What it does

Fabrik is a web IDE for deep learning. You drag layers around in a browser, wire them together, and export the result to Caffe, Keras, or TensorFlow. It also imports existing models from those frameworks so you can inspect or tweak them visually. The project is built on Django, with Celery handling background tasks and a React frontend.

The interesting bit

The real work isn’t the GUI—it’s the translation layer. Fabrik parses framework-specific formats (Caffe’s .prototxt, Keras JSON, TensorFlow’s pbtxt) into a common graph representation, then serializes back out to whichever format you need. That’s harder than it sounds; the README includes a compatibility matrix showing which of 14 tested models round-trip cleanly between which frameworks.

Key highlights

• Supports import/export across Caffe, Keras, and TensorFlow
• 14 tested reference models including Inception V3/V4, ResNet-101, VGG-16, YoloNet, and VQA
• Docker Compose setup available alongside manual virtualenv instructions
• GitHub OAuth integration for collaborative features
• Hot-reloading webpack dev server for UI work

Caveats

• Requires Python 2.7, which is end-of-life
• Several models (GoogLeNet, SqueezeNet, DenseNet, Pix2Pix, FCN32 Pascal, AllCNN) only round-trip through Caffe; Keras and TensorFlow export is marked as unsupported
• AlexNet’s custom LRN layer needs manual handling when loading exported Keras JSON
• Setup is involved: Postgres, Redis, Celery, Node, and framework-specific dependencies all need configuration

Verdict

Worth a look if you’re teaching deep learning or need to quickly communicate model architecture without passing around protobuf files. Skip it if you want a maintained, batteries-included tool—this is explicitly “under active development” with open issues, and the Python 2.7 dependency is a yellow flag for new projects.