{"cells":[{"cell_type":"code","execution_count":1,"id":"67a85157-37a9-4aea-aaef-f1802ac133c2","metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["Using device: cuda\n","Loading taxonomy data...\n","Loading training data...\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"cd8b2332bdf94a13ab6e87762447b74c","version_major":2,"version_minor":0},"text/plain":["Validating files: 0%| | 0/28564 [00:00"]},"metadata":{},"output_type":"display_data"},{"name":"stdout","output_type":"stream","text":["Saved final model.\n","Training completed in 112.87 minutes\n"]}],"source":"\"\"\"\n# BirdCLEF 2025 - RegNet Full Training (No Validation Split)\n\nThis notebook implements a RegNet model for the BirdCLEF 2025 competition.\nIt uses the entire dataset for training without a validation split to maximize\nthe amount of training data for the final model.\n\"\"\"\n\nimport os\nimport gc\nimport warnings\nimport logging\nimport time\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport librosa\nimport cv2\nfrom tqdm.auto import tqdm\nfrom pathlib import Path\nfrom collections import Counter\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nfrom torch.utils.data import Dataset, DataLoader\nimport timm\n\nwarnings.filterwarnings(\"ignore\")\nlogging.basicConfig(level=logging.INFO)\n\n# Set random seeds for reproducibility\nSEED = 42\ntorch.manual_seed(SEED)\nnp.random.seed(SEED)\n\n# Configuration class\nclass CFG:\n # Paths\n train_audio_dir = \"C:/Users/lohit/Desktop/Radboud University/Machine Learning in Practice (NWI-IMC030)/Challenge 2/birdclef-2025/train_audio\"\n train_csv = \"C:/Users/lohit/Desktop/Radboud University/Machine Learning in Practice (NWI-IMC030)/Challenge 2/birdclef-2025/train.csv\"\n taxonomy_csv = \"C:/Users/lohit/Desktop/Radboud University/Machine Learning in Practice (NWI-IMC030)/Challenge 2/birdclef-2025/taxonomy.csv\"\n model_save_path = \"./models\"\n \n # Audio parameters\n sample_rate = 32000\n duration = 5 # seconds\n \n # Mel spectrogram parameters\n n_fft = 1024\n hop_length = 512\n n_mels = 64\n fmin = 50\n fmax = 14000\n \n # Image parameters\n img_size = 224\n \n # Training parameters\n model_name = 'regnety_008'\n batch_size = 32\n num_workers = 0 # Avoid dataloader deadlocks\n epochs = 15\n learning_rate = 3e-4\n weight_decay = 1e-5\n \n # Mixed precision training\n use_amp = True\n \n # Debug options\n debug_mode = False # Set to False for full training\n debug_samples = 500 # Number of samples to use in debug mode\n \n # Device\n device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n\ncfg = CFG()\n\n# Create model directory if it doesn't exist\nos.makedirs(cfg.model_save_path, exist_ok=True)\n\n# Load and prepare data\ndef load_data():\n print(\"Loading taxonomy data...\")\n taxonomy_df = pd.read_csv(cfg.taxonomy_csv)\n species_map = {row['primary_label']: idx for idx, row in taxonomy_df.iterrows()}\n \n print(\"Loading training data...\")\n train_df = pd.read_csv(cfg.train_csv)\n \n # Filter out missing files\n valid_files = []\n for idx, row in tqdm(train_df.iterrows(), total=len(train_df), desc=\"Validating files\"):\n file_path = os.path.join(cfg.train_audio_dir, row['filename'])\n if os.path.exists(file_path):\n valid_files.append(idx)\n \n train_df = train_df.iloc[valid_files].reset_index(drop=True)\n print(f\"Found {len(train_df)} valid files.\")\n \n # Debug mode: use smaller dataset\n if cfg.debug_mode:\n print(f\"Debug mode enabled. Using up to {cfg.debug_samples} samples.\")\n \n # Take a random sample\n sample_size = min(cfg.debug_samples, len(train_df))\n train_df = train_df.sample(sample_size, random_state=SEED)\n \n # Update species map to only include classes in our filtered dataset\n present_classes = set(train_df['primary_label'].unique())\n species_map = {species: idx for idx, species in enumerate(present_classes)}\n \n num_classes = len(species_map)\n print(f\"Working with {num_classes} classes in debug mode\")\n else:\n # Full training mode\n species_list = taxonomy_df['primary_label'].tolist()\n num_classes = len(species_list)\n print(f\"Number of classes: {num_classes}\")\n \n return train_df, taxonomy_df, species_map, num_classes\n\n# Audio processing functions\ndef audio_to_melspec(audio, cfg):\n \"\"\"Convert audio data to mel spectrogram with optimized settings\"\"\"\n # Handle NaN values\n if np.isnan(audio).any():\n audio = np.nan_to_num(audio)\n \n # Generate mel spectrogram\n mel_spec = librosa.feature.melspectrogram(\n y=audio,\n sr=cfg.sample_rate,\n n_fft=cfg.n_fft,\n hop_length=cfg.hop_length,\n n_mels=cfg.n_mels,\n fmin=cfg.fmin,\n fmax=cfg.fmax,\n power=2.0\n )\n \n # Convert to dB scale\n mel_spec_db = librosa.power_to_db(mel_spec, ref=np.max)\n \n # Faster normalization\n mel_spec_norm = (mel_spec_db + 80) / 80 # Typical dB range\n \n return np.clip(mel_spec_norm, 0, 1) # Clip to [0, 1]\n\n# Dataset class\nclass BirdCLEFDataset(Dataset):\n def __init__(self, df, species_map, cfg):\n self.df = df\n self.cfg = cfg\n self.species_map = species_map\n self.total_samples = len(df)\n print(f\"Dataset initialized with {self.total_samples} samples\")\n \n def __len__(self):\n return self.total_samples\n \n def __getitem__(self, idx):\n row = self.df.iloc[idx]\n file_path = os.path.join(self.cfg.train_audio_dir, row['filename'])\n \n try:\n # Faster audio loading\n audio, _ = librosa.load(\n file_path, \n sr=self.cfg.sample_rate, \n res_type='kaiser_fast', # Faster resampling\n duration=self.cfg.duration # Only load what we need\n )\n \n # Make sure we have exactly cfg.duration seconds\n expected_length = self.cfg.sample_rate * self.cfg.duration\n if len(audio) < expected_length:\n # Pad if shorter\n audio = np.pad(audio, (0, expected_length - len(audio)), mode='constant')\n elif len(audio) > expected_length:\n # Trim if longer\n audio = audio[:expected_length]\n \n # Convert to mel spectrogram\n mel_spec = audio_to_melspec(audio, self.cfg)\n \n # Resize to target dimensions\n mel_spec = cv2.resize(mel_spec, (self.cfg.img_size, self.cfg.img_size))\n \n # Add channel dimension and convert to tensor\n mel_spec = torch.tensor(mel_spec, dtype=torch.float32).unsqueeze(0) # Shape: [1, H, W]\n \n # Get label index\n label_idx = self.species_map[row['primary_label']]\n label = torch.zeros(len(self.species_map), dtype=torch.float32)\n label[label_idx] = 1.0\n \n return mel_spec, label\n \n except Exception as e:\n print(f\"Error loading file {file_path}: {e}\")\n # Return a zero spectrogram and label if there's an error\n mel_spec = torch.zeros((1, self.cfg.img_size, self.cfg.img_size), dtype=torch.float32)\n label = torch.zeros(len(self.species_map), dtype=torch.float32)\n return mel_spec, label\n\n# RegNet model with automatic feature detection\nclass BirdCLEFModel(nn.Module):\n def __init__(self, model_name, num_classes, in_channels=1):\n super().__init__()\n \n # Load the RegNet model\n try:\n self.backbone = timm.create_model(\n model_name,\n pretrained=True,\n in_chans=in_channels,\n num_classes=0 # Remove classifier head\n )\n \n # Get the feature dimension automatically\n with torch.no_grad():\n dummy_input = torch.zeros(1, in_channels, cfg.img_size, cfg.img_size)\n features = self.backbone(dummy_input)\n feature_dim = features.shape[1]\n print(f\"Feature dimension detected: {feature_dim}\")\n \n # Create a new classifier head\n self.classifier = nn.Sequential(\n nn.Dropout(0.2),\n nn.Linear(feature_dim, num_classes)\n )\n except Exception as e:\n print(f\"Error initializing RegNet: {e}\")\n print(\"Check available models with: print(timm.list_models('regnet*'))\")\n raise e\n \n def forward(self, x):\n features = self.backbone(x)\n output = self.classifier(features)\n return output\n\ndef train_one_epoch(model, dataloader, criterion, optimizer, scheduler, device, use_amp, epoch):\n model.train()\n \n running_loss = 0.0\n correct_pred = 0\n total_pred = 0\n processed_batches = 0\n \n scaler = torch.cuda.amp.GradScaler() if use_amp else None\n \n pbar = tqdm(enumerate(dataloader), total=len(dataloader), desc=f\"Epoch {epoch+1}/{cfg.epochs}\")\n \n for i, (inputs, targets) in pbar:\n # Print batch shape for debugging (first batch only)\n if i == 0 and epoch == 0:\n print(f\"Batch shapes - inputs: {inputs.shape}, targets: {targets.shape}\")\n \n inputs = inputs.to(device)\n targets = targets.to(device)\n \n # Zero the parameter gradients\n optimizer.zero_grad()\n \n try:\n if use_amp:\n # Forward pass with mixed precision\n with torch.cuda.amp.autocast():\n outputs = model(inputs)\n loss = criterion(outputs, targets)\n \n # Backward and optimize\n scaler.scale(loss).backward()\n scaler.step(optimizer)\n scaler.update()\n else:\n # Standard forward and backward pass\n outputs = model(inputs)\n loss = criterion(outputs, targets)\n loss.backward()\n optimizer.step()\n \n # Calculate accuracy metrics\n predicted = (outputs > 0.5).float()\n correct_pred += (predicted == targets).sum().item()\n total_pred += targets.numel()\n \n # Update running loss\n running_loss += loss.item() * inputs.size(0)\n processed_batches += 1\n \n # Update progress bar\n pbar.set_postfix({\n 'loss': f\"{loss.item():.4f}\",\n 'acc': f\"{(predicted == targets).sum().item() / targets.numel():.4f}\"\n })\n except Exception as e:\n print(f\"Error in batch {i}: {e}\")\n continue\n \n # Step the scheduler\n if scheduler is not None:\n scheduler.step()\n \n # Calculate epoch metrics\n epoch_loss = running_loss / len(dataloader.dataset) if processed_batches > 0 else float('inf')\n epoch_acc = correct_pred / total_pred if total_pred > 0 else 0\n \n print(f\"Epoch {epoch+1} - Loss: {epoch_loss:.4f}, Acc: {epoch_acc:.4f}\")\n return epoch_loss, epoch_acc\n\ndef train_full_model():\n # Load data\n train_df, taxonomy_df, species_map, num_classes = load_data()\n \n # Create dataset and dataloader\n train_dataset = BirdCLEFDataset(train_df, species_map, cfg)\n \n train_loader = DataLoader(\n train_dataset,\n batch_size=cfg.batch_size,\n shuffle=True,\n num_workers=cfg.num_workers,\n pin_memory=True if torch.cuda.is_available() else False,\n drop_last=True\n )\n \n # Initialize model\n print(f\"Initializing model: {cfg.model_name}...\")\n try:\n # List available RegNet models for reference\n available_regnets = [m for m in timm.list_models() if 'regnet' in m.lower()]\n print(f\"Available RegNet models: {available_regnets[:5]}... (showing first 5)\")\n \n # Initialize RegNet model\n model = BirdCLEFModel(cfg.model_name, num_classes)\n except Exception as e:\n print(f\"Error initializing RegNet: {e}\")\n raise e\n \n # Move model to device\n model = model.to(cfg.device)\n print(f\"Model moved to {cfg.device}\")\n \n # Set up loss, optimizer and scheduler\n criterion = nn.BCEWithLogitsLoss()\n optimizer = optim.AdamW(model.parameters(), lr=cfg.learning_rate, weight_decay=cfg.weight_decay)\n scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=cfg.epochs)\n \n # Training loop\n train_losses = []\n train_accs = []\n \n print(f\"Starting full training for {cfg.epochs} epochs...\")\n \n for epoch in range(cfg.epochs):\n # Train\n epoch_start_time = time.time()\n train_loss, train_acc = train_one_epoch(\n model, train_loader, criterion, optimizer, scheduler, \n cfg.device, cfg.use_amp, epoch\n )\n epoch_time = time.time() - epoch_start_time\n \n # Print epoch results\n print(f\"Epoch {epoch+1} completed in {epoch_time:.1f}s\")\n print(f\"Train Loss: {train_loss:.4f}, Train Acc: {train_acc:.4f}\")\n \n # Track metrics\n train_losses.append(train_loss)\n train_accs.append(train_acc)\n \n # Save model checkpoint\n torch.save({\n 'epoch': epoch,\n 'model_state_dict': model.state_dict(),\n 'optimizer_state_dict': optimizer.state_dict(),\n 'scheduler_state_dict': scheduler.state_dict() if scheduler else None,\n 'loss': train_loss,\n 'accuracy': train_acc,\n 'species_map': species_map, # Save mapping for inference\n }, os.path.join(cfg.model_save_path, f\"model_epoch_{epoch+1}.pth\"))\n \n # Save as best model too (for compatibility with inference code)\n if epoch == cfg.epochs - 1 or (epoch > 0 and train_loss < min(train_losses[:-1])):\n torch.save({\n 'epoch': epoch,\n 'model_state_dict': model.state_dict(),\n 'optimizer_state_dict': optimizer.state_dict(),\n 'scheduler_state_dict': scheduler.state_dict() if scheduler else None,\n 'loss': train_loss,\n 'accuracy': train_acc,\n 'species_map': species_map, # Save mapping for inference\n }, os.path.join(cfg.model_save_path, \"best_model.pth\"))\n print(f\"Saved best model at epoch {epoch+1}\")\n \n # Clean up memory\n gc.collect()\n if torch.cuda.is_available():\n torch.cuda.empty_cache()\n \n # Plot training curves\n plt.figure(figsize=(12, 5))\n \n plt.subplot(1, 2, 1)\n plt.plot(train_losses, label='Train Loss')\n plt.xlabel('Epoch')\n plt.ylabel('Loss')\n plt.legend()\n plt.title('Loss Curve')\n \n plt.subplot(1, 2, 2)\n plt.plot(train_accs, label='Train Accuracy')\n plt.xlabel('Epoch')\n plt.ylabel('Accuracy')\n plt.legend()\n plt.title('Accuracy Curve')\n \n plt.tight_layout()\n plt.savefig('training_curves.png')\n plt.show()\n \n # Save final model\n torch.save({\n 'epoch': cfg.epochs-1,\n 'model_state_dict': model.state_dict(),\n 'species_map': species_map, # Save mapping for inference\n }, os.path.join(cfg.model_save_path, \"final_model.pth\"))\n \n print(\"Saved final model.\")\n \n return model, train_losses, train_accs\n\nif __name__ == \"__main__\":\n start_time = time.time()\n print(f\"Using device: {cfg.device}\")\n try:\n model, train_losses, train_accs = train_full_model()\n print(f\"Training completed in {(time.time() - start_time)/60:.2f} minutes\")\n except Exception as e:\n print(f\"Error during training: {e}\")\n import traceback\n traceback.print_exc()"}],"metadata":{"kernelspec":{"display_name":"Python 3 (ipykernel)","language":"python","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.12.7"}},"nbformat":4,"nbformat_minor":5}