MMdnn

A comprehensive, cross-framework solution to convert, visualize and diagnose deep neural network models. The "MM" in MMdnn stands for model management and "dnn" is an acronym for deep neural network.

Typically, people use deep neural networks with following steps:

--------------                                               --------------
| Find model | --------------------------------------------> | Deployment |
--------------      |                                        --------------
                    |                                            ^    ^
                    |        --------------                      |    |
                    -------> | Conversion | ----------------------    |
                             --------------                           |
                                   |                                  |
                                   |           -----------            |
                                   ----------> | Retrain | ------------
                                               -----------

In MMdnn, we focus on helping users handle their work better.

• Find model

  • We provide a model collection to help you find some popular models.
  • We provide a model visualizer to display the network architecture more intuitively.

• Conversion

  • We implement an universal convertor to convert DNN models between frameworks, which means you can train on one framework and deploy on another.

• Retrain

  • In convertor, we can generate some training/inference code snippet to simplify the retrain/evaluate work.

• Deployment

  • We provide some guidelines to help you deploy your models to other hardware platform.

    • Android
    • Serving

  • We provide a guide to help you accelerate inference with TensorRT.

    • TensorRT

This project is designed and developed by Microsoft Research (MSR). We also encourage researchers and students leverage this project to analyse DNN models. We welcome any new ideas to extend this project.

Installation

Install manually

You can get stable version of MMdnn by

pip install mmdnn

And make sure to have Python installed or you can try the newest version by

pip install -U git+https://github.com/Microsoft/MMdnn.git@master

Install with docker image

MMdnn provides a docker image, which packages MMdnn, deep learning frameworks we support and other dependencies in one image. You can easily get the image in several steps:

1. Install Docker Community Edition(CE)

   Learn more about how to install docker

2. Pull MMdnn docker image

   docker pull mmdnn/mmdnn:cpu.small

3. Run image in interactive mode

   docker run -it mmdnn/mmdnn:cpu.small

Features

Model Conversion

Across the industry and academia, there are a number of existing frameworks available for developers and researchers to design a model, where each framework has its own network structure definition and saving model format. The gaps between frameworks impede the inter-operation of the models.

We provide a model converter to help developers convert models between frameworks through an intermediate representation format.

Support frameworks

[Note] You can click the links to get detail README of each framework

• Caffe
• Microsoft Cognitive Toolkit (CNTK)
• CoreML
• Keras
• MXNet
• ONNX (Destination only)
• PyTorch
• TensorFlow (Experimental) (We highly recommend you read the README of TensorFlow first)
• DarkNet (Source only, Experiment)

Tested models

The model conversion between currently supported frameworks is tested on some ImageNet models.

Models	Caffe	Keras	TensorFlow	CNTK	MXNet	PyTorch	CoreML	ONNX
VGG 19	√	√	√	√	√	√	√	√
Inception V1	√	√	√	√	√	√	√	√
Inception V3	√	√	√	√	√	√	√	√
Inception V4	√	√	√	o	√	√	√	√
ResNet V1	×	√	√	o	√	√	√	√
ResNet V2	√	√	√	√	√	√	√	√
MobileNet V1	×	√	√	o	√	√	√	√
MobileNet V2	×	√	√	o	√	√	√	√
Xception	√	√	√	o	×	√	√	√
SqueezeNet	√	√	√	√	√	√	√	√
DenseNet	√	√	√	√	√	√	√	√
NASNet	x	√	√	o	√	√	√	x
ResNext	√	√	√	√	√	√	√	√
voc FCN			√	√
Yolo3		√		√

Usage

One command to achieve the conversion. Using TensorFlow ResNet V2 152 to PyTorch as our example.

$ mmdownload -f tensorflow -n resnet_v2_152 -o ./
$ mmconvert -sf tensorflow -in imagenet_resnet_v2_152.ckpt.meta -iw imagenet_resnet_v2_152.ckpt --dstNodeName MMdnn_Output -df pytorch -om tf_resnet_to_pth.pth

Done.

On-going frameworks

• Torch7 (help wanted)
• Chainer (help wanted)

On-going Models

• Face Detection
• Semantic Segmentation
• Image Style Transfer
• Object Detection
• RNN

---

Model Visualization

You can use the MMdnn model visualizer and submit your IR json file to visualize your model. In order to run the commands below, you will need to install requests, keras, and TensorFlow using your favorite package manager.

Use the Keras "inception_v3" model as an example again.

1. Download the pre-trained models

$ mmdownload -f keras -n inception_v3

2. Convert the pre-trained model files into intermediate representation

$ mmtoir -f keras -w imagenet_inception_v3.h5 -o keras_inception_v3

3. Open the MMdnn model visualizer and choose file keras_inception_v3.json

---

Examples

Official Tutorial

• Keras "inception V3" to CNTK and related issue

• TensorFlow slim model "ResNet V2 152" to PyTorch

• Mxnet model "LResNet50E-IR" to TensorFlow and related issue

Users' Examples

• MXNet "ResNet-152-11k" to PyTorch

• MXNet "ResNeXt" to Keras

• TensorFlow "ResNet-101" to PyTorch

• TensorFlow "mnist mlp model" to CNTK

• TensorFlow "Inception_v3" to MXNet

• Caffe "voc-fcn" to TensorFlow

• Caffe "AlexNet" to TensorFlow

• Caffe "inception_v4" to TensorFlow

• Caffe "VGG16_SOD" to TensorFlow

• Caffe "SqueezeNet v1.1" to CNTK

---

Contributing

Intermediate Representation

The intermediate representation stores the network architecture in protobuf binary and pre-trained weights in NumPy native format.

[Note!] Currently the IR weights data is in NHWC (channel last) format.

Details are in ops.txt and graph.proto. New operators and any comments are welcome.

Frameworks

We are working on other frameworks conversion and visualization, such as PyTorch, CoreML and so on. We're investigating more RNN related operators. Any contributions and suggestions are welcome! Details in Contribution Guideline.

License

Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.microsoft.com.

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.

Authors

Cheng CHEN (Microsoft Research Asia): Project Manager; Caffe, CNTK, CoreML Emitter, Keras, MXNet, TensorFlow

Jiahao YAO (Peking University): CoreML, MXNet Emitter, PyTorch Parser; HomePage

Ru ZHANG (Chinese Academy of Sciences): CoreML Emitter, DarkNet Parser, Keras, TensorFlow frozen graph Parser; Yolo and SSD models; Tests

Yuhao ZHOU (Shanghai Jiao Tong University): MXNet

Tingting QIN (Microsoft Research Asia): Caffe Emitter

Tong ZHAN (Microsoft): ONNX Emitter

Qianwen WANG (Hong Kong University of Science and Technology): Visualization

Acknowledgements

Thanks to Saumitro Dasgupta, the initial code of caffe -> IR converting is references to his project caffe-tensorflow.