Keras2cpp

Keras2cpp is a small library for running trained Keras models from a C++ application without any dependences.

Design goals:

• Compatibility with networks generated by Keras using TensorFlow backend.
• CPU only, no GPU.
• No external dependencies, standard library, C++17.
• Model stored on disk in binary format and can be quickly read.
• Model stored in memory in contiguous block for better cache performance.

Not not layer and activation types are supported yet. Work in progress

Supported Keras layers:

• Dense
• Convolution1D
• Convolution2D
• Convolution3D
• Flatten
• ELU
• Activation
• MaxPooling2D
• Embedding
• LocallyConnected1D
• LocallyConnected2D
• LSTM
• GRU
• CNN
• BatchNormalization

Supported activation:

• linear
• relu
• softplus
• tanh
• sigmoid
• hard_sigmoid
• elu
• softsign
• softmax

Other tasks:

• Create unit tests
• Create Makefile
• Code refactoring (in progress)

The project is compatible with Keras 2.x (all versions) and Python 3.x

Example

python_model.py:

import numpy as np
from keras import Sequential
from keras.layers import Dense

#create random data
test_x = np.random.rand(10, 10).astype('f')
test_y = np.random.rand(10).astype('f')
model = Sequential([
    Dense(1, input_dim=10)
])
model.compile(loss='mse', optimizer='adam')

#train model by 1 iteration
model.fit(test_x, test_y, epochs=1, verbose=False)

#predict
data = np.array([[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])
prediction = model.predict(data)
print(prediction)

#save model
from keras2cpp import export_model
export_model(model, 'example.model')

cpp_model.cc:

#include "src/model.h"

using keras2cpp::Model;
using keras2cpp::Tensor;

int main() {
    // Initialize model.
    auto model = Model::load("example.model");

    // Create a 1D Tensor on length 10 for input data.
    Tensor in{10};
    in.data_ = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9};

    // Run prediction.
    Tensor out = model(in);
    out.print();
    return 0;
}

How to build and run

Tested with Keras 2.2.1, Python 3.6

$ git clone https://github.com/gosha20777/keras2cpp.git
$ cd keras2cpp
$ mkdir build && cd build
$ python3 ../python_model.py
[[-1.85735667]]

$ cmake ..
$ cmake --build .
$ ./keras2cpp
[ -1.857357 ]

License

MIT

Similar projects

I found another similar projects on Github:

• https://github.com/pplonski/keras2cpp/;
• https://github.com/moof2k/kerasify
• https://github.com/Dobiasd/frugally-deep

But It works only with Keras 1 and didn’t work for me. That's why I wrote my own implementation.