| license: apache-2.0 | |
| tags: | |
| - flight-planning | |
| - transformer | |
| - coordinate-prediction | |
| - sequence-to-sequence | |
| - count-classification | |
| # Flight Plan Coordinate Prediction Model (Seq2SeqCoordsTransformer) | |
| Encoder-Decoder Transformer model trained for AI flight planning project. Predicts normalized coordinates directly and waypoint count via classification. | |
| ## Model Description | |
| Seq2SeqCoordsTransformer architecture using `torch.nn.Transformer`. Predicts normalized lat/lon coordinates autoregressively and waypoint count (0-10) via classification head on encoder output. | |
| * Embed Dim: 256, Heads: 8, Enc Layers: 4, Dec Layers: 4, Max Waypoints: 10 | |
| ## Intended Use | |
| Research prototype. **Not for real-world navigation.** | |
| ## Limitations | |
| Accuracy depends on data/tuning. Fixed max waypoints (10). Not certified. **Architecture differs significantly from previous versions in this repo.** | |
| ## How to Use | |
| Requires loading the custom `Seq2SeqCoordsTransformer` class and weights. Generation requires autoregressive decoding and taking argmax of count logits. | |
| ## Training Data | |
| Trained on `frankmorales2020/flight_plan_waypoints` - https://huggingface.co/datasets/frankmorales2020/flight_plan_waypoints. | |
| ## Contact | |
| Frank Morales, BEng, MEng, SMIEEE (Boeing ATF) - https://www.linkedin.com/in/frank-morales1964/ |