MMdnn - Utilization for Keras2Tensorflow

MMdnn is a model converter and enables us to transform model from one framework to another. This repository introduces the utilization of MMdnn and provides a simple GUI for inference task of image classification.

Firstly, download weights(.h5) and model structure(.json) form Keras and then convert them to Intermediate Representation(IR). You can visualize IR if you'd like by MMdnn model visualizer .

Secondly, convert IR to Tensorflow weights and model structure(.ckpt) as well as a runable python module(.py).

Thirdly, create module try_inference_tf for inference by module inference_tf.

Last but not least, use python module try_inference_tf for inference.

All you need to do is to set : 1) the path of your folder, 2) which model you'd like to use, 3) the path of image to be infered, 4) your self-defined name for this inference.

If you have already finished evironment settings (step 1&2 of the following paragraph) and you'd like to run it by only using one command, you can modify the file infer_keras2tf.sh and run below command (for step 3-10):

sh infer_keras2tf.sh  # for linux

For the first time you run this code, it's nessasary to use the above command OR step 1-10 of the following paragraph. Afterward, you could do inference derectly by the command such like:

python -c "from try_inference_tf import inference;print(inference('cat1.jpeg','inception_v3'))"

And get the inference result by model 'inception_v3' for image 'cat1.jpeg':

72.76% : tiger cat

13.49% : tabby, tabby cat

08.87% : Egyptian cat

00.84% : plastic bag

00.33% : lynx, catamount

Or, you could do inference by a simple GUI. Use below command to open GUI.

python GUI_sim_1.py

Change model and ImageURL by editting the texts. Click Inference button to obtain the the inference result:

Step-by-Step Commands

Step 1. Confirm Toolkit

• keras >= 2.0
• tensorflow >= 1.4
• Pillow >= 4.2
• Numpy >= 1.14
• Tkinter == '$Revision: 81008 $'
• Develop under Ubuntu 16.04

Step 2. Install MMdnn

https://github.com/Microsoft/MMdnn

sudo pip install -U git+https://github.com/Microsoft/MMdnn.git@master 
#sudo pip install https://github.com/Microsoft/MMdnn/releases/download/0.1.2/mmdnn-0.1.2-py2.py3-none-any.whl # This version has error!! #sudo pip uninstall MMdnn

It returns

Successfully installed mmdnn-0.1.2 protobuf-3.5.1 setuptools-38.5.1

Step 3. Decide the path name

• Download this repository as a new folder.
• Open terminal in this folder. Run commands below step by step.

selfDefinedFileName="try" #<your self-defined name for this inference>
mainPath="/home/phoebehuang/20180214/" #<your folder path>
cd ${mainPath}
oriModelPath=${mainPath}"ori_model_meta/" 
IRpath=${mainPath}"gen_pb_json_npy/"
HWIRpath=${mainPath}"json_hw/" 
CodePath=${mainPath}"gen_code/";echo "" >> ${CodePath}__init__.py
genModelPath=${mainPath}"gen_model/"
mkdir "ori_model_meta/" "gen_pb_json_npy/" "gen_code/"  "gen_model/" "json_hw/"

Step 4. Decide the model name

Choose which model you'd like to use.

model="inception_v3"  
#model="vgg16"
#model="vgg19"
#model="resnet"
#model="mobilenet"
#model="xception"

Step 5. Download the pre-trained Keras models

python -m mmdnn.conversion.examples.keras.extract_model -n ${model}
#move model to a certain folder
mv imagenet_${model}.h5 ${oriModelPath}; mv imagenet_${model}.json ${oriModelPath}

It returns

Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.5/inception_v3_weights_tf_dim_ordering_tf_kernels.h5

96116736/96112376 [==============================] - 129s 1us/step

96124928/96112376 [==============================] - 129s 1us/step

Network structure is saved as [imagenet_inception_v3.json].

Network weights are saved as [imagenet_inception_v3.h5].

Step 6. Convert model to IR

#Convert architecture from Keras to IR
python -m mmdnn.conversion._script.convertToIR -f keras -d ${IRpath}${selfDefinedFileName}_${model} -n ${oriModelPath}imagenet_${model}.json
#Convert model (including architecture and weights) from Keras to IR
python -m mmdnn.conversion._script.convertToIR -f keras -d ${IRpath}${selfDefinedFileName}_${model} -n ${oriModelPath}imagenet_${model}.json -w ${oriModelPath}imagenet_${model}.h5

It returns :

Network file [ori_model_meta/imagenet_inception_v3.json] and [ori_model_meta/imagenet_inception_v3.h5] is loaded successfully.

IR network structure is saved as [gen_pb_json_npy/try_inception_v3.json].

IR network structure is saved as [gen_pb_json_npy/try_inception_v3.pb].

IR weights are saved as [gen_pb_json_npy/try_inception_v3.npy].

(Optional) Visualization:

Open the MMdnn model visualizer and choose .json file under folder ${IRpath}

Step 7. Convert models from IR to Tensorflow code snippet

python -m mmdnn.conversion._script.IRToCode -f tensorflow --IRModelPath ${IRpath}${selfDefinedFileName}_${model}.pb --IRWeightPath ${IRpath}${selfDefinedFileName}_${model}.npy --dstModelPath ${CodePath}${selfDefinedFileName}_tensorflow_${model}.py
# test 
#python -m mmdnn.conversion.examples.tensorflow.imagenet_test -s tensorflow -p ${model} -n ${selfDefinedFileName}_tensorflow_${model} -w ${IRpath}${selfDefinedFileName}_${model}.npy

It returns :

Parse file [gen_pb_json_npy/try_inception_v3.pb] with binary format successfully. Target network code snippet is saved as [gen_code/try_tensorflow_inception_v3.py].

Step 8. Convert models from IR to Tensorflow model

cd ${CodePath}
python -m mmdnn.conversion.examples.tensorflow.imagenet_test -n ${selfDefinedFileName}_tensorflow_${model}.py -w ${IRpath}${selfDefinedFileName}_${model}.npy --dump ${genModelPath}${selfDefinedFileName}_tf_${model}.ckpt
cd ${mainPath}

It returns :

Tensorflow file is saved as [/home/phoebehuang/itri/20180214/gen_model/try_tf_inception_v3.ckpt], generated by [try_tensorflow_inception_v3.py] and [/home/phoebehuang/itri/20180214/gen_pb_json_npy/try_inception_v3.npy].

Step 9. Create Inference Python Module

python inference_tf.py -n ${selfDefinedFileName}

It will create ${selfDefinedFileName}_inference_tf.py, which is try_inference_tf.py in this case.

Step 10. Use Tensorflow model for Inference

Set image path, can be local or a URL.

inferImgPath="cat.jpeg" # you can change to the other local path or a URL
python -c "from try_inference_tf import inference;print(inference('${inferImgPath}','${model}' ))"

And get the inference result by model 'inception_v3' for image 'cat.jpeg':

92.01% : Egyptian cat

06.03% : tabby, tabby cat

00.81% : tiger cat

00.26% : window screen

00.04% : window shade

Or, you could do inference by a simple GUI. Use below command to open GUI.

python GUI_sim_1.py

Change model and ImageURL by editting the texts. Click Inference button to obtain the the inference result: