINING =========
-        for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Training epoch {generator_epoch+1}/{generator_epochs}, Current epoch {epoch+1}"):
+        for training_data 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:
+            ## Data structure:
-            high_quality_sample = (high_quality_clip[0], high_quality_clip[1])
+            # [[[float..., float..., float...], [float..., float..., float...]], [original_sample_rate, mangled_sample_rate]]
-            low_quality_sample = (low_quality_clip[0], low_quality_clip[1])
 
             # ========= LABELS =========
-            batch_size = high_quality_clip[0].size(0)
+            good_quality_data = training_data[0][0].to(device)
+            bad_quality_data = training_data[0][1].to(device)
+            original_sample_rate = training_data[1][0]
+            mangled_sample_rate = training_data[1][1]
+
+            batch_size = good_quality_data.size(0)
             real_labels = torch.ones(batch_size, 1).to(device)
             fake_labels = torch.zeros(batch_size, 1).to(device)
 
+            high_quality_audio = (good_quality_data, original_sample_rate)
+            low_quality_audio = (bad_quality_data, mangled_sample_rate)
+
             # ========= DISCRIMINATOR =========
             discriminator.train()
             d_loss = discriminator_train(
-                high_quality_sample,
+                good_quality_data,
-                low_quality_sample,
+                bad_quality_data,
                 real_labels,
                 fake_labels,
                 discriminator,
@@ -148,8 +167,8 @@ def start_training():
             # ========= GENERATOR =========
             generator.train()
             generator_output, combined_loss, adversarial_loss, mel_l1_tensor, log_stft_l1_tensor, mfcc_l_tensor = generator_train(
-                low_quality_sample,
+                bad_quality_data,
-                high_quality_sample,
+                good_quality_data,
                 real_labels,
                 generator,
                 discriminator,
@@ -167,23 +186,14 @@ def start_training():
             scheduler_g.step(adversarial_loss.detach())
 
             # ========= SAVE LATEST AUDIO =========
-            high_quality_audio = (high_quality_clip[0][0], high_quality_clip[1][0])
+            high_quality_audio = (good_quality_data, original_sample_rate)
-            low_quality_audio = (low_quality_clip[0][0], low_quality_clip[1][0])
+            low_quality_audio = (bad_quality_data, original_sample_rate)
-            ai_enhanced_audio = (generator_output[0], high_quality_clip[1][0])
+            ai_enhanced_audio = (generator_output, original_sample_rate)
 
-        new_epoch = generator_epoch+epoch
-
-        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])
-
-        #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")
+
+        new_epoch = generator_epoch+epoch
         Data.write_data(f"{models_dir}/epoch_data.json", {"epoch": new_epoch})
 
 

training_utils.py

@@ -20,12 +20,10 @@ def mel_spectrogram_l1_loss(mel_transform: T.MelSpectrogram, y_true: torch.Tenso
     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
 
@@ -69,11 +67,11 @@ def discriminator_train(high_quality, low_quality, real_labels, fake_labels, dis
     optimizer.zero_grad()
 
     # Forward pass for real samples
-    discriminator_decision_from_real = discriminator(high_quality[0])
+    discriminator_decision_from_real = discriminator(high_quality)
     d_loss_real = criterion(discriminator_decision_from_real, real_labels)
 
     with torch.no_grad():
-        generator_output = generator(low_quality[0])
+        generator_output = generator(low_quality)
     discriminator_decision_from_fake = discriminator(generator_output)
     d_loss_fake = criterion(discriminator_decision_from_fake, fake_labels.expand_as(discriminator_decision_from_fake))
 
@@ -105,7 +103,7 @@ def generator_train(
 ):
     g_optimizer.zero_grad()
 
-    generator_output = generator(low_quality[0])
+    generator_output = generator(low_quality)
 
     discriminator_decision = discriminator(generator_output)
     adversarial_loss = adv_criterion(discriminator_decision, real_labels.expand_as(discriminator_decision))
@@ -116,15 +114,15 @@ def generator_train(
 
     # 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)
+        mel_l1 = mel_spectrogram_l1_loss(mel_transform, high_quality, 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)
+        log_stft_l1 = log_stft_magnitude_loss(stft_transform, high_quality, 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)
+        mfcc_l = gpu_mfcc_loss(mfcc_transform, high_quality, 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