synapse-lang-core 1.0.0

A declarative ML language that compiles to real PyTorch code

🧠 Synapse Language

A declarative, human-friendly programming language for Machine Learning — compiles to real PyTorch code.

---

✨ What is Synapse?

Instead of writing 50+ lines of complex PyTorch code, you write this:

tensor x = [[1,2],[3,4]]

model TestNet:
    layer dense(8, relu)
    layer dense(4, sigmoid)

train TestNet on x:
    epochs = 2
    lr = 0.001

Synapse reads it, understands it, and trains a real neural network. That's it.

---

🚀 Why Synapse?

• Beginners can build neural networks without learning PyTorch boilerplate
• Researchers can prototype models in seconds
• Teachers can explain ML concepts with clean, readable syntax
• Developers can write ML code that reads like plain English

---

⚙️ How It Works

Your .syn file
     ↓
  Lexer          reads your code word by word
     ↓
  Parser         understands the structure
     ↓
  Transpiler     writes real PyTorch Python
     ↓
  Runtime        executes it and shows results

---

📦 Installation

1. Clone the repo

git clone https://github.com/YOUR_USERNAME/synapse-lang.git
cd synapse-lang

2. Create a virtual environment

# Mac/Linux
python -m venv venv
source venv/bin/activate

# Windows
python -m venv venv
venv\Scripts\activate

3. Install PyTorch (CPU)

pip install torch --index-url https://download.pytorch.org/whl/cpu

---

▶️ Running Synapse

Run a program

python synapse.py run examples/test.syn

See generated Python code

python synapse.py transpile examples/test.syn

See the token stream

python synapse.py tokenize examples/test.syn

See the AST

python synapse.py parse examples/test.syn

Run with full pipeline trace

python synapse.py run examples/test.syn --verbose

---

📝 Language Reference

Tensor Declaration

tensor mydata = [[1, 2, 3], [4, 5, 6]]

Model Declaration

model MyModel:
    layer dense(64, relu)
    layer dense(32, relu)
    layer dense(10, sigmoid)

Train Block

train MyModel on mydata:
    epochs = 10
    lr = 0.001
    optimizer = adam
    loss = mse

Supported Layers

Synapse	PyTorch
dense	nn.LazyLinear
dropout	nn.Dropout
flatten	nn.Flatten
conv2d	nn.LazyConv2d
batchnorm	nn.LazyBatchNorm1d

Supported Activations

relu · sigmoid · tanh · softmax · gelu · selu · leakyrelu

Supported Optimizers

adam · sgd · adamw · rmsprop

Supported Loss Functions

mse · crossentropy · bce · l1 · huber

---

📁 Project Structure

synapse/
├── synapse.py        # CLI entry point
├── lexer.py          # Tokenizer
├── parser.py         # Recursive-descent parser
├── ast_nodes.py      # AST node class hierarchy
├── transpiler.py     # AST → PyTorch transpiler
├── runtime.py        # Execution engine
├── requirements.txt
├── README.md
└── examples/
    ├── test.syn      # Basic example
    ├── advanced.syn  # Multi-layer example
    └── myfirst.syn   # Demo example

---

🧪 Example Output

Running examples/test.syn:

Training TestNet for 2 epoch(s)...
  Epoch 1/2 --- loss: 0.000000
  Epoch 2/2 --- loss: 0.000000
Training complete.

Generated Python (via python synapse.py transpile examples/test.syn):

import torch
import torch.nn as nn
import torch.optim as optim

x = torch.tensor([[1, 2], [3, 4]], dtype=torch.float32)

class TestNet(nn.Module):
    def __init__(self):
        super(TestNet, self).__init__()
        self.network = nn.Sequential(
            nn.LazyLinear(8),
            nn.ReLU(),
            nn.LazyLinear(4),
            nn.Sigmoid(),
        )
    def forward(self, x):
        return self.network(x)

def main():
    _model = TestNet()
    _criterion = nn.MSELoss()
    _optimizer = optim.Adam(_model.parameters(), lr=0.001)
    # ... training loop

---

🔮 Roadmap

• Web playground (write Synapse in the browser)
• VS Code extension with syntax highlighting
• AI blocks — plug in LLMs with one line
• Memory system — vector stores and retrieval
• Multi-model pipelines
• PyPI package (pip install synapse-lang)

---

🤝 Contributing

Pull requests are welcome! Here's how to add a new feature:

1. Add a node class in ast_nodes.py
2. Add a parse rule in parser.py
3. Add a visit_<NodeName> method in transpiler.py

That's all — the architecture handles the rest.

---

📄 License

MIT License — free to use, modify, and distribute.

---

👨‍💻 Author

Built with ❤️ — a declarative ML language for everyone.

"Synapse makes AI accessible — not just for people who memorised PyTorch." Update README with full documentation