neural-dsl 0.2.3

A domain-specific language and debugger for neural networks

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

Neural: A Neural Network Programming Language

Neural is a domain-specific language (DSL) designed for defining, training, debugging, and deploying neural networks. With declarative syntax, cross-framework support, and built-in execution tracing (NeuralDbg), it simplifies deep learning development.

🚀 Features

• YAML-like Syntax: Define models intuitively without framework boilerplate.
• Shape Propagation: Catch dimension mismatches before runtime.
• Multi-Backend Export: Generate code for TensorFlow, PyTorch, or ONNX.
• Training Orchestration: Configure optimizers, schedulers, and metrics in one place.
• Visual Debugging: Render interactive 3D architecture diagrams.
• Extensible: Add custom layers/losses via Python plugins.

🛠 NeuralDbg: Built-in Neural Network Debugger

NeuralDbg provides real-time execution tracing, profiling, and debugging, allowing you to visualize and analyze deep learning models in action.

✅ Real-Time Execution Monitoring – Track activations, gradients, memory usage, and FLOPs.

✅ Shape Propagation Debugging – Visualize tensor transformations at each layer.
✅ Gradient Flow Analysis – Detect vanishing & exploding gradients.
✅ Dead Neuron Detection – Identify inactive neurons in deep networks.
✅ Anomaly Detection – Spot NaNs, extreme activations, and weight explosions.
✅ Step Debugging Mode – Pause execution and inspect tensors manually.

📦 Installation

Clone the repository

git clone https://github.com/yourusername/neural.git cd neural

Create a virtual environment (recommended)

python -m venv venv source venv/bin/activate # Linux/macOS venv\Scripts\activate # Windows

Install dependencies

pip install -r requirements.txt

pip install neural-dsl

see v0.1.1 for bug fixes

Prerequisites: Python 3.8+, pip

🛠️ Quick Start

1. Define a Model

Create mnist.neural:

network MNISTClassifier {
  input: (28, 28, 1)  # Channels-last format
  layers:
    Conv2D(filters=32, kernel_size=(3,3), activation="relu")
    MaxPooling2D(pool_size=(2,2))
    Flatten()
    Dense(units=128, activation="relu")
    Dropout(rate=0.5)
    Output(units=10, activation="softmax")
  
  loss: "sparse_categorical_crossentropy"
  optimizer: Adam(learning_rate=0.001)
  metrics: ["accuracy"]
  
  train {
    epochs: 15
    batch_size: 64
    validation_split: 0.2
  }
}

3. Run Or Compile The Model

neural run mnist.neural --backend tensorflow --output mnist_tf.py
# Or for PyTorch:
neural run mnist.neural --backend pytorch --output mnist_torch.py

4. Visualize Architecture

neural visualize mnist.neural --format png

This will create architecture.png, shape_propagation.html, and tensor_flow.html for inspecting the network structure and shape propagation.

5. Debug with NeuralDbg

neural debug mnist.neural

Open your browser to http://localhost:8050 to monitor execution traces, gradients, and anomalies interactively.

6. Use The No-Code Interface

neural --no_code

Open your browser to http://localhost:8051 to build and compile models via a graphical interface.

---

🛠 Debugging with NeuralDbg

🔹 1️⃣ Start Real-Time Execution Tracing

python neural.py debug mnist.neural

Features:
✅ Layer-wise execution trace
✅ Memory & FLOP profiling
✅ Live performance monitoring

🔹 2️⃣ Analyze Gradient Flow

python neural.py debug --gradients mnist.neural

🚀 Detect vanishing/exploding gradients with interactive charts.

🔹 3️⃣ Identify Dead Neurons

python neural.py debug --dead-neurons mnist.neural

🛠 Find layers with inactive neurons (common in ReLU networks).

🔹 4️⃣ Detect Training Anomalies

python neural.py debug --anomalies mnist.neural

🔥 Flag NaNs, weight explosions, and extreme activations.

🔹 5️⃣ Step Debugging (Interactive Tensor Inspection)

python neural.py debug --step mnist.neural

🔍 Pause execution at any layer and inspect tensors manually.

---

🌟 Why Neural?

Feature	Neural	Raw TensorFlow/PyTorch
Shape Validation	✅ Auto	❌ Manual
Framework Switching	1-line flag	Days of rewriting
Architecture Diagrams	Built-in	Third-party tools
Training Config	Unified	Fragmented configs

🔄 Cross-Framework Code Generation

Neural DSL	TensorFlow Output	PyTorch Output
Conv2D(filters=32)	tf.keras.layers.Conv2D(32)	nn.Conv2d(in_channels, 32)
Dense(units=128)	tf.keras.layers.Dense(128)	nn.Linear(in_features, 128)

🏆 Benchmarks

Task	Neural	Baseline (TF/PyTorch)
MNIST Training	1.2x ⚡	1.0x
Debugging Setup	5min 🕒	2hr+

📚 Documentation

• DSL Documentation

Explore advanced features:

• Custom Layers Guide
• ONNX Export Tutorial
• Training Configuration
• NeuralDbg Debugging Features

📚 Examples

Explore common use cases in examples/ with step-by-step guides in docs/examples/:

• MNIST Classifier Guide
• Sentiment Analysis Guide
• Transformer for NLP Guide

---

🤝 Contributing

We welcome contributions! See our:

• Contributing Guidelines
• Code of Conduct
• Roadmap

To set up a development environment:

git clone https://github.com/yourusername/neural.git
cd neural
pip install -r requirements-dev.txt  # Includes linter, formatter, etc.
pre-commit install  # Auto-format code on commit

📬 Community

• Discord Server: Chat with developers
• Twitter @NLang4438: Updates & announcements

Note: This is an early release (v0.1.0) with known bugs—see GitHub for updates!