{"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":"none","dataSources":[{"sourceId":59156,"databundleVersionId":13874099,"sourceType":"competition"}],"dockerImageVersionId":31089,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"
\n
\n

Published on September 20, 2025

\n

Author: Jocelyn C. Dumlao

\n

\n LinkedIn\n GitHub\n YouTube\n Kaggle\n
\n
","metadata":{}},{"cell_type":"markdown","source":"
\n \"image\"\n
","metadata":{}},{"cell_type":"markdown","source":"#

Import Libraries

\n","metadata":{}},{"cell_type":"code","source":"import pandas as pd\nimport numpy as np\nfrom tqdm.notebook import tqdm \nfrom sklearn.metrics import f1_score","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:19.160081Z","iopub.execute_input":"2025-12-09T01:06:19.160424Z","iopub.status.idle":"2025-12-09T01:06:20.285729Z","shell.execute_reply.started":"2025-12-09T01:06:19.160396Z","shell.execute_reply":"2025-12-09T01:06:20.284673Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#

Data Loading and Filtering

\n","metadata":{}},{"cell_type":"code","source":"# Data Loading and Filtering\ntrain_df = pd.read_csv('/kaggle/input/MABe-mouse-behavior-detection/train.csv')\ntrain_subset = train_df.query(\"~ lab_id.str.startswith('MABe22_')\")","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:25.798282Z","iopub.execute_input":"2025-12-09T01:06:25.798797Z","iopub.status.idle":"2025-12-09T01:06:25.900641Z","shell.execute_reply.started":"2025-12-09T01:06:25.798771Z","shell.execute_reply":"2025-12-09T01:06:25.899767Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"def score_class_specific(solution, submission):\n \"\"\"\n Calculates F1 score separately for each behavior class and averages them (Macro F1).\n This matches the leaderboard logic much closer than the binary check.\n \"\"\"\n print(\"Calculating Class-Specific Frame-level F1 Score...\")\n \n # Get all unique videos\n video_ids = solution['video_id'].unique()\n \n # Store F1 scores per behavior class\n behavior_scores = {}\n \n for vid in tqdm(video_ids, desc=\"Scoring Videos\"):\n sol_vid = solution[solution['video_id'] == vid]\n sub_vid = submission[submission['video_id'] == vid]\n \n # Get max frame for this video\n if sub_vid.empty and sol_vid.empty:\n continue\n max_frame_sol = sol_vid['stop_frame'].max() if not sol_vid.empty else 0\n max_frame_sub = sub_vid['stop_frame'].max() if not sub_vid.empty else 0\n max_frame = max(max_frame_sol, max_frame_sub)\n \n # Identify all unique behaviors in this video's solution\n unique_behaviors = sol_vid['action'].unique()\n \n for action in unique_behaviors:\n # Create boolean arrays for this specific action\n y_true = np.zeros(max_frame + 1, dtype=int)\n y_pred = np.zeros(max_frame + 1, dtype=int)\n \n # Fill True values\n sol_action = sol_vid[sol_vid['action'] == action]\n for row in sol_action.itertuples():\n y_true[row.start_frame:row.stop_frame] = 1\n \n sub_action = sub_vid[sub_vid['action'] == action]\n for row in sub_action.itertuples():\n # We perform a clip here to ensure we don't go out of bounds if prediction is longer\n start = min(row.start_frame, max_frame)\n stop = min(row.stop_frame, max_frame)\n y_pred[start:stop] = 1\n \n # Calculate F1 for this action in this video\n score = f1_score(y_true, y_pred, zero_division=0)\n \n if action not in behavior_scores:\n behavior_scores[action] = []\n behavior_scores[action].append(score)\n \n # Calculate Macro Average\n final_scores = [np.mean(scores) for scores in behavior_scores.values()]\n macro_f1 = np.mean(final_scores) if final_scores else 0\n \n print(f\"Class-Specific F1 Score: {macro_f1:.4f}\")\n return macro_f1\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:29.331243Z","iopub.execute_input":"2025-12-09T01:06:29.331618Z","iopub.status.idle":"2025-12-09T01:06:29.343005Z","shell.execute_reply.started":"2025-12-09T01:06:29.331581Z","shell.execute_reply":"2025-12-09T01:06:29.341941Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#

Ground Truth Generation

\n","metadata":{}},{"cell_type":"code","source":"# Ground Truth Generation (Validation Baseline)\n\ndef generate_ground_truth(dataset):\n ground_truth_list = []\n # We limit to first 100 for speed in local testing, remove .head(100) for full run\n for _, row in tqdm(dataset.iterrows(), total=len(dataset), desc=\"Processing Annotations\"):\n lab_id = row['lab_id']\n video_id = row['video_id']\n annotation_file_path = f\"/kaggle/input/MABe-mouse-behavior-detection/train_annotation/{lab_id}/{video_id}.parquet\"\n try:\n annotations_df = pd.read_parquet(annotation_file_path)\n annotations_df['video_id'] = video_id\n annotations_df['lab_id'] = lab_id\n annotations_df['behaviors_labeled'] = row['behaviors_labeled']\n annotations_df['agent_id'] = annotations_df['agent_id'].apply(lambda x: f\"mouse{x}\")\n annotations_df['target_id'] = annotations_df['target_id'].apply(lambda x: f\"mouse{x}\")\n ground_truth_list.append(annotations_df)\n except FileNotFoundError:\n continue\n return pd.concat(ground_truth_list, ignore_index=True)\n\n# Generate GT for Validation\nsolution = generate_ground_truth(train_subset)\n\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:36.616296Z","iopub.execute_input":"2025-12-09T01:06:36.616674Z","iopub.status.idle":"2025-12-09T01:06:41.701543Z","shell.execute_reply.started":"2025-12-09T01:06:36.616647Z","shell.execute_reply":"2025-12-09T01:06:41.700758Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#

Prediction

","metadata":{}},{"cell_type":"code","source":"# Prediction \ndef create_all_in_predictions(dataset, traintest):\n \"\"\"\n Strategy: For every behavior listed in 'behaviors_labeled', \n predict it for the FULL duration of the video.\n \"\"\"\n predictions = []\n\n for _, row in tqdm(dataset.iterrows(), total=len(dataset), desc=\"Generating Predictions\"):\n lab_id = row['lab_id']\n video_id = row['video_id']\n \n tracking_file_path = f\"/kaggle/input/MABe-mouse-behavior-detection/{traintest}_tracking/{lab_id}/{video_id}.parquet\"\n \n try:\n # Get video duration\n tracking_meta = pd.read_parquet(tracking_file_path, columns=['video_frame'])\n stop_frame_total = tracking_meta['video_frame'].max() + 1\n except FileNotFoundError:\n # Fallback if tracking file missing\n stop_frame_total = 1800 # default fallback\n # print(f\"Warning: Tracking missing for {video_id}\")\n\n # Parse behaviors\n behaviors_labeled_str = row['behaviors_labeled']\n try:\n behaviors_labeled = eval(behaviors_labeled_str)\n except:\n continue\n \n cleaned_behaviors = [b.replace(\"'\", \"\").strip() for b in behaviors_labeled]\n behavior_tuples = [tuple(b.split(',')) for b in cleaned_behaviors]\n \n # Iterate over EVERY behavior tuple found in the metadata\n for b_tuple in behavior_tuples:\n if len(b_tuple) == 3:\n agent, target, action = b_tuple\n \n # Predict this action for the WHOLE video [0, stop_frame_total]\n predictions.append({\n 'video_id': video_id,\n 'agent_id': agent.strip(),\n 'target_id': target.strip(),\n 'action': action.strip(),\n 'start_frame': 0,\n 'stop_frame': stop_frame_total\n })\n\n predictions_df = pd.DataFrame(predictions)\n return predictions_df\n","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:44.698047Z","iopub.execute_input":"2025-12-09T01:06:44.698471Z","iopub.status.idle":"2025-12-09T01:06:44.708736Z","shell.execute_reply.started":"2025-12-09T01:06:44.698448Z","shell.execute_reply":"2025-12-09T01:06:44.707457Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# Generate predictions\nsubmission = create_all_in_predictions(train_subset, 'train')","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:48.336738Z","iopub.execute_input":"2025-12-09T01:06:48.337099Z","iopub.status.idle":"2025-12-09T01:06:54.032986Z","shell.execute_reply.started":"2025-12-09T01:06:48.337071Z","shell.execute_reply":"2025-12-09T01:06:54.031839Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#

Evaluation

","metadata":{}},{"cell_type":"code","source":"# Evaluation\nprint(\"Evaluating Predictions...\")\nscore_class_specific(solution, submission)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:06:57.243818Z","iopub.execute_input":"2025-12-09T01:06:57.244177Z","iopub.status.idle":"2025-12-09T01:07:23.977920Z","shell.execute_reply.started":"2025-12-09T01:06:57.244148Z","shell.execute_reply":"2025-12-09T01:07:23.976628Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#

Test Data Submission

","metadata":{}},{"cell_type":"code","source":"# Test Data Submission\ntest_df = pd.read_csv('/kaggle/input/MABe-mouse-behavior-detection/test.csv')\ntest_submission = create_all_in_predictions(test_df, 'test')\n\ntest_submission.index.name = 'row_id'\ntest_submission.to_csv('submission.csv')\n\nprint(\"Submission file created successfully!\")\nprint(test_submission.head())","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-12-09T01:07:27.828195Z","iopub.execute_input":"2025-12-09T01:07:27.828519Z","iopub.status.idle":"2025-12-09T01:07:27.869590Z","shell.execute_reply.started":"2025-12-09T01:07:27.828494Z","shell.execute_reply":"2025-12-09T01:07:27.868385Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}