{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"gpu","dataSources":[],"dockerImageVersionId":31154,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# This Python 3 environment comes with many helpful analytics libraries installed\n# It is defined by the kaggle/python Docker image: https://github.com/kaggle/docker-python\n# For example, here's several helpful packages to load\n\nimport numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\n\n# Input data files are available in the read-only \"../input/\" directory\n# For example, running this (by clicking run or pressing Shift+Enter) will list all files under the input directory\n\nimport os\nfor dirname, _, filenames in os.walk('/kaggle/input'):\n for filename in filenames:\n print(os.path.join(dirname, filename))\n\n# You can write up to 20GB to the current directory (/kaggle/working/) that gets preserved as output when you create a version using \"Save & Run All\" \n# You can also write temporary files to /kaggle/temp/, but they won't be saved outside of the current session","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2025-10-07T18:50:11.214761Z","iopub.execute_input":"2025-10-07T18:50:11.215070Z","iopub.status.idle":"2025-10-07T18:50:11.220692Z","shell.execute_reply.started":"2025-10-07T18:50:11.215047Z","shell.execute_reply":"2025-10-07T18:50:11.219652Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\"\"\"\nRetinal Disease Progression Forecasting Using Longitudinal Fundus Images\n---------------------------------------------------------------------------\nA temporal vision model with time-aware attention, temporal convolution,\nand contrastive progression loss for diabetic retinopathy progression prediction.\n\nDatasets: EyePACS, DDR, IDRiD with time-stamped patient records\n\"\"\"\n\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\nfrom torchvision import transforms, models\nimport cv2\nfrom PIL import Image\nimport os\nfrom pathlib import Path\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import accuracy_score, f1_score, roc_auc_score\nimport matplotlib.pyplot as plt\nfrom tqdm import tqdm\nimport warnings\nwarnings.filterwarnings('ignore')\n\n# Set device\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nprint(f\"Using device: {device}\")\n\n# ============================================================================\n# 1. DATA PREPROCESSING AND LOADING\n# ============================================================================\n\nclass RetinalTimeSeriesDataset(Dataset):\n \"\"\"Dataset for longitudinal retinal images with temporal information\"\"\"\n \n def __init__(self, df, img_dir, max_sequence_length=5, transform=None):\n \"\"\"\n Args:\n df: DataFrame with columns ['patient_id', 'image_path', 'timestamp', 'severity_level']\n img_dir: Root directory containing images\n max_sequence_length: Maximum number of images in sequence\n transform: Image transformations\n \"\"\"\n self.df = df\n self.img_dir = img_dir\n self.max_seq_len = max_sequence_length\n self.transform = transform\n \n # Group by patient_id and sort by timestamp\n self.patient_groups = df.groupby('patient_id').apply(\n lambda x: x.sort_values('timestamp')\n ).reset_index(drop=True)\n \n self.patient_ids = df['patient_id'].unique()\n \n def __len__(self):\n return len(self.patient_ids)\n \n def __getitem__(self, idx):\n patient_id = self.patient_ids[idx]\n patient_data = self.patient_groups[\n self.patient_groups['patient_id'] == patient_id\n ]\n \n images = []\n timestamps = []\n severity_levels = []\n \n for _, row in patient_data.iterrows():\n img_path = os.path.join(self.img_dir, row['image_path'])\n \n # Load and preprocess image\n if os.path.exists(img_path):\n img = Image.open(img_path).convert('RGB')\n else:\n # Create dummy image if file doesn't exist (for demo purposes)\n img = Image.new('RGB', (224, 224), color='black')\n \n if self.transform:\n img = self.transform(img)\n \n images.append(img)\n timestamps.append(row['timestamp'])\n severity_levels.append(row['severity_level'])\n \n # Pad or truncate sequence\n seq_len = len(images)\n if seq_len < self.max_seq_len:\n # Pad with zeros\n padding = self.max_seq_len - seq_len\n images += [torch.zeros_like(images[0]) for _ in range(padding)]\n timestamps += [0] * padding\n severity_levels += [0] * padding\n else:\n # Take last max_seq_len images\n images = images[-self.max_seq_len:]\n timestamps = timestamps[-self.max_seq_len:]\n severity_levels = severity_levels[-self.max_seq_len:]\n \n images = torch.stack(images)\n timestamps = torch.tensor(timestamps, dtype=torch.float32)\n severity_levels = torch.tensor(severity_levels, dtype=torch.long)\n \n # Mask for valid positions\n mask = torch.zeros(self.max_seq_len, dtype=torch.bool)\n mask[:seq_len] = 1\n \n return {\n 'images': images,\n 'timestamps': timestamps,\n 'severity_levels': severity_levels,\n 'mask': mask,\n 'seq_len': seq_len\n }\n\n\n# ============================================================================\n# 2. TEMPORAL VISION MODEL ARCHITECTURE\n# ============================================================================\n\nclass TimeAwareAttention(nn.Module):\n \"\"\"Time-aware attention mechanism that considers temporal relationships\"\"\"\n \n def __init__(self, hidden_dim, num_heads=8):\n super().__init__()\n self.hidden_dim = hidden_dim\n self.num_heads = num_heads\n self.head_dim = hidden_dim // num_heads\n \n self.qkv = nn.Linear(hidden_dim, hidden_dim * 3)\n self.time_embed = nn.Linear(1, hidden_dim)\n self.out_proj = nn.Linear(hidden_dim, hidden_dim)\n \n def forward(self, x, timestamps, mask=None):\n \"\"\"\n Args:\n x: (batch, seq_len, hidden_dim)\n timestamps: (batch, seq_len)\n mask: (batch, seq_len) - True for valid positions\n \"\"\"\n B, T, C = x.shape\n \n # Generate Q, K, V\n qkv = self.qkv(x).reshape(B, T, 3, self.num_heads, self.head_dim)\n qkv = qkv.permute(2, 0, 3, 1, 4) # (3, B, heads, T, head_dim)\n q, k, v = qkv[0], qkv[1], qkv[2]\n \n # Time embeddings\n time_emb = self.time_embed(timestamps.unsqueeze(-1)) # (B, T, C)\n time_emb = time_emb.reshape(B, T, self.num_heads, self.head_dim)\n time_emb = time_emb.permute(0, 2, 1, 3) # (B, heads, T, head_dim)\n \n # Attention with temporal bias\n attn = (q @ k.transpose(-2, -1)) / np.sqrt(self.head_dim)\n time_bias = (time_emb @ time_emb.transpose(-2, -1)) / np.sqrt(self.head_dim)\n attn = attn + 0.1 * time_bias # Scale temporal bias\n \n # Apply mask\n if mask is not None:\n mask = mask.unsqueeze(1).unsqueeze(2) # (B, 1, 1, T)\n attn = attn.masked_fill(~mask, float('-inf'))\n \n attn = F.softmax(attn, dim=-1)\n out = attn @ v # (B, heads, T, head_dim)\n \n out = out.permute(0, 2, 1, 3).reshape(B, T, C)\n out = self.out_proj(out)\n \n return out\n\n\nclass TemporalConvBlock(nn.Module):\n \"\"\"1D temporal convolution for capturing progression patterns\"\"\"\n \n def __init__(self, in_channels, out_channels, kernel_size=3):\n super().__init__()\n self.conv = nn.Conv1d(in_channels, out_channels, kernel_size, \n padding=kernel_size//2)\n self.bn = nn.BatchNorm1d(out_channels)\n self.relu = nn.ReLU(inplace=True)\n \n def forward(self, x):\n \"\"\"\n Args:\n x: (batch, seq_len, channels)\n Returns:\n (batch, seq_len, channels)\n \"\"\"\n x = x.permute(0, 2, 1) # (B, C, T)\n x = self.conv(x)\n x = self.bn(x)\n x = self.relu(x)\n x = x.permute(0, 2, 1) # (B, T, C)\n return x\n\n\nclass ProgressionAwareEncoder(nn.Module):\n \"\"\"Encoder with progression-aware embeddings\"\"\"\n \n def __init__(self, feature_dim=512, hidden_dim=256, num_layers=3):\n super().__init__()\n \n # Feature encoder (ResNet backbone)\n resnet = models.resnet34(pretrained=True)\n self.backbone = nn.Sequential(*list(resnet.children())[:-1])\n self.feature_proj = nn.Linear(512, feature_dim)\n \n # Temporal processing\n self.temporal_convs = nn.ModuleList([\n TemporalConvBlock(feature_dim, feature_dim)\n for _ in range(num_layers)\n ])\n \n # Time-aware attention layers\n self.attention_layers = nn.ModuleList([\n TimeAwareAttention(feature_dim, num_heads=8)\n for _ in range(num_layers)\n ])\n \n self.norm = nn.LayerNorm(feature_dim)\n \n def forward(self, images, timestamps, mask=None):\n \"\"\"\n Args:\n images: (batch, seq_len, C, H, W)\n timestamps: (batch, seq_len)\n mask: (batch, seq_len)\n \"\"\"\n B, T, C, H, W = images.shape\n \n # Encode each image\n images = images.reshape(B * T, C, H, W)\n features = self.backbone(images) # (B*T, 512, 1, 1)\n features = features.squeeze(-1).squeeze(-1) # (B*T, 512)\n features = self.feature_proj(features) # (B*T, feature_dim)\n features = features.reshape(B, T, -1) # (B, T, feature_dim)\n \n # Temporal processing\n for conv, attn in zip(self.temporal_convs, self.attention_layers):\n # Temporal convolution\n features_conv = conv(features)\n \n # Time-aware attention\n features_attn = attn(features, timestamps, mask)\n \n # Residual connection\n features = features + features_conv + features_attn\n features = self.norm(features)\n \n return features\n\n\nclass ProgressionForecaster(nn.Module):\n \"\"\"Complete model for disease progression forecasting\"\"\"\n \n def __init__(self, num_classes=5, feature_dim=512, hidden_dim=256):\n super().__init__()\n \n self.encoder = ProgressionAwareEncoder(feature_dim, hidden_dim)\n \n # Progression predictor\n self.progression_head = nn.Sequential(\n nn.Linear(feature_dim, hidden_dim),\n nn.ReLU(),\n nn.Dropout(0.3),\n nn.Linear(hidden_dim, num_classes)\n )\n \n # Embedding projector for contrastive learning\n self.projection_head = nn.Sequential(\n nn.Linear(feature_dim, hidden_dim),\n nn.ReLU(),\n nn.Linear(hidden_dim, 128)\n )\n \n def forward(self, images, timestamps, mask=None):\n # Encode temporal sequence\n features = self.encoder(images, timestamps, mask)\n \n # Use last valid feature for prediction\n if mask is not None:\n # Get last valid index for each sequence\n seq_lengths = mask.sum(dim=1)\n batch_indices = torch.arange(features.size(0), device=features.device)\n last_features = features[batch_indices, seq_lengths - 1]\n else:\n last_features = features[:, -1]\n \n # Predict progression\n predictions = self.progression_head(last_features)\n \n # Get embeddings for contrastive learning\n embeddings = self.projection_head(features)\n \n return predictions, embeddings, features\n\n\n# ============================================================================\n# 3. CONTRASTIVE PROGRESSION LOSS\n# ============================================================================\n\nclass ContrastiveProgressionLoss(nn.Module):\n \"\"\"Contrastive loss that enforces progression-aware embeddings\"\"\"\n \n def __init__(self, temperature=0.07, margin=0.5):\n super().__init__()\n self.temperature = temperature\n self.margin = margin\n \n def forward(self, embeddings, severity_levels, mask):\n \"\"\"\n Args:\n embeddings: (batch, seq_len, embed_dim)\n severity_levels: (batch, seq_len)\n mask: (batch, seq_len)\n \"\"\"\n B, T, D = embeddings.shape\n \n # Normalize embeddings\n embeddings = F.normalize(embeddings, dim=-1)\n \n loss = 0.0\n count = 0\n \n for b in range(B):\n valid_len = mask[b].sum().item()\n if valid_len < 2:\n continue\n \n # Get valid embeddings and labels\n embs = embeddings[b, :valid_len] # (T', D)\n severities = severity_levels[b, :valid_len] # (T',)\n \n # Compute pairwise similarities\n sim_matrix = embs @ embs.T # (T', T')\n \n # Create progression labels (1 if severity increases, 0 otherwise)\n severity_diff = severities.unsqueeze(1) - severities.unsqueeze(0)\n progression_labels = (severity_diff > 0).float()\n \n # Contrastive loss: similar embeddings for progression pairs\n for i in range(valid_len - 1):\n for j in range(i + 1, valid_len):\n if progression_labels[i, j] > 0:\n # Progression pair: maximize similarity\n loss += F.relu(self.margin - sim_matrix[i, j])\n else:\n # Non-progression pair: push apart if too similar\n loss += F.relu(sim_matrix[i, j] - self.margin)\n count += 1\n \n return loss / max(count, 1)\n\n\n# ============================================================================\n# 4. TRAINING PIPELINE\n# ============================================================================\n\ndef train_epoch(model, dataloader, optimizer, device, alpha=0.5):\n \"\"\"Train for one epoch\"\"\"\n model.train()\n total_loss = 0\n ce_loss_fn = nn.CrossEntropyLoss()\n contrast_loss_fn = ContrastiveProgressionLoss()\n \n pbar = tqdm(dataloader, desc='Training')\n for batch in pbar:\n images = batch['images'].to(device)\n timestamps = batch['timestamps'].to(device)\n severity_levels = batch['severity_levels'].to(device)\n mask = batch['mask'].to(device)\n \n optimizer.zero_grad()\n \n # Forward pass\n predictions, embeddings, _ = model(images, timestamps, mask)\n \n # Classification loss (on last prediction)\n batch_indices = torch.arange(predictions.size(0), device=device)\n seq_lengths = mask.sum(dim=1)\n target = severity_levels[batch_indices, seq_lengths - 1]\n ce_loss = ce_loss_fn(predictions, target)\n \n # Contrastive progression loss\n contrast_loss = contrast_loss_fn(embeddings, severity_levels, mask)\n \n # Combined loss\n loss = ce_loss + alpha * contrast_loss\n \n loss.backward()\n optimizer.step()\n \n total_loss += loss.item()\n pbar.set_postfix({'loss': loss.item()})\n \n return total_loss / len(dataloader)\n\n\ndef evaluate(model, dataloader, device):\n \"\"\"Evaluate model\"\"\"\n model.eval()\n all_preds = []\n all_targets = []\n \n with torch.no_grad():\n for batch in tqdm(dataloader, desc='Evaluating'):\n images = batch['images'].to(device)\n timestamps = batch['timestamps'].to(device)\n severity_levels = batch['severity_levels'].to(device)\n mask = batch['mask'].to(device)\n \n predictions, _, _ = model(images, timestamps, mask)\n \n # Get targets (last severity level)\n batch_indices = torch.arange(predictions.size(0), device=device)\n seq_lengths = mask.sum(dim=1)\n targets = severity_levels[batch_indices, seq_lengths - 1]\n \n preds = predictions.argmax(dim=1)\n \n all_preds.extend(preds.cpu().numpy())\n all_targets.extend(targets.cpu().numpy())\n \n all_preds = np.array(all_preds)\n all_targets = np.array(all_targets)\n \n accuracy = accuracy_score(all_targets, all_preds)\n f1 = f1_score(all_targets, all_preds, average='weighted')\n \n return accuracy, f1\n\n\n# ============================================================================\n# 5. DEMO DATA GENERATION (FOR KAGGLE TESTING)\n# ============================================================================\n\ndef generate_demo_data(num_patients=100, output_dir='./demo_data'):\n \"\"\"Generate synthetic demo data for testing\"\"\"\n os.makedirs(output_dir, exist_ok=True)\n os.makedirs(os.path.join(output_dir, 'images'), exist_ok=True)\n \n data = []\n \n for patient_id in range(num_patients):\n num_visits = np.random.randint(2, 6)\n base_severity = np.random.randint(0, 3)\n \n for visit in range(num_visits):\n # Simulate progression\n severity = min(base_severity + visit // 2, 4)\n timestamp = visit * 30 # Days between visits\n \n # Create dummy image\n img_name = f'patient_{patient_id}_visit_{visit}.jpg'\n img_path = os.path.join(output_dir, 'images', img_name)\n \n # Generate random fundus-like image\n img = np.random.randint(0, 255, (224, 224, 3), dtype=np.uint8)\n cv2.imwrite(img_path, img)\n \n data.append({\n 'patient_id': patient_id,\n 'image_path': f'images/{img_name}',\n 'timestamp': timestamp,\n 'severity_level': severity\n })\n \n df = pd.DataFrame(data)\n df.to_csv(os.path.join(output_dir, 'metadata.csv'), index=False)\n \n return df, output_dir\n\n\n# ============================================================================\n# 6. MAIN EXECUTION\n# ============================================================================\n\ndef main():\n # Configuration\n CONFIG = {\n 'batch_size': 8,\n 'num_epochs': 20,\n 'learning_rate': 1e-4,\n 'num_classes': 5,\n 'max_seq_length': 5,\n 'feature_dim': 512,\n 'hidden_dim': 256,\n 'alpha': 0.5 # Weight for contrastive loss\n }\n \n print(\"=\"*70)\n print(\"Retinal Disease Progression Forecasting\")\n print(\"=\"*70)\n \n # Generate demo data\n print(\"\\n1. Generating demo data...\")\n df, data_dir = generate_demo_data(num_patients=100)\n print(f\"Generated data for {len(df)} records from {df['patient_id'].nunique()} patients\")\n \n # Split data\n train_patients, val_patients = train_test_split(\n df['patient_id'].unique(), test_size=0.2, random_state=42\n )\n train_df = df[df['patient_id'].isin(train_patients)]\n val_df = df[df['patient_id'].isin(val_patients)]\n \n print(f\"Train: {len(train_df)} records, Val: {len(val_df)} records\")\n \n # Data transforms\n transform = transforms.Compose([\n transforms.Resize((224, 224)),\n transforms.ToTensor(),\n transforms.Normalize(mean=[0.485, 0.456, 0.406], \n std=[0.229, 0.224, 0.225])\n ])\n \n # Create datasets\n print(\"\\n2. Creating datasets...\")\n train_dataset = RetinalTimeSeriesDataset(\n train_df, data_dir, \n max_sequence_length=CONFIG['max_seq_length'],\n transform=transform\n )\n val_dataset = RetinalTimeSeriesDataset(\n val_df, data_dir,\n max_sequence_length=CONFIG['max_seq_length'],\n transform=transform\n )\n \n train_loader = DataLoader(train_dataset, batch_size=CONFIG['batch_size'], \n shuffle=True, num_workers=2)\n val_loader = DataLoader(val_dataset, batch_size=CONFIG['batch_size'], \n shuffle=False, num_workers=2)\n \n # Create model\n print(\"\\n3. Initializing model...\")\n model = ProgressionForecaster(\n num_classes=CONFIG['num_classes'],\n feature_dim=CONFIG['feature_dim'],\n hidden_dim=CONFIG['hidden_dim']\n ).to(device)\n \n print(f\"Model parameters: {sum(p.numel() for p in model.parameters()):,}\")\n \n # Optimizer\n optimizer = torch.optim.AdamW(model.parameters(), lr=CONFIG['learning_rate'])\n \n # Training loop\n print(\"\\n4. Training model...\")\n best_f1 = 0\n history = {'train_loss': [], 'val_acc': [], 'val_f1': []}\n \n for epoch in range(CONFIG['num_epochs']):\n print(f\"\\nEpoch {epoch+1}/{CONFIG['num_epochs']}\")\n \n train_loss = train_epoch(model, train_loader, optimizer, device, \n alpha=CONFIG['alpha'])\n val_acc, val_f1 = evaluate(model, val_loader, device)\n \n history['train_loss'].append(train_loss)\n history['val_acc'].append(val_acc)\n history['val_f1'].append(val_f1)\n \n print(f\"Train Loss: {train_loss:.4f}\")\n print(f\"Val Accuracy: {val_acc:.4f}, Val F1: {val_f1:.4f}\")\n \n if val_f1 > best_f1:\n best_f1 = val_f1\n torch.save(model.state_dict(), 'best_model.pth')\n print(\"✓ Saved best model\")\n \n # Plot results\n print(\"\\n5. Plotting results...\")\n fig, axes = plt.subplots(1, 3, figsize=(15, 4))\n \n axes[0].plot(history['train_loss'])\n axes[0].set_title('Training Loss')\n axes[0].set_xlabel('Epoch')\n axes[0].set_ylabel('Loss')\n axes[0].grid(True)\n \n axes[1].plot(history['val_acc'])\n axes[1].set_title('Validation Accuracy')\n axes[1].set_xlabel('Epoch')\n axes[1].set_ylabel('Accuracy')\n axes[1].grid(True)\n \n axes[2].plot(history['val_f1'])\n axes[2].set_title('Validation F1 Score')\n axes[2].set_xlabel('Epoch')\n axes[2].set_ylabel('F1 Score')\n axes[2].grid(True)\n \n plt.tight_layout()\n plt.savefig('training_results.png', dpi=150, bbox_inches='tight')\n print(\"Saved training plots to 'training_results.png'\")\n \n print(\"\\n\" + \"=\"*70)\n print(f\"Training Complete! Best F1 Score: {best_f1:.4f}\")\n print(\"=\"*70)\n \n\n# 1. Print model architecture summary\n print_model_summary(model)\n\n# 2. Detailed evaluation on validation set\n eval_results = detailed_evaluation(\n model, val_loader, device, \n num_classes=5,\n class_names=['No DR', 'Mild', 'Moderate', 'Severe', 'Proliferative']\n )\n\n# 3. Analyze progression patterns\n progression_df = analyze_progression_patterns(model, val_loader, device)\n\n# 4. Print final summary\n print(\"\\n\" + \"=\"*80)\n print(\"FINAL MODEL PERFORMANCE SUMMARY\")\n print(\"=\"*80)\n print(f\"Best Validation F1 Score: {best_f1:.4f}\")\n print(f\"Final Validation Accuracy: {eval_results['accuracy']:.4f}\")\n print(f\"Final Validation Precision: {eval_results['precision']:.4f}\")\n print(f\"Final Validation Recall: {eval_results['recall']:.4f}\")\n print(\"=\"*80)\n\nif __name__ == '__main__':\n main()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-07T18:50:11.798165Z","iopub.execute_input":"2025-10-07T18:50:11.798475Z","iopub.status.idle":"2025-10-07T18:51:01.173458Z","shell.execute_reply.started":"2025-10-07T18:50:11.798454Z","shell.execute_reply":"2025-10-07T18:51:01.172406Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# modified_retinal_forecaster.py\n\"\"\"\nModified script:\n- Fixes dataset masking/truncation bug\n- Adds reproducibility (seeds)\n- Uses ResNet18 backbone for speed\n- Normalizes timestamps\n- Makes demo images visually correlated with severity (so model can learn)\n- Adds scheduler, stronger optimizer, and safer pretrained handling\n- Removes undefined helper calls at the end\n- Keeps: time-aware attention, temporal convs, contrastive progression loss\n\"\"\"\n\nimport random\nimport numpy as np\nimport pandas as pd\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\nfrom torchvision import transforms, models\nimport cv2\nfrom PIL import Image\nimport os\nfrom pathlib import Path\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score\nimport matplotlib.pyplot as plt\nfrom tqdm import tqdm\nimport warnings\nwarnings.filterwarnings('ignore')\n\n# reproducibility\ndef set_seed(seed=42):\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed_all(seed)\n\nset_seed(42)\n\n# Set device\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nprint(f\"Using device: {device}\")\n\n# ============================================================================\n# 1. DATA PREPROCESSING AND LOADING (fixed)\n# ============================================================================\n\nclass RetinalTimeSeriesDataset(Dataset):\n \"\"\"Dataset for longitudinal retinal images with temporal information\"\"\"\n \n def __init__(self, df, img_dir, max_sequence_length=5, transform=None):\n \"\"\"\n Args:\n df: DataFrame with columns ['patient_id', 'image_path', 'timestamp', 'severity_level']\n img_dir: Root directory containing images\n max_sequence_length: Maximum number of images in sequence\n transform: Image transformations\n \"\"\"\n self.df = df.copy()\n self.img_dir = img_dir\n self.max_seq_len = max_sequence_length\n self.transform = transform\n \n # Group by patient_id and sort by timestamp, keep groups in list for quick access\n grouped = self.df.groupby('patient_id')\n self.patient_groups = {}\n for pid, g in grouped:\n g_sorted = g.sort_values('timestamp').reset_index(drop=True)\n self.patient_groups[pid] = g_sorted\n \n self.patient_ids = list(self.patient_groups.keys())\n \n def __len__(self):\n return len(self.patient_ids)\n \n def __getitem__(self, idx):\n patient_id = self.patient_ids[idx]\n patient_data = self.patient_groups[patient_id]\n \n images = []\n timestamps = []\n severity_levels = []\n \n # Original sequence length before truncation/padding\n original_len = len(patient_data)\n \n for _, row in patient_data.iterrows():\n img_path = os.path.join(self.img_dir, row['image_path'])\n \n # Load and preprocess image\n if os.path.exists(img_path):\n img = Image.open(img_path).convert('RGB')\n else:\n # Create dummy image if file doesn't exist (for demo purposes)\n img = Image.new('RGB', (224, 224), color='black')\n \n if self.transform:\n img = self.transform(img)\n \n images.append(img)\n timestamps.append(float(row['timestamp']))\n severity_levels.append(int(row['severity_level']))\n \n # Truncate to last max_seq_len visits (prefer recent)\n if len(images) > self.max_seq_len:\n images = images[-self.max_seq_len:]\n timestamps = timestamps[-self.max_seq_len:]\n severity_levels = severity_levels[-self.max_seq_len:]\n \n # After truncation: valid_len\n valid_len = len(images)\n \n # Pad if shorter\n if valid_len < self.max_seq_len:\n padding = self.max_seq_len - valid_len\n # if no images (shouldn't happen), create zero image\n pad_img = torch.zeros_like(images[0]) if images else torch.zeros(3,224,224)\n images += [pad_img for _ in range(padding)]\n timestamps += [0.0] * padding\n severity_levels += [0] * padding\n \n images = torch.stack(images) # (T, C, H, W)\n timestamps = torch.tensor(timestamps, dtype=torch.float32)\n severity_levels = torch.tensor(severity_levels, dtype=torch.long)\n \n # Mask for valid positions (after truncation/padding)\n mask = torch.zeros(self.max_seq_len, dtype=torch.bool)\n mask[:valid_len] = 1\n \n return {\n 'images': images, # (T, C, H, W)\n 'timestamps': timestamps, # (T,)\n 'severity_levels': severity_levels, # (T,)\n 'mask': mask,\n 'seq_len': valid_len\n }\n\n\n# ============================================================================\n# 2. TEMPORAL VISION MODEL ARCHITECTURE (minor adjustments)\n# ============================================================================\n\nclass TimeAwareAttention(nn.Module):\n \"\"\"Time-aware attention mechanism that considers temporal relationships\"\"\"\n \n def __init__(self, hidden_dim, num_heads=8):\n super().__init__()\n assert hidden_dim % num_heads == 0, \"hidden_dim must be divisible by num_heads\"\n self.hidden_dim = hidden_dim\n self.num_heads = num_heads\n self.head_dim = hidden_dim // num_heads\n \n self.qkv = nn.Linear(hidden_dim, hidden_dim * 3)\n self.time_embed = nn.Linear(1, hidden_dim)\n self.out_proj = nn.Linear(hidden_dim, hidden_dim)\n \n def forward(self, x, timestamps, mask=None):\n \"\"\"\n Args:\n x: (batch, seq_len, hidden_dim)\n timestamps: (batch, seq_len)\n mask: (batch, seq_len) - True for valid positions\n \"\"\"\n B, T, C = x.shape\n \n # Generate Q, K, V\n qkv = self.qkv(x).reshape(B, T, 3, self.num_heads, self.head_dim)\n qkv = qkv.permute(2, 0, 3, 1, 4) # (3, B, heads, T, head_dim)\n q, k, v = qkv[0], qkv[1], qkv[2]\n \n # Time embeddings (normalized timestamps)\n time_emb = self.time_embed(timestamps.unsqueeze(-1)) # (B, T, C)\n time_emb = time_emb.reshape(B, T, self.num_heads, self.head_dim)\n time_emb = time_emb.permute(0, 2, 1, 3) # (B, heads, T, head_dim)\n \n # Attention with temporal bias\n attn = (q @ k.transpose(-2, -1)) / np.sqrt(self.head_dim)\n time_bias = (time_emb @ time_emb.transpose(-2, -1)) / np.sqrt(self.head_dim)\n attn = attn + 0.1 * time_bias # Scale temporal bias\n \n # Apply mask to keys (mask indicates valid keys)\n if mask is not None:\n # mask: (B, T) -> make (B, 1, 1, T) so it masks last dim (keys)\n key_mask = mask.unsqueeze(1).unsqueeze(2) # (B,1,1,T)\n attn = attn.masked_fill(~key_mask, float('-inf'))\n \n attn = F.softmax(attn, dim=-1)\n out = attn @ v # (B, heads, T, head_dim)\n \n out = out.permute(0, 2, 1, 3).reshape(B, T, C)\n out = self.out_proj(out)\n \n return out\n\n\nclass TemporalConvBlock(nn.Module):\n \"\"\"1D temporal convolution for capturing progression patterns\"\"\"\n \n def __init__(self, in_channels, out_channels, kernel_size=3):\n super().__init__()\n self.conv = nn.Conv1d(in_channels, out_channels, kernel_size, \n padding=kernel_size//2)\n self.bn = nn.BatchNorm1d(out_channels)\n self.relu = nn.ReLU(inplace=True)\n \n def forward(self, x):\n \"\"\"\n Args:\n x: (batch, seq_len, channels)\n Returns:\n (batch, seq_len, channels)\n \"\"\"\n x = x.permute(0, 2, 1) # (B, C, T)\n x = self.conv(x)\n x = self.bn(x)\n x = self.relu(x)\n x = x.permute(0, 2, 1) # (B, T, C)\n return x\n\n\nclass ProgressionAwareEncoder(nn.Module):\n \"\"\"Encoder with progression-aware embeddings\"\"\"\n \n def __init__(self, feature_dim=256, hidden_dim=128, num_layers=2, use_pretrained=False):\n super().__init__()\n \n # Feature encoder (ResNet backbone) - use ResNet18 for speed\n # keep final avgpool removed, output dim = 512\n resnet = models.resnet18(pretrained=use_pretrained)\n self.backbone = nn.Sequential(*list(resnet.children())[:-1]) # (B,512,1,1)\n self.feature_proj = nn.Linear(512, feature_dim)\n \n # Temporal processing\n self.temporal_convs = nn.ModuleList([\n TemporalConvBlock(feature_dim, feature_dim)\n for _ in range(num_layers)\n ])\n \n # Time-aware attention layers\n self.attention_layers = nn.ModuleList([\n TimeAwareAttention(feature_dim, num_heads=8)\n for _ in range(num_layers)\n ])\n \n self.norm = nn.LayerNorm(feature_dim)\n \n def forward(self, images, timestamps, mask=None):\n \"\"\"\n Args:\n images: (batch, seq_len, C, H, W)\n timestamps: (batch, seq_len)\n mask: (batch, seq_len)\n \"\"\"\n B, T, C, H, W = images.shape\n \n # Encode each image\n images = images.reshape(B * T, C, H, W)\n features = self.backbone(images) # (B*T, 512, 1, 1)\n features = features.reshape(B*T, -1) # (B*T, 512)\n features = self.feature_proj(features) # (B*T, feature_dim)\n features = features.reshape(B, T, -1) # (B, T, feature_dim)\n \n # Temporal processing\n for conv, attn in zip(self.temporal_convs, self.attention_layers):\n # Temporal convolution\n features_conv = conv(features)\n \n # Time-aware attention\n features_attn = attn(features, timestamps, mask)\n \n # Residual connection\n features = features + features_conv + features_attn\n features = self.norm(features)\n \n return features\n\n\nclass ProgressionForecaster(nn.Module):\n \"\"\"Complete model for disease progression forecasting\"\"\"\n \n def __init__(self, num_classes=5, feature_dim=256, hidden_dim=128, use_pretrained_backbone=False):\n super().__init__()\n \n self.encoder = ProgressionAwareEncoder(feature_dim=feature_dim, hidden_dim=hidden_dim, use_pretrained=use_pretrained_backbone)\n \n # Progression predictor (operates on last valid feature)\n self.progression_head = nn.Sequential(\n nn.Linear(feature_dim, hidden_dim),\n nn.ReLU(),\n nn.Dropout(0.3),\n nn.Linear(hidden_dim, num_classes)\n )\n \n # Embedding projector for contrastive learning (applied per timestep)\n self.projection_head = nn.Sequential(\n nn.Linear(feature_dim, hidden_dim),\n nn.ReLU(),\n nn.Linear(hidden_dim, 128)\n )\n \n def forward(self, images, timestamps, mask=None):\n # Encode temporal sequence\n features = self.encoder(images, timestamps, mask) # (B, T, feat)\n \n # Use last valid feature for prediction\n if mask is not None:\n seq_lengths = mask.sum(dim=1)\n batch_indices = torch.arange(features.size(0), device=features.device)\n last_features = features[batch_indices, seq_lengths - 1]\n else:\n last_features = features[:, -1]\n \n # Predict progression\n predictions = self.progression_head(last_features)\n \n # Get embeddings for contrastive learning (per timestep)\n embeddings = self.projection_head(features) # (B, T, embed_dim)\n \n return predictions, embeddings, features\n\n\n# ============================================================================\n# 3. CONTRASTIVE PROGRESSION LOSS (unchanged conceptually but vectorized)\n# ============================================================================\n\nclass ContrastiveProgressionLoss(nn.Module):\n \"\"\"Contrastive loss that enforces progression-aware embeddings\"\"\"\n \n def __init__(self, margin=0.5):\n super().__init__()\n self.margin = margin\n \n def forward(self, embeddings, severity_levels, mask):\n \"\"\"\n Args:\n embeddings: (batch, seq_len, embed_dim)\n severity_levels: (batch, seq_len)\n mask: (batch, seq_len)\n \"\"\"\n B, T, D = embeddings.shape\n embeddings = F.normalize(embeddings, dim=-1)\n \n loss = embeddings.new_tensor(0.0)\n count = 0\n \n for b in range(B):\n valid_len = int(mask[b].sum().item())\n if valid_len < 2:\n continue\n embs = embeddings[b, :valid_len] # (L, D)\n severities = severity_levels[b, :valid_len] # (L,)\n \n sim = torch.matmul(embs, embs.T) # (L, L)\n # progression if later severity > earlier severity\n sev_diff = severities.unsqueeze(1) - severities.unsqueeze(0)\n prog_mask = (sev_diff > 0).float()\n \n # We only consider i 0:\n # want high similarity -> push margin up: loss = relu(margin - s)\n loss = loss + F.relu(self.margin - s)\n else:\n # want low similarity -> relu(s - margin)\n loss = loss + F.relu(s - self.margin)\n count += 1\n \n if count == 0:\n return torch.tensor(0.0, device=embeddings.device)\n return loss / count\n\n\n# ============================================================================\n# 4. TRAINING PIPELINE (fixes & improvements)\n# ============================================================================\n\ndef train_epoch(model, dataloader, optimizer, device, alpha=0.5):\n \"\"\"Train for one epoch\"\"\"\n model.train()\n total_loss = 0.0\n ce_loss_fn = nn.CrossEntropyLoss()\n contrast_loss_fn = ContrastiveProgressionLoss()\n \n pbar = tqdm(dataloader, desc='Training', leave=False)\n for batch in pbar:\n imgs = batch['images'].to(device) # (B, T, C, H, W)\n # Model expects images as (B, T, C, H, W) -> our dataset returns (T, C, H, W) per sample, dataloader stacks to (B, T, C, H, W)\n timestamps = batch['timestamps'].to(device) # (B, T)\n severity_levels = batch['severity_levels'].to(device)\n mask = batch['mask'].to(device)\n \n # Normalize timestamps (scale to years, avoid huge values)\n timestamps_norm = timestamps / 365.0\n \n optimizer.zero_grad()\n \n predictions, embeddings, _ = model(imgs, timestamps_norm, mask)\n \n # Classification loss (on last valid timestep)\n batch_indices = torch.arange(predictions.size(0), device=device)\n seq_lengths = mask.sum(dim=1)\n target = severity_levels[batch_indices, seq_lengths - 1]\n ce_loss = ce_loss_fn(predictions, target)\n \n # Contrastive progression loss\n contrast_loss = contrast_loss_fn(embeddings, severity_levels, mask)\n \n loss = ce_loss + alpha * contrast_loss\n \n loss.backward()\n optimizer.step()\n \n total_loss += loss.item()\n pbar.set_postfix({'loss': loss.item()})\n \n return total_loss / max(len(dataloader), 1)\n\n\ndef evaluate(model, dataloader, device):\n \"\"\"Evaluate model\"\"\"\n model.eval()\n all_preds = []\n all_targets = []\n \n with torch.no_grad():\n for batch in tqdm(dataloader, desc='Evaluating', leave=False):\n imgs = batch['images'].to(device)\n timestamps = batch['timestamps'].to(device)\n severity_levels = batch['severity_levels'].to(device)\n mask = batch['mask'].to(device)\n \n timestamps_norm = timestamps / 365.0\n \n predictions, _, _ = model(imgs, timestamps_norm, mask)\n \n batch_indices = torch.arange(predictions.size(0), device=device)\n seq_lengths = mask.sum(dim=1)\n targets = severity_levels[batch_indices, seq_lengths - 1]\n \n preds = predictions.argmax(dim=1)\n \n all_preds.extend(preds.cpu().numpy())\n all_targets.extend(targets.cpu().numpy())\n \n all_preds = np.array(all_preds)\n all_targets = np.array(all_targets)\n \n accuracy = accuracy_score(all_targets, all_preds)\n f1 = f1_score(all_targets, all_preds, average='weighted', zero_division=0)\n precision = precision_score(all_targets, all_preds, average='weighted', zero_division=0)\n recall = recall_score(all_targets, all_preds, average='weighted', zero_division=0)\n \n return {\n 'accuracy': accuracy,\n 'f1': f1,\n 'precision': precision,\n 'recall': recall\n }\n\n\n# ============================================================================\n# 5. DEMO DATA GENERATION (ENHANCED so model can learn visual + temporal cues)\n# ============================================================================\n\ndef generate_demo_data(num_patients=100, output_dir='./demo_data'):\n \"\"\"Generate synthetic demo data for testing with visual cues correlated to severity\"\"\"\n os.makedirs(output_dir, exist_ok=True)\n img_dir = os.path.join(output_dir, 'images')\n os.makedirs(img_dir, exist_ok=True)\n \n data = []\n \n for patient_id in range(num_patients):\n num_visits = np.random.randint(2, 6) # 2..5 visits\n base_severity = np.random.randint(0, 3) # 0,1,2\n \n for visit in range(num_visits):\n # deterministic-ish progression: every 2 visits severity increases by 1\n severity = int(min(base_severity + visit // 2, 4))\n timestamp = float(visit * 30) # days between visits\n \n # Create an image where intensity correlates with severity:\n # higher severity -> brighter red channel and a centered \"lesion\" circle\n img_name = f'patient_{patient_id}_visit_{visit}.jpg'\n img_path = os.path.join(img_dir, img_name)\n \n base_intensity = int(40 + severity * 50) # 40, 90, 140, 190, 240\n img = np.random.randint(0, 30, (224, 224, 3), dtype=np.uint8) # background noise\n img[..., 0] = np.clip(img[..., 0] + base_intensity, 0, 255) # red channel stronger with severity\n \n # add a circular lesion whose radius scales with severity\n center = (112 + np.random.randint(-10, 10), 112 + np.random.randint(-10, 10))\n radius = 6 + severity * 6\n color = (min(255, base_intensity + 20), 30, 30)\n cv2.circle(img, center, radius, color, -1)\n \n cv2.imwrite(img_path, img)\n \n data.append({\n 'patient_id': int(patient_id),\n 'image_path': f'images/{img_name}',\n 'timestamp': timestamp,\n 'severity_level': int(severity)\n })\n \n df = pd.DataFrame(data)\n df.to_csv(os.path.join(output_dir, 'metadata.csv'), index=False)\n \n return df, output_dir\n\n\n# ============================================================================\n# 6. MAIN EXECUTION (cleaned & runnable)\n# ============================================================================\n\ndef main():\n # Configuration - tuned for demo\n CONFIG = {\n 'batch_size': 16,\n 'num_epochs': 18,\n 'learning_rate': 3e-4,\n 'num_classes': 5,\n 'max_seq_length': 5,\n 'feature_dim': 256,\n 'hidden_dim': 128,\n 'alpha': 0.8, # weight for contrastive loss (higher helps sequence supervision)\n 'use_pretrained_backbone': False # set True if you have weights / internet\n }\n \n print(\"=\"*70)\n print(\"Retinal Disease Progression Forecasting (modified demo)\")\n print(\"=\"*70)\n \n # Generate demo data\n print(\"\\n1. Generating demo data...\")\n df, data_dir = generate_demo_data(num_patients=300, output_dir='./demo_data')\n n_patients = df['patient_id'].nunique()\n print(f\"Generated {len(df)} records from {n_patients} patients\")\n\n# ➕ ADD BELOW\n print(\"\\nShowing a few sample images from the generated dataset...\")\n sample_df = df.sample(9, random_state=42).reset_index(drop=True)\n fig, axes = plt.subplots(3, 3, figsize=(8, 8))\n for i, ax in enumerate(axes.flat):\n row = sample_df.iloc[i]\n img_path = os.path.join(data_dir, row['image_path'])\n img = cv2.imread(img_path)[..., ::-1] # BGR->RGB\n ax.imshow(img)\n ax.set_title(f\"PID:{row['patient_id']}\\nSeverity:{row['severity_level']}\", fontsize=8)\n ax.axis('off')\n plt.tight_layout()\n plt.savefig('sample_images.png', dpi=150, bbox_inches='tight')\n plt.show()\n print(\"Saved 9-sample grid to 'sample_images.png'\")\n# ➕ ADD ABOVE\n\n# Continue with splitting train/val...\n\n print(f\"Generated {len(df)} records from {n_patients} patients\")\n \n # Split patients (patient-wise split)\n train_patients, val_patients = train_test_split(\n df['patient_id'].unique(), test_size=0.20, random_state=42\n )\n train_df = df[df['patient_id'].isin(train_patients)].reset_index(drop=True)\n val_df = df[df['patient_id'].isin(val_patients)].reset_index(drop=True)\n \n print(f\"Train records: {len(train_df)}, Val records: {len(val_df)}\")\n \n # Data transforms\n transform = transforms.Compose([\n transforms.Resize((224, 224)),\n transforms.ToTensor(),\n transforms.Normalize(mean=[0.485, 0.456, 0.406], \n std=[0.229, 0.224, 0.225])\n ])\n \n # Create datasets\n print(\"\\n2. Creating datasets...\")\n train_dataset = RetinalTimeSeriesDataset(\n train_df, data_dir, \n max_sequence_length=CONFIG['max_seq_length'],\n transform=transform\n )\n val_dataset = RetinalTimeSeriesDataset(\n val_df, data_dir,\n max_sequence_length=CONFIG['max_seq_length'],\n transform=transform\n )\n \n train_loader = DataLoader(train_dataset, batch_size=CONFIG['batch_size'], \n shuffle=True, num_workers=2, pin_memory=True)\n val_loader = DataLoader(val_dataset, batch_size=CONFIG['batch_size'], \n shuffle=False, num_workers=2, pin_memory=True)\n \n # Create model\n print(\"\\n3. Initializing model...\")\n model = ProgressionForecaster(\n num_classes=CONFIG['num_classes'],\n feature_dim=CONFIG['feature_dim'],\n hidden_dim=CONFIG['hidden_dim'],\n use_pretrained_backbone=CONFIG['use_pretrained_backbone']\n ).to(device)\n \n total_params = sum(p.numel() for p in model.parameters())\n print(f\"Model parameters: {total_params:,}\")\n \n # Optimizer + scheduler\n optimizer = torch.optim.AdamW(model.parameters(), lr=CONFIG['learning_rate'], weight_decay=1e-4)\n scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=CONFIG['num_epochs'], eta_min=1e-6)\n \n # Training loop\n print(\"\\n4. Training model...\")\n best_f1 = 0.0\n history = {'train_loss': [], 'val_acc': [], 'val_f1': []}\n \n for epoch in range(CONFIG['num_epochs']):\n print(f\"\\nEpoch {epoch+1}/{CONFIG['num_epochs']}\")\n train_loss = train_epoch(model, train_loader, optimizer, device, alpha=CONFIG['alpha'])\n eval_res = evaluate(model, val_loader, device)\n scheduler.step()\n \n history['train_loss'].append(train_loss)\n history['val_acc'].append(eval_res['accuracy'])\n history['val_f1'].append(eval_res['f1'])\n \n print(f\"Train Loss: {train_loss:.4f}\")\n print(f\"Val Accuracy: {eval_res['accuracy']:.4f}, Val F1: {eval_res['f1']:.4f}\")\n print(f\"Precision: {eval_res['precision']:.4f}, Recall: {eval_res['recall']:.4f}\")\n \n if eval_res['f1'] > best_f1:\n best_f1 = eval_res['f1']\n torch.save(model.state_dict(), 'best_model.pth')\n print(\"✓ Saved best model\")\n \n # Plot results\n print(\"\\n5. Plotting results...\")\n fig, axes = plt.subplots(1, 3, figsize=(15, 4))\n axes[0].plot(history['train_loss'])\n axes[0].set_title('Training Loss')\n axes[0].set_xlabel('Epoch')\n axes[0].set_ylabel('Loss')\n axes[0].grid(True)\n \n axes[1].plot(history['val_acc'])\n axes[1].set_title('Validation Accuracy')\n axes[1].set_xlabel('Epoch')\n axes[1].set_ylabel('Accuracy')\n axes[1].grid(True)\n \n axes[2].plot(history['val_f1'])\n axes[2].set_title('Validation F1 Score')\n axes[2].set_xlabel('Epoch')\n axes[2].set_ylabel('F1 Score')\n axes[2].grid(True)\n \n plt.tight_layout()\n plt.savefig('training_results.png', dpi=150, bbox_inches='tight')\n print(\"Saved training plots to 'training_results.png'\")\n \n print(\"\\n\" + \"=\"*70)\n print(f\"Training Complete! Best Val F1 Score: {best_f1:.4f}\")\n print(\"=\"*70)\n \n # Final evaluation\n final_eval = evaluate(model, val_loader, device)\n print(\"\\nFinal evaluation on validation set:\")\n print(final_eval)\n \n\nif __name__ == '__main__':\n main()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-07T19:01:16.208603Z","iopub.execute_input":"2025-10-07T19:01:16.209335Z","iopub.status.idle":"2025-10-07T19:02:28.082091Z","shell.execute_reply.started":"2025-10-07T19:01:16.209304Z","shell.execute_reply":"2025-10-07T19:02:28.081068Z"}},"outputs":[],"execution_count":null}]}