carbono

a micro-library for feedforward neural networks that works everywhere:

latest changes

• ✦ for more information on the changes, please take a look at the changelog
• v7: now default exporting mode is json, with optional binary compression for larger neural networks (warning: v8 beta breaking changes due to quantization, only js version)
• v6: dramatically optimized file size; for instance, 18k parameters model was reduced from 376kb to 73kb
• v5: new metadata fields added; removed unnecesary "layers" metadata

Browser & Edge Ready

• Highly optimized models
• Fast browser inference
• Edge-optimized C++ server
• Fast training with Pytorch
• Simple JavaScript API

Build Once, Run Anywhere

// Train in browser
new carbono().layer(2,1).train([{ input:[1,1], output:[0] }])

// Or use PyTorch for larger models
// Then run anywhere - browser, edge, server

inference on edge devices

cd server
g++ -std=c++17 src/server.cpp -I./include -pthread -o server
./server xor_model.json 8080

# later...

curl -X POST http://localhost:8080/predict \
     -H "Content-Type: application/json" \
     -d '{"input": [1, 1]}'

# {"output":[0.0009481541869894417]}