keras2graph

Convert Keras model to Intel Movidius NCS graph file, using NCSDK 2.x.

Intel Movidius Neural Compute Stick

About Intel Movidius Neural Compute Stick:

Official Web Page: developer.movidius.com
Intel® Movidius™ Neural Compute SDK(NCSDK 2.x): github

Keras Model to NCS Graph:

keras2graph process turn used TensorFlow backend Keras models to NCS graph. Converts Keras model and weights to Intel Movidius technology internal compiled format.

Keras Model -> TensorFlow Session -> TensorFlow Meta File -> NCS Graph

Important: You need to install NCSDK 2.x.

Command usage:

keras2graph.py [-h] [--output_graph OUTPUT_GRAPH] [--take_tf_files]  
               [--shaves SHAVES]  
               [--do_profile] [--verbose]  
               model_path model_in model_out weights_path

This command creates and saves graph file as <OUTPUT_GRAPH>, default is './graph'.

Arguments:

model_path: Model location. Json file. Data Type: String
model_in: Name of model's input layer. Data Type: String
model_out: Name of model's output layer. Data Type: String
weights_path: Model weights files location. Data Type: String
output_graph: Location of output Intel Movidius technology internal compiled format graph. Data Type: String
take_tf_files: If you don't want to delete created TensorFlow model files(meta file inside), select False. Data Type: Bool
shaves: The number of available SHAVEs depends on your neural compute device. Data Type: Int
do_profile: Run mvNCProfile before mvNCCompile. Data Type: Bool
verbose: Show model details. Data Type: Bool

Example Run Command:

python3 keras2graph.py keras_model/model_CNN_1.json conv2d_1_input activation_5/Softmax keras_model/weights-best_CNN_1.h5 -o keras_model/CNN.graph --shaves=12 --do_profile

Check result:

Use pred_cpu.py for prediction on CPU, and use pred_movidius.py for prediction on Intel Movidius.

for Intel Movidius:

python3 pred_movidius.py keras_model/CNN.graph ./img 64

for CPU:

python3 pred_cpu.py keras_model/model_CNN_1.json keras_model/weights-best_CNN_1.h5 ./img 64

Important Notes:

Install necessary modules with sudo pip3 install -r requirements.txt command.

tak-s/keras2graph