| --- |
| license: mit |
| tags: |
| - computer-vision |
| - classification |
| - xgboost |
| - tensorflow |
| - military |
| - pipeline |
| --- |
| |
| # ATAS Model Weights |
|
|
| Three trained model files for the [ATAS (Aerial Threat Assessment System)](https://huggingface.co/spaces/Eakempreet/ATAS) pipeline. |
|
|
| ## Models |
|
|
| ### 1. Aircraft Classifier |
| - **File:** `aircraft_classifier/atas_final_fine_tuned_aircraft_classifier_model.keras` |
| - **Architecture:** EfficientNetV2-L + custom classification head |
| - **Dataset:** ~12k images, 101 aircraft classes |
| - **Top-1 Accuracy:** 78.08% | **Top-5 Accuracy:** 92.02% |
|
|
| ### 2. ETA Regressor |
| - **File:** `eta/atas_final_eta_regressor_model.joblib` |
| - **Architecture:** XGBoost Regressor (Optuna-tuned, ~944 trials) |
| - **Task:** Predicts time-to-impact in seconds |
| - **R²:** 0.9939 | **MAE:** 0.4552s |
|
|
| ### 3. Hit Classifier |
| - **File:** `hit/atas_final_hit_classifier_model.joblib` |
| - **Architecture:** XGBoost Classifier |
| - **Task:** Predicts missile hit probability after evasion |
| - **Recall:** 0.9966 | **F1:** 0.9968 | **ROC-AUC:** 0.9999 |
|
|
| ## Usage |
|
|
| These models are used together in the ATAS pipeline. See the live demo: |
| 👉 [https://huggingface.co/spaces/Eakempreet/ATAS](https://huggingface.co/spaces/Eakempreet/ATAS) |
|
|
