How cchenyuchi/huggingface-art-transformer works
Project type and purpose
• This is an open-source research/project repository implementing an EEG-based ART (Adaptive Response Transformer) classification pipeline in TensorFlow, leveraging Huggingface Transformer conventions.
• Key goals: load EEG/SLT data, preprocess it, configure and train an ART-style transformer model, and evaluate/test it via notebooks.
File structure
• Top-level artifacts:
• ART_model_testing.ipynb, GitVisionModel_testing.ipynb, Trainer.ipynb, eval.ipynb → interactive evaluation & experiment tracking.
• modeling_git.py → glue or packaging script for model versions.
• DataLoader/ → data ingestion & preprocessing code + JSON group definitions.
• FirstMultiModel/EEGART/ → core model code, config folder, pretrained weights.
• Notice a clear pipeline structure: DataLoader → model definition & configuration → training scripts → evaluation notebooks.
README
• Brief: “Huggingface Transformer format for ART”
• Implies that the model code follows transformer conventions (config.json, safetensors), consistent with Huggingface’s layout.
Main components and interactions (
a. Components:
1. Data Ingestion Layer (DataLoader/SLT_dataloader.py + JSONs)
2. Data Processing & Batching (tf_data.py)
3. Model Configuration (artcls-config/config.json + model.safetensors)
4. Model Definition (tf_model.py)
5. Loss & Optimization (tf_loss.py, tf_opt.py)
6. Training Orchestration (tf_train2.py)
7. Utilities (tf_utils.py, tf_config.py)
8. Experiment Scripts/Notebooks (ART_model_testing.ipynb, GitVisionModel_testing.ipynb, Trainer.ipynb, eval.ipynb)
9. Packaging/Versioning (modeling_git.py)
b. Relationships & Dataflow:
• tf_model ← config.json/model.safetensors
• Model outputs → tf_loss → tf_opt → tf_train2 (loop)
• Trained checkpoints → artcls-config/ (persisted)
• Trained model & metrics → eval.ipynb / testing notebooks
c. Architectural patterns/design principles:
• Pipeline-oriented ML architecture (modular layers for data, model, training, evaluation).
• Huggingface-style decoupling of config, weights, and code.
• Separation of concerns: data preprocessing vs. model logic vs. training loop vs. evaluation.
d. Key technologies/frameworks:
• Python & Jupyter Notebooks for experimentation
• TensorFlow for model implementation
• Huggingface Transformer conventions (config.json, safetensors)
• Custom data loader parsing JSON subject groups
Detailed guidance
• Next: optimizer & loss modules feeding into the training orchestration block.
• Finally: outputs feeding into evaluation/test notebooks and the modeling_git packaging step.