How fanghuachen/aednet works
Project type and purpose
• This is a research‐oriented deep‐learning pipeline (“AEDNet”) for asynchronous event denoising. It consumes irregular spatiotemporal event data, trains a neural network to remove noise, and evaluates denoising performance.
File structure
• Top‐level directories and files:
• BA_noise_removal/ — core Python modules
– aednet.py (model definition)
– dataprocess.py (data loader & preprocessor)
– train_net.py (training script)
– test_net.py (inference script)
– results/ (directory for outputs: plots, logs)
– __pycache__/
• data/ — raw and split datasets
– AEDNetDataset_BA/ (training data + file lists)
– AEDNetTestset_BA/ (test data)
• models/ — saved model artifacts
– BA_noise_removal_model/
· AEDNet_params.pth (trained weights)
· AEDNet_description.txt
• pic/ — example visualizations
• README.md
README for additional insights
• Purpose: implement and evaluate the AEDNet denoising network.
• Key components:
• Data organization and naming conventions
• Training command‑line interface (CLI)
• Testing CLI
• Dependencies: Python 3.7, PyTorch, NumPy, Matplotlib, tensorboardX
a. Main components:
1. Data Layer
• Raw event files under data/AEDNetDataset_BA & data/AEDNetTestset_BA
• File lists (training_file_name.txt, test_file_name.txt)
2. Data Processing Module
• dataprocess.py: loads .npy files, constructs batches, applies any normalization
3. Model Definition
• aednet.py: PyTorch neural network graph implementing spatial‑temporal correlation layers
4. Training Engine
• train_net.py:
· Reads CLI args (train/test file lists, hyperparameters)
· Invokes dataprocess → model → loss functions → optimizer
· Logs metrics, checkpoints to models/BA_noise_removal_model & results/
5. Inference Engine
• test_net.py:
· Loads saved weights from models/BA_noise_removal_model/AEDNet_params.pth
· Processes test data via dataprocess + aednet → outputs denoised event streams → saves to results/
6. Outputs & Visualization
• results/ folder: performance logs, plots
• pic/ folder: example GIFs demonstrating denoising quality
7. External Dependencies & Environment
• PyTorch, CUDA, NumPy, Matplotlib, tensorboardX
b. Relationships and interactions
• Data Layer feeds into Data Processing Module
• Data Processing Module pipelines batches into Model Definition
• Training Engine orchestrates the forward/backward passes and writes checkpoints
• Inference Engine reads checkpoints, drives the same data‑processing → model → output flow
• Results are visualized and compared against ground truth
c. Architectural patterns and design principles
• Modular separation: data I/O, preprocessing, model, training loop, testing loop
• Clear CLI interfaces for training vs. testing
• Checkpointing for reproducibility
• Use of standard deep‑learning framework patterns (dataset → dataloader → model → optimizer)
d. Key technologies / frameworks
• Python 3.7
• PyTorch (+ CUDA)
• NumPy
• Matplotlib & tensorboardX for logging and visualization
Final checklist
• Clear, descriptive labels on every component
• Legends or annotations for file‐locations and key parameters
• Emphasis on capturing all five modules (data, preprocess, model, train, test) without over‑splitting