About

refrakt_core is a deep learning research framework that re-implements key computer vision papers with high modularity, extensibility, and reproducibility. Designed for education, experimentation, and benchmarking, Refrakt enables rapid iteration on architectures, loss functions, and training routines — all driven via configuration.

This project is part of a broader initiative to understand, unify, and visualize foundational and modern architectures through clean code, clear abstractions, and rigorous logging.

📌 Goals

• Faithfully reproduce influential CV papers
• Modularize training/evaluation pipelines using YAML + registry patterns
• Enable quick debugging and metric logging via TensorBoard and Weights & Biases (W&B)
• Serve as a platform for personal research, fine-tuning, or new model design

📚 Implemented Papers

• Vision Transformer (ViT) – An Image is Worth 16x16 Words
• ResNet – Deep Residual Learning for Image Recognition
• Autoencoders – Learning Representations via Reconstruction
• Swin Transformer – Hierarchical Vision Transformer with Shifted Windows
• Attention is All You Need
• ConvNeXt – A ConvNet for the 2020s
• SRGAN – Photo-Realistic Single Image Super-Resolution with GANs
• SimCLR – A Simple Framework for Contrastive Learning
• DINO – Self-Supervised Vision Transformers
• MAE – Masked Autoencoders
• MSN – Masked Siamese Networks

⚙️ Setup

git clone https://github.com/refrakt-hub/refrakt_core.git
cd refrakt_core

# Create and activate a virtual environment
conda create -n refrakt python=3.10 -y
conda activate refrakt

# Install dependencies
pip install -r requirements.txt

🧪 Running Experiments

# Calling __main__.py
python -m refrakt_core.api --config refrakt_core/config/vit.yaml

# Calling `refrakt` as a command. 
refrakt --config ./refrakt_core/config/resnet.yaml


# Override on-the-fly
python -m refrakt_core.api.train \
    config.model.name=ResNet \
    config.optimizer.lr=0.0005 \
    config.trainer.epochs=20

CLI Usage

usage: refrakt [-h] --config CONFIG --mode {train,test,inference,pipeline}
               [--log_dir LOG_DIR]
               [--log_type [{tensorboard,wandb} ...]]
               [--console]
               [--model_path MODEL_PATH]
               [--debug]

Flag	Description
--config	Path to YAML config file
--mode	Execution mode: train, test, inference, or pipeline
--log_dir	Path to log directory (e.g., ./runs)
--log_type	Logging backend: tensorboard, wandb, or both
--console	Print logs to console as well
--model_path	Path to model weights/checkpoint for inference/test
--debug	Enable debug mode with extra verbosity

🔧 Config Structure (YAML)

All components (model, trainer, loss, optimizer, scheduler) are defined in modular YAML files under refrakt_core/config/.

Example: vit.yaml

dataset:
  name: MNIST
  params:
    root: ./data
    train: true
    download: true
  transform:
    - name: Resize
      params: { size: [28, 28] }
    - name: ToTensor
    - name: Normalize
      params:
        mean: [0.1307]
        std: [0.3081]

dataloader:
  params:
    batch_size: 32
    shuffle: true
    num_workers: 4
    drop_last: false

model:
  name: vit
  params:
    in_channels: 1
    num_classes: 10
    image_size: 28 
    patch_size: 7

loss:
  name: cross_entropy
  params: {}

optimizer:
  name: adamw
  params:
    lr: 0.0003

scheduler: null

trainer:
  name: supervised
  params:
    save_dir: "./checkpoints"
    model_name: "vit"
    num_epochs: 1
    device: cuda

📈 Logging & Monitoring

✅ Supported

• TensorBoard (logs saved in runs/)
• Weights & Biases (auto-logged via config)

Setup

# TensorBoard
tensorboard --logdir=./logs/<model_name>/tensorboard/

# Weights & Biases
export WANDB_API_KEY=your_key_here

🧱 Project Structure

refrakt_core/
├── api/                  # CLI: train.py, test.py, inference.py
│   └── builders/         # Builders for models, losses, optimizers, datasets
│
├── config/               # YAML configurations for each experiment
│   └── schema.py         # Pydantic validation for configs
│
├── losses/               # Contrastive, GAN, MAE, VAE, etc.
│
├── models/               # Vision architectures (ViT, ResNet, MAE, etc.)
│   └── templates/        # Base model templates and abstractions
│
├── trainer/              # Task-specific training logic (SimCLR, SRGAN, etc.)
│
├── registry/             # Decorator-based plugin system
│
├── utils/                # Helper modules (encoders, decoders, data classes)
│
├── transforms.py         # Data augmentation logic
├── datasets.py           # Dataset definitions and loader helpers
├── logging.py            # Logger wrapper for stdout + W&B/TensorBoard

🧩 Extending Refrakt

Add a New Model

1. Create the architecture in models/your_model.py
2. Inherit from a base class in models/templates/models.py
3. Register it using:

from refrakt_core.registry.model_registry import register_model

@register_model("your_model")
class YourModel(BaseClassifier):
    ...

4. Add a YAML config: config/your_model.yaml
5. Write a custom trainer if needed (trainer/your_model.py)s

🔍 Example Output

Once training begins, expect:

• Printed progress bar (via tqdm)
• Metrics like loss/accuracy printed and logged
• ./logs/<model_name>/ with events.out.tfevents for TensorBoard
• W&B dashboard with training curves, if enabled

📬 Contributing

1. Clone and install:

   git clone ...
   pip install -r requirements-dev.txt
   pre-commit install

2. Follow formatting (black, isort, pylint)

3. Write tests for any new feature

4. Run:

   pytest tests/

PRs and creating issues are absolutely welcome.

🔭 Future Scope

Milestone	Description
✅ Stage 1	Paper re-implementations in notebooks
✅ Stage 2	Modular training + model pipelines
🔜 Stage 3	Convert to a Python library (refrakt train, etc.)
🔜 Stage 4	Web app: visual dashboards, interactive inference
🔜 Stage 5	No-code/low-code config engine for multimodal pipelines
🔜 Stage 6	Explainability: saliency maps, attention visualizations

Planned additions include:

• Mixed-modality support (e.g., image + text)
• More contrastive/self-supervised frameworks
• System-level benchmarks across GPUs
• Evaluation on downstream tasks (detection, segmentation)

📄 License

This repository is licensed under the MIT License. See LICENSE for full details.

👤 Maintainer

Akshath Mangudi If you find issues, raise them. If you learn from this, share it. Built with love and curiosity :)