Introducing Neural: A DSL and Debugger for Neural Networks

Hi everyone! 👾 I’m excited to share "Neural," a project I’ve been working on to simplify neural network development. Neural is a domain-specific language (DSL) and debugger that lets you define, train, and debug models with ease—whether via code, CLI, or a no-code interface. 🎛

Why Neural?

Building neural networks can be complex—boilerplate code, shape mismatches, and debugging woes slow us down. Neural tackles this with:

• A YAML-like DSL: Define models concisely (e.g., Conv2D(filters=32, kernel_size=(3,3))).
• NeuralDbg: Real-time monitoring of gradients, execution traces, and resources, with a --hacky mode for security analysis.
• Cross-Framework Support: Export to TensorFlow, PyTorch, or ONNX.

Example: MNIST Classifier

Here’s a quick .neural file:

network MNISTClassifier {
  input: (28, 28, 1)
  layers:
    Conv2D(filters=32, kernel_size=(3,3), activation="relu")
    MaxPooling2D(pool_size=(2,2))
    Flatten()
    Dense(units=128, activation="relu")
    Output(units=10, activation="softmax")
  loss: "sparse_categorical_crossentropy"
  optimizer: Adam
}

---

Compile And Run

pip install neural-dsl
neural compile mnist.neural --backend pytorch
neural run mnist_pytorch.py

---

Debug And Test

neural debug mnist.neural --hacky

Visit http://localhost:8050 to see gradients, resource usage, and potential security breaches.

---

 Neural-dsl is a WIP DSL and debugger, bugs exist, feedback is welcome!

---

What’s Next?

• I’m adding automatic hyperparameter optimization (HPO), research paper generation, and TensorBoard integration.

---

Try it out on GitHub and let me know what you think!

🦾 Share your feedback—I’d love to hear from the community!

Comments

[Silvia O'Dwyer]

This looks like a very useful library, I'm just wondering if you could link to the GitHub repo? Would love to take a look at it!