Counting 2,129 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Dec. 6, 2017
Jul. 15, 2016


Build Status (License MIT 1.0)


Use Keras models in C++ with ease

Table of contents


Would you like to use/deploy your already-trained Keras models in C++? And would like to avoid linking your application against TensorFlow? Then frugally-deep is exactly for you.


  • is a small header-only library written in modern and pure C++.
  • is very easy to integrate and use.
  • depends only on FunctionalPlus, Eigen and json - also header-only libraries.
  • supports inference (model.predict) not only for sequential models but also for computational graphs with a more complex topology, created with the functional API.
  • re-implements a (small) subset of TensorFlow, i.e. the operations needed to support prediction.
  • utterly ignores even the most powerful GPU in your system and uses only one CPU core. ;-)
  • but is quite fast on one CPU core compared to TensorFlow.

Supported layer types

Layer types typically used in image recognition/generation are supported, making many popular model architectures possible (see Performance section).

  • Add, Concatenate
  • AveragePooling1D/2D, GlobalAveragePooling1D/2D
  • Conv1D/2D, SeparableConv2D
  • Cropping1D/2D, ZeroPadding1D/2D
  • BatchNormalization, Dense, Dropout, Flatten
  • MaxPooling1D/2D, GlobalMaxPooling1D/2D
  • ELU, LeakyReLU, ReLU, SeLU
  • Sigmoid, Softmax, Softplus, Tanh
  • UpSampling1D/2D

Also supported

  • multiple inputs and outputs
  • nested models
  • residual connections
  • shared layers
  • arbitrary complex model architectures / computational graphs

Currently not supported are the following layer types: ActivityRegularization, AlphaDropout, Average, AveragePooling3D, Bidirectional, Conv2DTranspose, Conv3D, ConvLSTM2D, CuDNNGRU, CuDNNLSTM, Cropping3D, DepthwiseConv2D, Dot, Embedding, GaussianDropout, GaussianNoise, GRU, GRUCell, Lambda, LocallyConnected1D, LocallyConnected2D, LSTM, LSTMCell, Masking, Maximum, MaxPooling3D, Multiply, Permute, PReLU, RepeatVector, Reshape, RNN, SimpleRNN, SimpleRNNCell, StackedRNNCells, Subtract, ThresholdedReLU, TimeDistributed, Upsampling3D, any custom layers


  1. Use Keras/Python to build (model.compile(...)), train ( and test (model.evaluate(...)) your model as usual. Then save it to a single HDF5 file using'....h5'). The image_data_format in your model must be channels_last, which is the default when using the TensorFlow backend. Models created with a different image_data_format and other backends are not supported.

  2. Now convert it to the frugally-deep file format with keras_export/

  3. Finally load it in C++ (fdeep::load_model(...)) and use model.predict(...) to invoke a forward pass with your data.

The following minimal example shows the full workflow:

import numpy as np
from keras.layers import Input, Dense
from keras.models import Model

inputs = Input(shape=(4,))
x = Dense(5, activation='relu')(inputs)
predictions = Dense(3, activation='softmax')(x)
model = Model(inputs=inputs, outputs=predictions)
model.compile(loss='categorical_crossentropy', optimizer='nadam')
    np.asarray([[1,2,3,4], [2,3,4,5]]),
    np.asarray([[1,0,0], [0,0,1]]), epochs=10)'keras_model.h5')
python3 keras_export/ keras_model.h5 fdeep_model.json
// main.cpp
#include <fdeep/fdeep.hpp>
int main()
    const auto model = fdeep::load_model("fdeep_model.json");
    const auto result = model.predict(
        {fdeep::tensor3(fdeep::shape3(4, 1, 1), {1, 2, 3, 4})});
    std::cout << fdeep::show_tensor3s(result) << std::endl;

When using a test case (input and corresponding output values) is generated automatically and saved along with your model. fdeep::load_model runs this test to make sure the results of a forward pass in frugally-deep are the same as in Keras.

In order to convert images to fdeep::tensor3 the convenience function tensor3_from_bytes is provided.


Below you can find the durations of one isolated forward pass for some popular models run on a single core of an Intel Core i5-6600 CPU @ 3.30GHz. frugally-deep was compiled with g++ -O3.

| Model             | Keras + TensorFlow | frugally-deep |
| InceptionV3       |             1.11 s |        0.57 s |
| ResNet50          |             0.92 s |        0.39 s |
| VGG16             |             1.15 s |        1.35 s |
| VGG19             |             1.46 s |        1.64 s |
| Xception          |             1.70 s |        0.90 s |

versions: Keras 2.1.1, TensorFlow 1.4.0 (default packages from pip)

Using -march=native when compiling frugally-deep brings the times down to 0.31 s, 0.20 s, 0.56 s, 0.67 s and 0.58 s but this values would have to be compared with the ones resulting from a TensorFlow version with the same optimizations.

Requirements and Installation

A C++14-compatible compiler is needed. Compilers from these versions on are fine: GCC 4.9, Clang 3.7 (libc++ 3.7) and Visual C++ 2015.

You can install frugally-deep using cmake as shown below, or (if you prefer) download the code (and the code of FunctionalPlus), extract it and tell your compiler to use the include directories.

git clone
cd FunctionalPlus
mkdir -p build && cd build
cmake ..
make && sudo make install
cd ../..

sudo apt install mercurial
hg clone
cd eigen
mkdir -p build && cd build
cmake ..
make && sudo make install
cd ../..

git clone
cd json
mkdir -p build && cd build
cmake ..
make && sudo make install
cd ../..

git clone
cd frugally-deep
mkdir -p build && cd build
cmake ..
make && sudo make install
cd ../..

Building the tests (optional) requires doctest. Unit Tests are disabled by default – they are enabled and executed by:

cmake -DUNITTEST=ON ..
make unittest


frugally-deep uses channels_first (depth/channels, height, width) as its image_data_format internally. takes care of all necessary conversions. From then on everything is handled as a float32 tensor with rank 3. Dense layers for example take its input flattened to a shape of (n, 1, 1). This is also the shape you will receive as the output of a final softmax layer for example.

A frugally-deep model is thread-safe, i.e. you can call model.predict on the same model instance from different threads simultaneously. This way you may utilize up to as many CPU cores as you have predictions to make. With model::predict_multi there is a convenience function available to handle the parallelism for you.

Convolution is done using im2col per default. You can disable it in the call of model.predict in case it is not suited for you application, e.g. due to tight memory constraints.


The API of this library still might change in the future. If you have any suggestions, find errors or want to give general feedback/criticism, I'd love to hear from you. Of course, contributions are also very welcome.


Distributed under the MIT License. (See accompanying file LICENSE or at