tf_cpp

This toy project demonstrates how to:

• Use TensorFlow's C++ API to generate training and test data sets, in TFRecord format;
• Load the data sets in Python with TFRecordDataset, train a model with Keras and export the model to a H5 file.
• Convert the model in .H5 to Tensorflow's GraphDef.
• Load the model and run inference with TensorFlow's C++ API.

The demo is based on the works in:

• minigo: for configuring TensorFlow in Bazel. ./cc/configure.sh is copied from the project with minor modifications;
• keras_to_tensorflow: for converting a Keras model to a TensorFlow model. ./keras_to_tensorflow.py is copied from the project with minor modifications.
• MNIST: the training and test datasets are generated from the MNIST database.

Run the Demo

To run the demo, you need to install the following:

• Python
• TensorFlow
• Bazel

Configure and Build Tensorflow

chmod +x cc/configure_tf.sh \
./cc/configure_tf.sh

Generate TFRecords from C++

Download the MNIST dataset, extract the files to a work directory, e.g. /tmp/mnist.

work_dir=/tmp/mnist
mkdir -p ${work_dir}

bazel build -c opt cc:gen_dataset

bazel-bin/cc/gen_dataset \
  --images=${work_dir}/train-images-idx3-ubyte \
  --labels=${work_dir}/train-labels-idx1-ubyte \
  --records=${work_dir}/mnist_train.rio

bazel-bin/cc/gen_dataset \
  --images=${work_dir}/t10k-images-idx3-ubyte \
  --labels=${work_dir}/t10k-labels-idx1-ubyte \
  --records=${work_dir}/mnist_test.rio

Train with Keras

python mnist_nn.py \
  --train_records=${work_dir}/mnist_train.rio \
  --test_records=${work_dir}/mnist_test.rio \
  --use_cnn={true|false} \
  --trained_model=${work_dir}/my_model.h5 \
  --first_layer_name=mnist

Choose a unique value for the flag --first_layer_name. The input node's name will be "${first_layer_name}_input" and run_model.cc needs to know this name.

Convert the Keras Model to TensorFlow Format

python keras_to_tensorflow.py \
  --input_model=${work_dir}/my_model.h5 \
  --output_model=${work_dir}/my_model.pb \
  --output_nodes_prefix=mnist_output/

If the model has only one output, the final output node's name will be "${output_nodes_prefix}0$" and run_model.cc will use this name.

Run the Model in C++

bazel build -c opt cc:run_model

bazel-bin/cc/run_model \
  --model=./tmp/my_model.pb \
  --input_layer_name=mnist_input \
  --output_layer_name=mnist_output/0 \
  --v=1

If the program complains about any compatibility issue, try to re-install tensorflow with the package built from cc/configure_tf.sh.

pip uninstall ${your-tensorflow-version}
pip install ${tmp_pkg_dir}/tensorflow-{your-version-tags}.whl