How cyf-cuber/fdp works
FDP is a PyTorch micro-expression recognition research model. Its workflow prepares subject-split video datasets from face-frame folders, then trains and evaluates the model on those saved datasets. The model combines a local/global transformer frame representation with 3D classification and dynamic-image components. README documentation establishes dataset preparation and training commands; excerpts show preprocessing, serialization, model construction, and training/evaluation behavior. Some model-forward details and unsampled dataset logic are omitted rather than inferred.
Main entry points: fdp.py FDP network · train.py Training loop · data.py Video preprocessing · dataset Subject datasets
Components 10
Dataset preparation
- Video preprocessing
data.py - Subject datasets
dataset
FDP model
- FDP network
fdp.py - Frame transformer
fdp.py - 3D classifier
fdp.py - Dynamic-image network
fdp.py
Training and evaluation
- Training loop
train.py - Evaluation metrics
train.py
Other components
- Researcher
- Face-frame folders
Connections 10
- Researcher → Video preprocessing: runs preparation
README.md - Face-frame folders → Video preprocessing: provides frames
README.md - Video preprocessing → Subject datasets: saves splits
data.py - Researcher → Training loop: runs training
README.md - Subject datasets → Training loop: loads datasets
train.py - Training loop → FDP network: calls model
train.py - FDP network → Frame transformer: constructs frame encoder
fdp.py - FDP network → 3D classifier: constructs classifier
fdp.py - FDP network → Dynamic-image network: constructs image network
fdp.py - Training loop → Evaluation metrics: evaluates model
train.py