How a-pru/descent works
DESCENT is a research system for predicting airport surface movement trajectories from Amelia-10 scenes and directed lane context. Its main workflow prepares airport-specific scene data, transforms agents and lane segments into an ego frame, predicts multimodal trajectories, and trains or evaluates them with displacement and lane-adherence objectives. Separate entry points generate lane maps and visualize predictions. The README documents these capabilities; some script and configuration wiring is not shown in the excerpts, so those links are cited to the README rather than inferred from module names.
Main entry points: dataset.py Scene dataset · descent.py DESCENT network · trajpred.py Prediction task · eval.py Evaluation command
Components 20
Scene and Maps
- Data module
descent/data/datamodule.py - Scene dataset
descent/data/dataset.py - Data helpers
descent/utils/data_utils.py - Feature masks
descent/utils/global_masks.py - Directed lane maps
maps
Prediction Model
- Lane encoder
descent/models/layers/lane_embedding.py - Attention blocks
descent/models/layers/transformer_blocks.py - Trajectory decoder
descent/models/layers/decoder.py - DESCENT network
descent/models/descent.py
Training and Evaluation
- Prediction task
descent/models/trajpred.py - Training losses
descent/utils/losses.py - Trajectory metrics
descent/utils/metrics.py - Model checkpoints
checkpoints
User Workflows
- Training command
scripts/train.py - Evaluation command
scripts/eval.py - Visualization command
scripts/visualize.py - Map generation
scripts/generate_maps.py - Prediction figures
scripts/visualize.py
Other components
- Researcher
- Amelia-10 scenes
Connections 25
- Researcher → Training command: starts training
README.md - Researcher → Evaluation command: requests evaluation
README.md - Researcher → Visualization command: requests visualization
README.md - Researcher → Map generation: generates maps
README.md - Amelia-10 scenes → Scene dataset: provides scenes
README.md - Map generation → Directed lane maps: writes maps
README.md - Scene dataset → Data helpers: uses listing helpers
descent/data/dataset.py - Data module → Data helpers: uses utilities
descent/data/datamodule.py - Scene dataset → Feature masks: uses feature indices
descent/data/dataset.py - Directed lane maps → Scene dataset: supplies lane context
README.md - Training command → Data module: configures data
README.md - Evaluation command → Data module: evaluates test data
README.md - Data module → Scene dataset: provides datasets
descent/data/datamodule.py - Scene dataset → DESCENT network: provides scene inputs
- DESCENT network → Lane encoder: encodes lane segments
descent/models/descent.py - DESCENT network → Attention blocks: applies attention blocks
descent/models/descent.py - DESCENT network → Trajectory decoder: decodes trajectories
descent/models/descent.py - Prediction task → DESCENT network: runs network
descent/models/trajpred.py - Prediction task → Training losses: computes loss
descent/models/trajpred.py - Prediction task → Trajectory metrics: computes metrics
descent/models/trajpred.py - Training command → Prediction task: trains task
README.md - Evaluation command → Model checkpoints: loads checkpoint
README.md - Evaluation command → Prediction task: evaluates task
README.md - Visualization command → Model checkpoints: uses predictions
- Visualization command → Prediction figures: writes figures
README.md