How vietnh1009/very-deep-cnn-pytorch works
Project type and purpose
• This is a research‐oriented machine-learning training pipeline implemented in PyTorch.
• It implements the “Very Deep Convolutional Networks for Text Classification” (VDCNN) model, orchestrates data loading, model construction, training loops, logging, and visualization for multiple text‐classification datasets.
File structure
• Top level:
• train.py – entry point / CLI for training
• src/ – core Python modules
– dataset.py: Dataset loading and batching
– utils.py: shared utilities (e.g., text preprocessing, tokenization)
– very_deep_cnn.py: definition of the VDCNN model architectures (depth 9, 17, 29, 49)
• input/ – raw datasets (CSV & metadata)
• output/ – per‐experiment log directories (loss/accuracy metrics, confusion matrices in logs.txt)
• tensorboard/ – TensorBoard event files per run
• visualization/ – pre‐rendered accuracy/loss plots
• README.md & LICENSE
• Patterns & configuration:
• CLI pattern (train.py parses arguments: dataset name, depth, optimizer, input/output paths)
• Modular separation: dataset loader, model definition, training loop, logging
• Standard PyTorch training architecture
README for additional insights
• Describes supported datasets, hyperparameter settings (SGD/Adam, learning rates)
• Usage examples for training on standard & custom datasets
• Directory conventions for logs and TensorBoard
a. Data Layer
• Input storage (input/* CSVs)
• Dataset Loader (src/dataset.py) – reads CSV → token sequences → batches
b. Model Layer
• VDCNN model definitions (src/very_deep_cnn.py)
• Configurable depth parameter (9, 17, 29, 49)
c. Training Controller
• train.py – CLI
• Builds DataLoader + Model + Optimizer + Scheduler
• Training Loop: forward → loss → backward → update
d. Logging & Monitoring
• File logging (output/<dataset>_depth_<n>/logs.txt)
• TensorBoard writer (tensorboard/vdcnn_<depth>_<dataset>/events.*)
e. Visualization
• Static plots in visualization/ (PNG output of training/test curves)
f. External/Infrastructure
• GPU accelerator (NVIDIA Titan X / Quadro)
• Python environment with PyTorch
Relationships and data flows
• Training Loop drives compute on GPU → produces metrics
• Metrics branch to (a) File Logger and (b) TensorBoard Writer
• After training, Visualization scripts read logs/TensorBoard events → generate PNGs
Architectural patterns & design principles
• Separation of concerns: data loading vs. model definition vs. training orchestration
• Pluggable model depth and optimizer via CLI arguments
• Logging abstraction: unified interfaces for text files and TensorBoard
• Reproducibility: directory structure isolates each experiment
Don’t overthink—maximize clarity
• Break out each logical component
• How CLI arguments configure the pipeline
• Clearly distinguish experiment artifacts from core modules