/mnist-NCS

Practice NCS using minst dataset with Keras

Primary LanguageJupyter Notebook

Intel® Movidius™ NCS MNIST example

Practice NCS using MNIST dataset with Keras

Train a simple CNN for MNIST using script

$ python3 train-mnist.py

Train a simple CNN for MNIST using jupyter

train-mnist.ipynb

Convert Keras model to Tensorflow model using script (model.json and weights.h5 file)

$ python3 convert-mnist-json-h5.py

Convert Keras model to Tensorflow model using jupyter

convert-mnist-json-h5.ipynb

Convert Keras model to Tensorflow model using script (model.h5 file)

$ python3 convert-mnist-only-h5.py

Convert Keras model to Tensorflow model using jupyter

convert-mnist-only-h5.ipynb

Compile MNIST model using mvNC Toolkit

$ mvNCCompile TF_Model/tf_model.meta -in=conv2d_1_input -on=dense_2/Softmax

Refer: https://movidius.github.io/ncsdk/tools/compile.html

If ImportError: /usr/local/lib/python3.5/dist-packages/pygraphviz/_graphviz.cpython-35m-x86_64-linux-gnu.so: undefined symbol: Agundirected when you using NCSDK v2.x: You should force reinstall your pygraphviz with direct path. Install command below:

$ sudo -H pip3 install --force-reinstall pygraphviz --install-option="--include-path=/usr/include/graphviz" --install-option="--library-path=/usr/lib/graphviz/"

Check, Profile model using mvNC Toolkit

$ mvNCCheck TF_Model/tf_model.meta -in=conv2d_1_input -on=dense_2/Softmax
$ mvNCProfile TF_Model/tf_model.meta -in=conv2d_1_input -on=dense_2/Softmax

If tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'conv2d_1_input' with dtype float and shape [?,28,28,1] occur on execute command above, please edit ncsdk source in /usr/local/bin/ncsdk/Controllers/TensorFlowParser.py line 1059, add a feed_dict to eval:

# desired_shape = node.inputs[1].eval() 
desired_shape = node.inputs[1].eval(feed_dict={inputnode + ':0' : input_data}) 

CAUTION:Graph file(blob) compiled by NCSDK 1.x not support NCSDK 2.x!!

Do prediction on a random image using NCSDK 1.x if you want use mnist.load_data() provided by TF, you should remark line 2,8~11 and edit line 6 or you must install mnist from PyPi using $pip3 install mnist .

$ python3 predict-mnist-ncsdk1.py

Do prediction on a random image using NCSDK 2.x

$ python3 predict-mnist-ncsdk2.py

or run predict-mnist-ncsdk*.py file directly:

$ chmod +x predict-mnist-ncsdk*.py
$ ./predict-mnist-ncsdk*.py

Do prediction on a random image using Keras

$ python3 predict-mnist-keras.py

or run predict-mnist-keras.py file directly:

$ chmod +x predict-mnist-keras.py
$ ./predict-mnist-keras.py

model.json Only contain model graph (Keras Format).

weights.h5 Only contain model weights (Keras Format).

model.h5 Both contain model graph and weights (Keras Format).

graph Intel neural network graph file for ncsdk v2.

References