microsoft/mmdnn

MMdnn is a Python toolkit from Microsoft that lets you convert trained AI models from one deep learning framework to another. Deep learning frameworks are the software environments that researchers and engineers use to build and train neural networks, and each one saves models in its own file format. MMdnn bridges those gaps so that a model trained in one framework can be loaded and used in a different one. The frameworks it supports include TensorFlow, PyTorch, Keras, MXNet, Caffe, CNTK, CoreML, ONNX, and DarkNet. Conversion works by translating a model into a neutral intermediate format first, and then converting from that format into the target framework. This approach means you do not need a direct path between every possible pair of frameworks. During conversion, MMdnn also generates code snippets to help you retrain or run inference with the converted model. Beyond conversion, MMdnn includes a model visualizer that draws the network architecture as a diagram, which can help you understand the structure of a model you did not originally build. It also has a model collection that points to popular pre-trained models you can download and work with. The tool is installable via pip or through a Docker image that bundles MMdnn together with the supported frameworks. There is also documentation covering how to deploy converted models on Android devices, via TensorFlow Serving, or with TensorRT for faster inference on compatible hardware. MMdnn is a research project from Microsoft and is released under the MIT license. Conversion support varies by framework pair, and the README includes a compatibility table showing which model architectures have been tested between which frameworks.