How deeppsp/cinc2022 works
Main Components Identification
The system should be divided into these major components:
a) Data Processing Layer
• Data Reader (data_reader.py)
• Dataset Handler (dataset.py)
• Augmentation & Transforms (utils/transforms.py, utils/augmentations.py)
b) Model Layer
• Core Models
• CRNN (models/crnn.py)
• Wav2Vec2 (models/wav2vec2.py)
• MLP (models/mlp.py)
• Model Heads (models/heads.py)
• Multi-task Learning Components (models/model_ml.py)
c) Training Infrastructure
• Trainer (trainer.py)
• Configuration (cfg.py, cfg_models.py)
• Evaluation (evaluate_model.py)
d) Feature Engineering
• Signal Processing Utils
• Springer DWT (utils/springer_dwt.py)
• Heart Rate Detection (utils/schmidt_heart_rate.py)
• Spike Removal (utils/schmidt_spike_removal.py)
[Input Data] → [Data Processing Layer] → [Feature Engineering] → [Model Layer] → [Training Infrastructure] → [Evaluation & Output]
Specific Elements
a) Data Flow Path:
Raw PCG Data → Preprocessing → Feature Extraction → Model Training → Predictions
b) Model Architecture Details:
• Parallel paths for different model types (CRNN, Wav2Vec2)
• Multi-task learning branches
• Model head connections
c) Training Pipeline:
• Configuration injection points
• Evaluation feedback loop
• Metrics collection points