How one-ear/ucdlcode works
This repository implements a research workflow for classifying gait-related movement and affect from skeleton sequences. The principal path is to load paired pose/movement features and labels, prepare batches, classify them with a graph-temporal neural network, and train/evaluate the model. The excerpts support the feeder, graph, model, and training entry point; wiring for some configured features is inferred only where the entry-point/config names make it explicit.
Main entry points: main_diff_combine_double_fagg.py Train and evaluate · agcn_diff_combine_score_fagg.py FAGG classifier · feeder_same_combine.py Paired data feeder · train_diff_combine_double_score_fagg.yaml Run configuration
Components 15
Training workflow
- Run configuration
config/EGait_journal/train_diff_combine_double_score_fagg.yaml - Train and evaluate
main_diff_combine_double_fagg.py - Optimization loop
main_diff_combine_double_fagg.py
Sequence data
- Skeleton data
- Paired data feeder
feeders/feeder_same_combine.py - Sequence transforms
feeders/tools.py
Graph model
- EGait topology
graph/EGait.py - Adjacency builder
graph/tools.py - FAGG classifier
model/agcn_diff_combine_score_fagg.py - Training losses
model/loss.py
Training results
- Model checkpoint
output_pAST/egait_runs/runs_diff_combine_fagg/fagg_att_cascade_st/best_acc.pt - Run logs
output_pAST/egait_runs/work_diff_combine_fagg/fagg_att_cascade_st/log.txt - Training summaries
output_pAST/egait_runs/runs_diff_combine_fagg/fagg_att_cascade_st/train - Class and affect scores
Other components
- Researcher
Connections 16
- Researcher → Train and evaluate: starts run
- Researcher → Run configuration: sets parameters
- Run configuration → Train and evaluate: configures
- Skeleton data → Paired data feeder: loads samples
- Train and evaluate → Paired data feeder: creates batches
- Paired data feeder → Sequence transforms: transforms sequences
- Train and evaluate → FAGG classifier: runs model
- FAGG classifier → EGait topology: loads graph
- EGait topology → Adjacency builder: builds adjacency
- Train and evaluate → Training losses: computes objective
- Train and evaluate → Optimization loop: updates weights
- Optimization loop → FAGG classifier: updates parameters
- Train and evaluate → Model checkpoint: saves weights
- Train and evaluate → Run logs: writes logs
- Train and evaluate → Training summaries: writes summaries
- FAGG classifier → Class and affect scores: returns scores