{"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":"nvidiaTeslaT4","dataSources":[{"sourceId":4104,"databundleVersionId":46661,"sourceType":"competition"},{"sourceId":7866129,"sourceType":"datasetVersion","datasetId":4614938}],"dockerImageVersionId":31090,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"!nvidia-smi","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:37.903214Z","iopub.execute_input":"2025-08-01T11:32:37.903762Z","iopub.status.idle":"2025-08-01T11:32:38.139043Z","shell.execute_reply.started":"2025-08-01T11:32:37.903742Z","shell.execute_reply":"2025-08-01T11:32:38.138245Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Libraries","metadata":{}},{"cell_type":"code","source":"import pandas as pd\nfrom glob import glob\nimport matplotlib.pyplot as plt\nimport cv2\nimport tensorflow as tf\nimport numpy as np\nimport os\nfrom sklearn.model_selection import train_test_split\nfrom tensorflow.keras import layers, models, callbacks, mixed_precision\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:38.140247Z","iopub.execute_input":"2025-08-01T11:32:38.140549Z","iopub.status.idle":"2025-08-01T11:32:38.145122Z","shell.execute_reply.started":"2025-08-01T11:32:38.140513Z","shell.execute_reply":"2025-08-01T11:32:38.144401Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import wandb\nfrom kaggle_secrets import UserSecretsClient\nuser_secrets = UserSecretsClient()\nsecret_value_0 = user_secrets.get_secret(\"WANDB_API_KEY\")\n\nwandb.login(key=secret_value_0)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:38.146222Z","iopub.execute_input":"2025-08-01T11:32:38.146547Z","iopub.status.idle":"2025-08-01T11:32:38.303069Z","shell.execute_reply.started":"2025-08-01T11:32:38.146529Z","shell.execute_reply":"2025-08-01T11:32:38.302508Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# import tensorflow as tf\nprint(tf.__version__)\n# Check if a GPU is available\ngpus = tf.config.list_physical_devices('GPU')\nif gpus:\n    try:\n        # Set memory growth to prevent TensorFlow from allocating all memory at once\n        for gpu in gpus:\n            tf.config.experimental.set_memory_growth(gpu, True)\n        print(\"GPU is available and will be used.\")\n    except RuntimeError as e:\n        # Memory growth must be set before GPUs have been initialized\n        print(f\"Error: {e}\")\nelse:\n    print(\"GPU is not available, running on CPU.\")\n\ntf.test.is_gpu_available()\n\npolicy = mixed_precision.Policy('mixed_float16')\nmixed_precision.set_global_policy(policy)\nprint(\"Mixed precision enabled:\", policy)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:38.303898Z","iopub.execute_input":"2025-08-01T11:32:38.304426Z","iopub.status.idle":"2025-08-01T11:32:39.144272Z","shell.execute_reply.started":"2025-08-01T11:32:38.304401Z","shell.execute_reply":"2025-08-01T11:32:39.143492Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Imports","metadata":{}},{"cell_type":"code","source":"IMG_SIZE = 224\nBATCH_SIZE = 8\nEPOCHS = 3\nFROZEN_EPOCHS = 2\n\ntrain_csv_path = \"/kaggle/input/diabetic-retinopathy-detection/trainLabels.csv.zip\"\ntrain_images_dir = \"/kaggle/input/diabetic-retinopathy-train-unzipped/train/\"\ntest_images_dir = \"/kaggle/input/diabetic-retinopathy-test-unzipped/test/\"\nsubmission_csv_path = \"/kaggle/input/diabetic-retinopathy-detection/sampleSubmission.csv.zip\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:39.147000Z","iopub.execute_input":"2025-08-01T11:32:39.147419Z","iopub.status.idle":"2025-08-01T11:32:39.154592Z","shell.execute_reply.started":"2025-08-01T11:32:39.147401Z","shell.execute_reply":"2025-08-01T11:32:39.153818Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"img_number = 10016\n\ndf = pd.read_csv(train_csv_path)\ndf_train = df[:img_number].copy()\ndf_train.head()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:39.155357Z","iopub.execute_input":"2025-08-01T11:32:39.155582Z","iopub.status.idle":"2025-08-01T11:32:39.215505Z","shell.execute_reply.started":"2025-08-01T11:32:39.155556Z","shell.execute_reply":"2025-08-01T11:32:39.214911Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"**Train Test Split**","metadata":{}},{"cell_type":"code","source":"# Convert multiclass level to binary: 0 = No DR, 1 = DR Present\ndf_train[\"Label\"] = df_train[\"level\"].apply(lambda x: False if x == 0 else True)\n\ndf_train[\"filepath\"] = df_train[\"image\"].apply(lambda x: os.path.join(train_images_dir, f\"{x}.jpeg\"))\n\ntrain_df, val_df = train_test_split(df_train, test_size=0.4, stratify=df_train[\"Label\"], random_state=42)\nval_df, test_df = train_test_split(val_df, test_size=0.5, stratify=val_df[\"Label\"], random_state=42)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:39.216247Z","iopub.execute_input":"2025-08-01T11:32:39.216510Z","iopub.status.idle":"2025-08-01T11:32:39.253665Z","shell.execute_reply.started":"2025-08-01T11:32:39.216487Z","shell.execute_reply":"2025-08-01T11:32:39.253132Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"print(train_df.shape, val_df.shape, test_df.shape)\n# train_df.filepath[0]\ntrain_df","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:39.254298Z","iopub.execute_input":"2025-08-01T11:32:39.254528Z","iopub.status.idle":"2025-08-01T11:32:39.265236Z","shell.execute_reply.started":"2025-08-01T11:32:39.254512Z","shell.execute_reply":"2025-08-01T11:32:39.264522Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Visualization","metadata":{}},{"cell_type":"code","source":"import cv2\nimport matplotlib.pyplot as plt\n\nplt.figure(figsize=(15, 15))\n\n# Show first 25 images from train_df\nfor idx, (i, row) in enumerate(train_df.head(25).iterrows()):\n    plt.subplot(5, 5, idx + 1)\n    plt.xticks([])\n    plt.yticks([])\n    plt.grid(False)\n    \n    # Load and preprocess image\n    img_path = row['filepath']\n    img = cv2.imread(img_path)\n    \n    if img is None:\n        print(f\"Warning: Image not found at {img_path}\")\n        continue\n    \n    img = cv2.resize(img, (224, 224))\n    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n    \n    # Set label from binary column\n    label = row.get('binary_label', row.get('Label', None))\n    label_text = 'Yes' if label else 'No'\n    \n    plt.imshow(img)\n    plt.xlabel(label_text, fontsize=12, color='green' if label else 'red')\n\nplt.tight_layout()\nplt.suptitle(\"Sample Images with Binary DR Labels\", fontsize=20, y=1.02)\nplt.show()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:39.265999Z","iopub.execute_input":"2025-08-01T11:32:39.266268Z","iopub.status.idle":"2025-08-01T11:32:43.067598Z","shell.execute_reply.started":"2025-08-01T11:32:39.266240Z","shell.execute_reply":"2025-08-01T11:32:43.066669Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Evaluation Functions","metadata":{}},{"cell_type":"markdown","source":"## Results","metadata":{}},{"cell_type":"code","source":"# from sklearn.metrics import (\n#     accuracy_score, precision_score, recall_score, f1_score, roc_auc_score,\n#     average_precision_score, confusion_matrix, cohen_kappa_score,\n#     precision_recall_curve, roc_curve\n# )\n# import tensorflow as tf\n# import pandas as pd\n# import numpy as np\n# import os\n\n# def Results(model, model_name,test_set, save_csv=False):\n#     # Evaluate the model\n#     results = model.evaluate(test_set)\n#     print(\"Evaluation results: \", results)\n\n#     y_true = []\n#     y_pred_probs = []\n\n#     print(\"Loading....\", flush=True)\n\n#     for images, labels in test_set:\n#         y_true.extend(labels.numpy())\n#         y_pred_probs.extend(model.predict(images, verbose=0).flatten())\n\n#     y_true = tf.convert_to_tensor(y_true)\n#     y_pred_probs = tf.convert_to_tensor(y_pred_probs)\n#     y_pred = (y_pred_probs > 0.5).numpy().astype(\"int32\")\n\n#     acc = accuracy_score(y_true, y_pred)\n#     precision = precision_score(y_true, y_pred, zero_division=1)\n#     recall = recall_score(y_true, y_pred)\n#     f1 = f1_score(y_true, y_pred)\n#     roc_auc = roc_auc_score(y_true, y_pred_probs)\n#     prc_auc = average_precision_score(y_true, y_pred_probs)\n#     conf_matrix = confusion_matrix(y_true, y_pred)\n#     kappa = cohen_kappa_score(y_true, y_pred)\n#     tn, fp, fn, tp = conf_matrix.ravel()\n#     npv = tn / (tn + fn)\n\n#     print(f\"Accuracy: {acc}\")\n#     print(f\"Precision: {precision}\")\n#     print(f\"Recall: {recall}\")\n#     print(f\"F1 Score: {f1}\")\n#     print(f\"ROC AUC: {roc_auc}\")\n#     print(f\"PRC AUC: {prc_auc}\")\n#     print(f\"Confusion Matrix: \\n{conf_matrix}\")\n#     print(f\"Kappa Coefficient: {kappa}\")\n#     print(f\"NPV: {npv}\")\n\n#     if save_csv:\n#         # Compute ROC and Precision-Recall curves\n#         fpr, tpr, _ = roc_curve(y_true, y_pred_probs)\n#         prec, rec, _ = precision_recall_curve(y_true, y_pred_probs)\n\n#         # Prepare DataFrame\n#         max_len = max(len(fpr), len(prec))\n#         fpr = np.pad(fpr, (0, max_len - len(fpr)), 'constant', constant_values=np.nan)\n#         tpr = np.pad(tpr, (0, max_len - len(tpr)), 'constant', constant_values=np.nan)\n#         prec = np.pad(prec, (0, max_len - len(prec)), 'constant', constant_values=np.nan)\n#         rec = np.pad(rec, (0, max_len - len(rec)), 'constant', constant_values=np.nan)\n\n#         df = pd.DataFrame({\n#             'FPR': fpr,\n#             'TPR': tpr,\n#             'Precision': prec,\n#             'Recall': rec\n#         })\n\n\n#         file_name = model_name\n#         df.to_csv(file_name, index=False)\n#         print(f\"Saved metrics to: {file_name}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.068895Z","iopub.execute_input":"2025-08-01T11:32:43.069280Z","iopub.status.idle":"2025-08-01T11:32:43.074117Z","shell.execute_reply.started":"2025-08-01T11:32:43.069249Z","shell.execute_reply":"2025-08-01T11:32:43.073421Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from sklearn.metrics import (\n    accuracy_score, precision_score, recall_score, f1_score, roc_auc_score,\n    average_precision_score, confusion_matrix, cohen_kappa_score,\n    precision_recall_curve, roc_curve\n)\nimport tensorflow as tf\nimport pandas as pd\nimport numpy as np\nimport os\nimport wandb\nimport matplotlib.pyplot as plt\n\ndef Results(model, test_set, model_name=\"Model\", save_csv=False, wandb_log=False):\n    # Evaluate the model\n    results = model.evaluate(test_set)\n    print(\"Evaluation results: \", results)\n\n    y_true = []\n    y_pred_probs = []\n\n    print(\"Loading....\", flush=True)\n\n    for images, labels in test_set:\n        y_true.extend(labels.numpy())\n        y_pred_probs.extend(model.predict(images, verbose=0).flatten())\n\n    y_true = tf.convert_to_tensor(y_true)\n    y_pred_probs = tf.convert_to_tensor(y_pred_probs)\n    y_pred = (y_pred_probs > 0.5).numpy().astype(\"int32\")\n\n    acc = accuracy_score(y_true, y_pred)\n    precision = precision_score(y_true, y_pred, zero_division=1)\n    recall = recall_score(y_true, y_pred)\n    f1 = f1_score(y_true, y_pred)\n    roc_auc = roc_auc_score(y_true, y_pred_probs)\n    prc_auc = average_precision_score(y_true, y_pred_probs)\n    conf_matrix = confusion_matrix(y_true, y_pred)\n    kappa = cohen_kappa_score(y_true, y_pred)\n    tn, fp, fn, tp = conf_matrix.ravel()\n    npv = tn / (tn + fn)\n\n    # Print metrics\n    print(f\"Accuracy: {acc}\")\n    print(f\"Precision: {precision}\")\n    print(f\"Recall: {recall}\")\n    print(f\"F1 Score: {f1}\")\n    print(f\"ROC AUC: {roc_auc}\")\n    print(f\"PRC AUC: {prc_auc}\")\n    print(f\"Confusion Matrix: \\n{conf_matrix}\")\n    print(f\"Kappa Coefficient: {kappa}\")\n    print(f\"NPV: {npv}\")\n\n    # Log metrics to W&B\n    if wandb_log:\n        wandb.log({\n            \"Accuracy\": acc,\n            \"Precision\": precision,\n            \"Recall\": recall,\n            \"F1 Score\": f1,\n            \"ROC AUC\": roc_auc,\n            \"PRC AUC\": prc_auc,\n            \"Kappa\": kappa,\n            \"NPV\": npv\n        })\n\n    if save_csv or wandb_log:\n        # Compute ROC and Precision-Recall curves\n        fpr, tpr, _ = roc_curve(y_true, y_pred_probs)\n        prec, rec, _ = precision_recall_curve(y_true, y_pred_probs)\n\n        # Pad arrays to same length\n        max_len = max(len(fpr), len(prec))\n        fpr = np.pad(fpr, (0, max_len - len(fpr)), 'constant', constant_values=np.nan)\n        tpr = np.pad(tpr, (0, max_len - len(tpr)), 'constant', constant_values=np.nan)\n        prec = np.pad(prec, (0, max_len - len(prec)), 'constant', constant_values=np.nan)\n        rec = np.pad(rec, (0, max_len - len(rec)), 'constant', constant_values=np.nan)\n\n        df = pd.DataFrame({\n            'FPR': fpr,\n            'TPR': tpr,\n            'Precision': prec,\n            'Recall': rec\n        })\n\n        file_name = model_name+'.metrices.csv'\n        df.to_csv(file_name, index=False)\n        print(f\"Saved metrics to: {file_name}\")\n\n        if wandb_log:\n            wandb.save(file_name)\n\n            # Plot ROC Curve\n            plt.figure()\n            plt.plot(fpr, tpr, label=f'ROC AUC = {roc_auc:.2f}')\n            plt.xlabel(\"False Positive Rate\")\n            plt.ylabel(\"True Positive Rate\")\n            plt.title(\"ROC Curve\")\n            plt.legend()\n            wandb.log({\"ROC Curve\": wandb.Image(plt)})\n            plt.close()\n\n            # Plot PR Curve\n            plt.figure()\n            plt.plot(rec, prec, label=f'PRC AUC = {prc_auc:.2f}')\n            plt.xlabel(\"Recall\")\n            plt.ylabel(\"Precision\")\n            plt.title(\"Precision-Recall Curve\")\n            plt.legend()\n            wandb.log({\"PR Curve\": wandb.Image(plt)})\n            plt.close()\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.074883Z","iopub.execute_input":"2025-08-01T11:32:43.075081Z","iopub.status.idle":"2025-08-01T11:32:43.150035Z","shell.execute_reply.started":"2025-08-01T11:32:43.075067Z","shell.execute_reply":"2025-08-01T11:32:43.149284Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## MACS & FLOPS","metadata":{}},{"cell_type":"code","source":"import tensorflow as tf\nfrom tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2_as_graph\n\ndef get_flops(model, input_shape):\n    # Convert Keras model to ConcreteFunction\n    concrete_func = tf.function(model).get_concrete_function(tf.TensorSpec(input_shape, tf.float32))\n    \n    # Convert ConcreteFunction to a frozen graph\n    frozen_func, graph_def = convert_variables_to_constants_v2_as_graph(concrete_func)\n    \n    # Create a session to run the frozen graph\n    with tf.compat.v1.Session(graph=tf.Graph()) as sess:\n        tf.import_graph_def(graph_def, name=\"\")\n        graph = tf.compat.v1.get_default_graph()\n        \n        # Calculate FLOPs using TensorFlow profiler\n        run_meta = tf.compat.v1.RunMetadata()\n        opts = tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()\n        flops = tf.compat.v1.profiler.profile(graph=graph, run_meta=run_meta, cmd='op', options=opts)\n        \n    return flops.total_float_ops","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.150820Z","iopub.execute_input":"2025-08-01T11:32:43.151044Z","iopub.status.idle":"2025-08-01T11:32:43.162811Z","shell.execute_reply.started":"2025-08-01T11:32:43.151027Z","shell.execute_reply":"2025-08-01T11:32:43.162125Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# # Define input shape\n# input_shape = (1, 224, 224,3 )\n# # Calculate FLOPs\n# flops = get_flops(model, input_shape)\n# macs = flops // 2  # MACs is half the FLOPs for Conv2D\n\n# print(f\"MACs: {macs}\")\n# print(f\"FLOPs: {flops}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.163594Z","iopub.execute_input":"2025-08-01T11:32:43.163820Z","iopub.status.idle":"2025-08-01T11:32:43.175937Z","shell.execute_reply.started":"2025-08-01T11:32:43.163799Z","shell.execute_reply":"2025-08-01T11:32:43.175210Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Parameters","metadata":{}},{"cell_type":"code","source":"import numpy as np\n\ndef model_parametersInfo(model):\n    \"\"\"\n    Prints total, trainable, and non-trainable parameters of a Keras model.\n    \"\"\"\n    total_params = model.count_params()\n    trainable_params = np.sum([np.prod(v.shape) for v in model.trainable_variables])\n    non_trainable_params = np.sum([np.prod(v.shape) for v in model.non_trainable_variables])\n\n    print(f\"Total params: {total_params:,}\")\n    print(f\"Trainable params: {trainable_params:,}\")\n    print(f\"Non-trainable params: {non_trainable_params:,}\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.176625Z","iopub.execute_input":"2025-08-01T11:32:43.176847Z","iopub.status.idle":"2025-08-01T11:32:43.182432Z","shell.execute_reply.started":"2025-08-01T11:32:43.176831Z","shell.execute_reply":"2025-08-01T11:32:43.181834Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"## Inference Report","metadata":{}},{"cell_type":"code","source":"import tensorflow as tf\nimport psutil\nimport os\nimport time\n\ndef inference_Report(model, test_set):\n    process = psutil.Process(os.getpid())\n\n    if tf.config.list_physical_devices('GPU'):\n        gpu_enabled = True\n        gpu_device = 'GPU:0'\n    else:\n        gpu_enabled = False\n\n    print(\"🔍 Capturing memory stats and inference time for 100 images...\\n\")\n\n    # --- Before Inference ---\n    ram_before = process.memory_info().rss\n    gpu_before = tf.config.experimental.get_memory_info(gpu_device)['current'] if gpu_enabled else 0\n\n    total_images = 0\n    start_time = time.time()\n\n    for images, _ in test_set:\n        for i in range(images.shape[0]):\n            if total_images >= 100:\n                break\n            image = tf.expand_dims(images[i], axis=0)\n            _ = model(image, training=False)\n            total_images += 1\n        if total_images >= 100:\n            break\n\n    end_time = time.time()\n\n    # --- After Inference ---\n    ram_after = process.memory_info().rss\n    gpu_after = tf.config.experimental.get_memory_info(gpu_device)['current'] if gpu_enabled else 0\n\n    # --- Results ---\n    print(\"📊 Inference Resource Usage Summary (100 images):\")\n    print(f\"CPU RAM Before: {ram_before / (1024 ** 2):.2f} MB\")\n    print(f\"CPU RAM After : {ram_after / (1024 ** 2):.2f} MB\")\n    print(f\"CPU RAM Used  : {(ram_after - ram_before) / (1024 ** 2):.2f} MB\\n\")\n\n    if gpu_enabled:\n        print(f\"GPU Mem Before: {gpu_before / (1024 ** 2):.2f} MB\")\n        print(f\"GPU Mem After : {gpu_after / (1024 ** 2):.2f} MB\")\n        print(f\"GPU Mem Used  : {(gpu_after - gpu_before) / (1024 ** 2):.2f} MB\\n\")\n    else:\n        print(\"GPU Not Detected.\\n\")\n    time_per_img = (end_time - start_time)/100\n    print(f\"🕒 Inference Time per image: {time_per_img:.4f} seconds\")\n\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.183128Z","iopub.execute_input":"2025-08-01T11:32:43.183449Z","iopub.status.idle":"2025-08-01T11:32:43.192554Z","shell.execute_reply.started":"2025-08-01T11:32:43.183429Z","shell.execute_reply":"2025-08-01T11:32:43.191991Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# Data Pipeline","metadata":{}},{"cell_type":"code","source":"def retreive_dataset(set_name):\n    images,labels=[],[]\n    def imgResize(img,width,height):\n        imgResize = cv2.resize(img,(width,height))\n        return imgResize\n    \n    for (img, imclass) in zip(set_name['filepath'], set_name['Label']):\n        img = cv2.imread(img)\n\n        # img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)   #BGR to RGB\n        img = cv2.resize(img, (224, 224))    \n        # img = img.astype(np.float32) / 255.0   #preprocess\n        images.append(img)\n        if(imclass==True):\n            labels.append(1)\n        else:\n            labels.append(0)\n    print(f\"Done\")\n    return np.array(images),np.array(labels)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.193291Z","iopub.execute_input":"2025-08-01T11:32:43.193606Z","iopub.status.idle":"2025-08-01T11:32:43.204626Z","shell.execute_reply.started":"2025-08-01T11:32:43.193579Z","shell.execute_reply":"2025-08-01T11:32:43.204057Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"X_train,y_train=retreive_dataset(train_df)\nX_val,y_val=retreive_dataset(val_df)\nX_test,y_test=retreive_dataset(test_df)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:32:43.205358Z","iopub.execute_input":"2025-08-01T11:32:43.205659Z","iopub.status.idle":"2025-08-01T11:41:41.996351Z","shell.execute_reply.started":"2025-08-01T11:32:43.205643Z","shell.execute_reply":"2025-08-01T11:41:41.995513Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\n\nindex = 0  # Change this to view another image\nimage = X_train[index]\nlabel = y_train[index]\n\nplt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))\nplt.title(f\"Label: {'DR Present' if label == 1 else 'No DR'}\")\nplt.axis('off')\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:41.997298Z","iopub.execute_input":"2025-08-01T11:41:41.997887Z","iopub.status.idle":"2025-08-01T11:41:42.160875Z","shell.execute_reply.started":"2025-08-01T11:41:41.997859Z","shell.execute_reply":"2025-08-01T11:41:42.160074Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"train_set_raw=tf.data.Dataset.from_tensor_slices((X_train,y_train))\nvalid_set_raw=tf.data.Dataset.from_tensor_slices((X_val,y_val))\ntest_set_raw=tf.data.Dataset.from_tensor_slices((X_test,y_test))","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:42.161711Z","iopub.execute_input":"2025-08-01T11:41:42.161921Z","iopub.status.idle":"2025-08-01T11:41:45.849007Z","shell.execute_reply.started":"2025-08-01T11:41:42.161904Z","shell.execute_reply":"2025-08-01T11:41:45.848309Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# MobileNet","metadata":{}},{"cell_type":"code","source":"tf.keras.backend.clear_session()  # extra code – resets layer name counter\n\nbatch_size = 32\npreprocess = tf.keras.applications.mobilenet.preprocess_input\ntrain_set = train_set_raw.map(lambda X, y: (preprocess(tf.cast(X, tf.float32)), y))\ntrain_set = train_set.shuffle(1000, seed=42).batch(batch_size).prefetch(1)\nvalid_set = valid_set_raw.map(lambda X, y: (preprocess(tf.cast(X, tf.float32)), y)).batch(batch_size)\ntest_set = test_set_raw.map(lambda X, y: (preprocess(tf.cast(X, tf.float32)), y)).batch(batch_size)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:45.852554Z","iopub.execute_input":"2025-08-01T11:41:45.853382Z","iopub.status.idle":"2025-08-01T11:41:47.764129Z","shell.execute_reply.started":"2025-08-01T11:41:45.853350Z","shell.execute_reply":"2025-08-01T11:41:47.763579Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"plt.figure(figsize=(12, 12))\nfor X_batch, y_batch in train_set.take(1):\n    for index in range(9):\n        plt.subplot(3, 3, index + 1)\n        img = (X_batch[index].numpy())\n        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n        plt.imshow(img /np.max(img)) # rescale to 0–1 for imshow()\n        if(y_batch[index]==1):\n            classt= 'DR Present'\n        else:\n            classt= \"No DR\"\n        plt.title(f\"Class: {classt}\")\n        plt.axis(\"off\")\n\nplt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:47.764768Z","iopub.execute_input":"2025-08-01T11:41:47.764963Z","iopub.status.idle":"2025-08-01T11:41:49.592778Z","shell.execute_reply.started":"2025-08-01T11:41:47.764946Z","shell.execute_reply":"2025-08-01T11:41:49.592131Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"tf.random.set_seed(42)  # Ensures reproducibility\n\ntf.keras.backend.clear_session()\n# Ensures reproducibility\n\n\nbase_model = tf.keras.applications.MobileNetV3Large(weights=\"imagenet\", include_top=False)\navg = tf.keras.layers.GlobalAveragePooling2D()(base_model.output)\noutput = tf.keras.layers.Dense(1, activation=\"sigmoid\")(avg)\nmodel = tf.keras.Model(inputs=base_model.input, outputs=output)\nfor layer in base_model.layers:\n    layer.trainable = False","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:49.593699Z","iopub.execute_input":"2025-08-01T11:41:49.593961Z","iopub.status.idle":"2025-08-01T11:41:51.755820Z","shell.execute_reply.started":"2025-08-01T11:41:49.593942Z","shell.execute_reply":"2025-08-01T11:41:51.755042Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)\nmodel.compile(loss=\"binary_crossentropy\", optimizer=optimizer,\n              metrics=[\"accuracy\"])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:51.756649Z","iopub.execute_input":"2025-08-01T11:41:51.756834Z","iopub.status.idle":"2025-08-01T11:41:51.769993Z","shell.execute_reply.started":"2025-08-01T11:41:51.756820Z","shell.execute_reply":"2025-08-01T11:41:51.769515Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.init(\n    project=\"Diabetic Retinopathy Detection\",   # change this\n    name=\"MobilenNetV3\",        # optional\n    config={                       # optional: log hyperparameters\n        \"epochs\": 10,\n        \"batch_size\": 32,\n        \"learning_rate\": 0.001\n    }\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:51.770713Z","iopub.execute_input":"2025-08-01T11:41:51.770965Z","iopub.status.idle":"2025-08-01T11:41:59.243055Z","shell.execute_reply.started":"2025-08-01T11:41:51.770944Z","shell.execute_reply":"2025-08-01T11:41:59.242473Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from tensorflow.keras.callbacks import ModelCheckpoint, CSVLogger\n# from wandb.keras import WandbMetricsLogger, WandbModelCheckpoint\n\nmodel_name = \"DR_MobileNetV3\"\n\n# checkpoint_callback = ModelCheckpoint(\n#     filepath='best_model.keras',  # Path to save the model\n#     monitor='val_accuracy',        # yMetric to monitor\n#     save_best_only=True,       # Save only the best model\n#     save_weights_only=False,   # Save the entire model, not just weights\n#     mode='max',                # Mode to minimize the monitored metric\n#     verbose=1                  # Verbosity mode\n# )\ncsv_logger = CSVLogger(model_name+'.csv',append = True)\n\n# callbacks = [checkpoint_callback,csv_logger]\ncallbacks = [csv_logger,wandb.keras.WandbMetricsLogger(log_freq=5)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:59.243845Z","iopub.execute_input":"2025-08-01T11:41:59.244458Z","iopub.status.idle":"2025-08-01T11:41:59.274331Z","shell.execute_reply.started":"2025-08-01T11:41:59.244432Z","shell.execute_reply":"2025-08-01T11:41:59.273598Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"history = model.fit(train_set, validation_data=valid_set, epochs=10, callbacks= callbacks)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:41:59.275167Z","iopub.execute_input":"2025-08-01T11:41:59.275371Z","iopub.status.idle":"2025-08-01T11:43:32.577625Z","shell.execute_reply.started":"2025-08-01T11:41:59.275355Z","shell.execute_reply":"2025-08-01T11:43:32.577017Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model.save_weights(model_name+'.weights.h5')\n\n# model.load_weights('/kaggle/working/DR_Xception.weights.h5')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:43:32.578456Z","iopub.execute_input":"2025-08-01T11:43:32.578690Z","iopub.status.idle":"2025-08-01T11:43:32.984107Z","shell.execute_reply.started":"2025-08-01T11:43:32.578665Z","shell.execute_reply":"2025-08-01T11:43:32.983545Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"Results(model,test_set,model_name=model_name,save_csv=True,wandb_log=True)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:43:32.984811Z","iopub.execute_input":"2025-08-01T11:43:32.985008Z","iopub.status.idle":"2025-08-01T11:43:54.344062Z","shell.execute_reply.started":"2025-08-01T11:43:32.984992Z","shell.execute_reply":"2025-08-01T11:43:54.343287Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.save(f\"/kaggle/working/{model_name}.metrices.csv\", policy=\"now\")\nwandb.save(f\"/kaggle/working/{model_name}.weights.h5\", policy=\"now\")\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:43:54.345191Z","iopub.execute_input":"2025-08-01T11:43:54.345424Z","iopub.status.idle":"2025-08-01T11:43:54.354780Z","shell.execute_reply.started":"2025-08-01T11:43:54.345407Z","shell.execute_reply":"2025-08-01T11:43:54.354056Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"inference_Report(model,test_set)\nmodel_parametersInfo(model)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:43:54.355590Z","iopub.execute_input":"2025-08-01T11:43:54.356249Z","iopub.status.idle":"2025-08-01T11:44:10.449702Z","shell.execute_reply.started":"2025-08-01T11:43:54.356230Z","shell.execute_reply":"2025-08-01T11:44:10.448813Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Define input shape\ninput_shape = (1, 224, 224,3 )\n# Calculate FLOPs\nflops = get_flops(model, input_shape)\nmacs = flops // 2  # MACs is half the FLOPs for Conv2D\n\nprint(f\"MACs: {macs}\")\nprint(f\"FLOPs: {flops}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:10.450558Z","iopub.execute_input":"2025-08-01T11:44:10.451368Z","iopub.status.idle":"2025-08-01T11:44:14.890756Z","shell.execute_reply.started":"2025-08-01T11:44:10.451342Z","shell.execute_reply":"2025-08-01T11:44:14.889980Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.finish()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:14.891580Z","iopub.execute_input":"2025-08-01T11:44:14.892128Z","iopub.status.idle":"2025-08-01T11:44:15.516656Z","shell.execute_reply.started":"2025-08-01T11:44:14.892080Z","shell.execute_reply":"2025-08-01T11:44:15.516124Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"# MobileNet fine tuned","metadata":{}},{"cell_type":"code","source":"n =len(base_model.layers)\nL = int(0.3*n)  # 30% of n\nf\"{L} trainable of {n} layers\"","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:15.517340Z","iopub.execute_input":"2025-08-01T11:44:15.517639Z","iopub.status.idle":"2025-08-01T11:44:15.522806Z","shell.execute_reply.started":"2025-08-01T11:44:15.517620Z","shell.execute_reply":"2025-08-01T11:44:15.522224Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"for layer in base_model.layers[-L:]:\n    layer.trainable = True","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:15.523552Z","iopub.execute_input":"2025-08-01T11:44:15.523718Z","iopub.status.idle":"2025-08-01T11:44:15.534321Z","shell.execute_reply.started":"2025-08-01T11:44:15.523705Z","shell.execute_reply":"2025-08-01T11:44:15.533640Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)\nmodel.compile(loss=\"binary_crossentropy\", optimizer=optimizer,\n              metrics=[\"accuracy\"])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:15.534957Z","iopub.execute_input":"2025-08-01T11:44:15.535249Z","iopub.status.idle":"2025-08-01T11:44:15.556518Z","shell.execute_reply.started":"2025-08-01T11:44:15.535223Z","shell.execute_reply":"2025-08-01T11:44:15.555995Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.init(\n    project=\"Diabetic Retinopathy Detection\", \n    name=\"MobileNet_ft\",   # Change this to \"VGG_Model_Run\", etc. for next model\n    resume=\"allow\"\n)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:15.560201Z","iopub.execute_input":"2025-08-01T11:44:15.560386Z","iopub.status.idle":"2025-08-01T11:44:22.138801Z","shell.execute_reply.started":"2025-08-01T11:44:15.560372Z","shell.execute_reply":"2025-08-01T11:44:22.138061Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model_name = model_name+'_ft'\nmodel_name","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:22.139641Z","iopub.execute_input":"2025-08-01T11:44:22.140255Z","iopub.status.idle":"2025-08-01T11:44:22.145002Z","shell.execute_reply.started":"2025-08-01T11:44:22.140227Z","shell.execute_reply":"2025-08-01T11:44:22.144498Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from tensorflow.keras.callbacks import ModelCheckpoint, CSVLogger\n\ncsv_logger = CSVLogger(model_name+'.csv',append = True)\n\ncallbacks = [csv_logger,wandb.keras.WandbMetricsLogger(log_freq=5)]","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:22.145912Z","iopub.execute_input":"2025-08-01T11:44:22.146178Z","iopub.status.idle":"2025-08-01T11:44:22.157754Z","shell.execute_reply.started":"2025-08-01T11:44:22.146158Z","shell.execute_reply":"2025-08-01T11:44:22.157138Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"history = model.fit(train_set, validation_data=valid_set, epochs=10, callbacks= callbacks)\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:44:22.158640Z","iopub.execute_input":"2025-08-01T11:44:22.158877Z","iopub.status.idle":"2025-08-01T11:46:28.078915Z","shell.execute_reply.started":"2025-08-01T11:44:22.158853Z","shell.execute_reply":"2025-08-01T11:46:28.078146Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model.save_weights(model_name+'.weights.h5')\n\n# model.load_weights('/kaggle/working/DR_Xception.weights.h5')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:46:28.079894Z","iopub.execute_input":"2025-08-01T11:46:28.080401Z","iopub.status.idle":"2025-08-01T11:46:28.528337Z","shell.execute_reply.started":"2025-08-01T11:46:28.080380Z","shell.execute_reply":"2025-08-01T11:46:28.527791Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"\nResults(model,test_set,model_name=model_name,save_csv=True,wandb_log=True)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:46:28.529132Z","iopub.execute_input":"2025-08-01T11:46:28.529390Z","iopub.status.idle":"2025-08-01T11:46:48.313354Z","shell.execute_reply.started":"2025-08-01T11:46:28.529372Z","shell.execute_reply":"2025-08-01T11:46:48.312766Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.save(f\"/kaggle/working/{model_name}.metrices.csv\", policy=\"now\")\nwandb.save(f\"/kaggle/working/{model_name}.weights.h5\", policy=\"now\")\nprint(f\"saved to wandb {model_name}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:46:48.314074Z","iopub.execute_input":"2025-08-01T11:46:48.314299Z","iopub.status.idle":"2025-08-01T11:46:48.322883Z","shell.execute_reply.started":"2025-08-01T11:46:48.314283Z","shell.execute_reply":"2025-08-01T11:46:48.322269Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"inference_Report(model,test_set)\nmodel_parametersInfo(model)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:46:48.323667Z","iopub.execute_input":"2025-08-01T11:46:48.323903Z","iopub.status.idle":"2025-08-01T11:47:04.269472Z","shell.execute_reply.started":"2025-08-01T11:46:48.323888Z","shell.execute_reply":"2025-08-01T11:47:04.268892Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Define input shape\ninput_shape = (1, 224, 224,3 )\n# Calculate FLOPs\nflops = get_flops(model, input_shape)\nmacs = flops // 2  # MACs is half the FLOPs for Conv2D\n\nprint(f\"MACs: {macs}\")\nprint(f\"FLOPs: {flops}\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:47:04.270317Z","iopub.execute_input":"2025-08-01T11:47:04.270557Z","iopub.status.idle":"2025-08-01T11:47:08.359794Z","shell.execute_reply.started":"2025-08-01T11:47:04.270541Z","shell.execute_reply":"2025-08-01T11:47:08.359042Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"wandb.finish()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-08-01T11:47:08.360650Z","iopub.execute_input":"2025-08-01T11:47:08.360881Z","iopub.status.idle":"2025-08-01T11:47:08.978965Z","shell.execute_reply.started":"2025-08-01T11:47:08.360864Z","shell.execute_reply":"2025-08-01T11:47:08.978259Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}