⚗️ | Experimenting...

training.py

@@ -55,6 +55,11 @@ def generator_train(low_quality, real_labels):
     optimizer_g.step()
     return generator_output
 
+def first(objects):
+    if len(objects) >= 1:
+        return objects[0]
+    return objects
+
 # Init script argument parser
 parser = argparse.ArgumentParser(description="Training script")
 parser.add_argument("--generator", type=str, default=None,
@@ -72,17 +77,6 @@ print(f"Using device: {device}")
 dataset_dir = './dataset/good'
 dataset = AudioDataset(dataset_dir)
 
-# ========= MULTIPLE =========
-
-# dataset_size = len(dataset)
-# train_size = int(dataset_size * .9)
-# val_size = int(dataset_size-train_size)
-
-#train_dataset, val_dataset = random_split(dataset, [train_size, val_size])
-
-# train_data_loader = DataLoader(train_dataset, batch_size=1, shuffle=True)
-# val_data_loader = DataLoader(val_dataset, batch_size=1, shuffle=True)
-
 # ========= SINGLE =========
 
 train_data_loader = DataLoader(dataset, batch_size=16, shuffle=True)
@@ -112,31 +106,6 @@ scheduler_g = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_g, mode='min'
 scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min', factor=0.5, patience=5)
 
 def start_training():
-
-    # Training loop
-
-    # ========= DISCRIMINATOR PRE-TRAINING =========
-    # discriminator_epochs = 1
-    # for discriminator_epoch in range(discriminator_epochs):
-
-    #     # ========= TRAINING =========
-    #     for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Epoch {discriminator_epoch+1}/{discriminator_epochs}"):
-    #         high_quality_sample = high_quality_clip[0].to(device)
-    #         low_quality_sample = low_quality_clip[0].to(device)
-
-    #         scale = high_quality_clip[0].shape[2]/low_quality_clip[0].shape[2]
-
-    #         # ========= LABELS =========
-    #         batch_size = high_quality_sample.size(0)
-    #         real_labels = torch.ones(batch_size, 1).to(device)
-    #         fake_labels = torch.zeros(batch_size, 1).to(device)
-
-    #         # ========= DISCRIMINATOR =========
-    #         discriminator.train()
-    #         discriminator_train(high_quality_sample, low_quality_sample, scale, real_labels, fake_labels)
-
-    #     torch.save(discriminator.state_dict(), "models/discriminator-single-shot-pre-train.pt")
-
     generator_epochs = 5000
     for generator_epoch in range(generator_epochs):
         low_quality_audio = (torch.empty((1)), 1)
@@ -165,9 +134,15 @@ def start_training():
             generator_output = generator_train(low_quality_sample, real_labels)
 
             # ========= SAVE LATEST AUDIO =========
-            high_quality_audio = high_quality_clip
-            low_quality_audio = low_quality_clip
-            ai_enhanced_audio = (generator_output, high_quality_clip[1])
+            high_quality_audio = (first(high_quality_clip[0]), high_quality_clip[1][0])
+            low_quality_audio = (first(low_quality_clip[0]), low_quality_clip[1][0])
+            ai_enhanced_audio = (first(generator_output[0]), high_quality_clip[1][0])
+            print(high_quality_audio)
+
+            print(f"Saved epoch {generator_epoch}!")
+            torchaudio.save(f"./output/epoch-{generator_epoch}-audio-crap.wav", low_quality_audio[0][0].cpu(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again.
+            torchaudio.save(f"./output/epoch-{generator_epoch}-audio-ai.wav", ai_enhanced_audio[0][0].cpu(), ai_enhanced_audio[1])
+            torchaudio.save(f"./output/epoch-{generator_epoch}-audio-orig.wav", high_quality_audio[0][0].cpu(), high_quality_audio[1])
 
         #metric = snr(high_quality_audio[0].to(device), ai_enhanced_audio[0])
         #print(f"Generator metric {metric}!")