Looking at the official docs: What do you see? The usual fare? Now, guess what: This is a bazel-free zone. We use CMake here!
This collection contains reliable and dead-simple examples to use TensorFlow in C, C++, Go and Python: load a pre-trained model or compile a custom operation with or without CUDA. All builds are tested against the most recent stable TensorFlow version and rely on CMake with a custom FindTensorFlow.cmake. This cmake file includes common work arounds for bugs in specific TF versions.
TensorFlow | Status |
---|---|
1.14.0 | |
1.13.1 | |
1.12.0 | |
1.11.0 | |
1.10.0 | |
1.9.0 |
The repository contains the following examples.
Example | Explanation |
---|---|
custom operation | build a custom operation for TensorFLow in C++/CUDA (requires only pip) |
inference (C++) | run inference in C++ |
inference (C) | run inference in C |
inference (Go) | run inference in Go |
event writer | write event files for TensorBoard in C++ |
keras cpp-inference example | run a Keras-model in C++ |
simple example | create and run a TensorFlow graph in C++ |
resize image example | resize an image in TensorFlow with/without OpenCV |
This example illustrates the process of creating a custom operation using C++/CUDA and CMake. It is not intended to show an implementation obtaining peak-performance. Instead, it is just a boilerplate-template.
user@host $ pip install tensorflow-gpu --user # solely the pip package is needed
user@host $ cd custom_op/user_ops
user@host $ cmake .
user@host $ make
user@host $ python test_matrix_add.py
user@host $ cd ..
user@host $ python example.py
This example illustrates the process of loading an image (using OpenCV or TensorFlow), resizing the image saving the image as a JPG or PNG (using OpenCV or TensorFlow).
user@host $ cd examples/resize
user@host $ export TENSORFLOW_BUILD_DIR=...
user@host $ export TENSORFLOW_SOURCE_DIR=...
user@host $ cmake .
user@host $ make
There are two examples demonstrating the handling of TensorFlow-Serving: using a vector input and using an encoded image input.
server@host $ CHOOSE=basic # or image
server@host $ cd serving/${CHOOSE}/training
server@host $ python create.py # create some model
server@host $ cd serving/server/
server@host $ ./run.sh # start server
# some some queries
client@host $ cd client/bash
client@host $ ./client.sh
client@host $ cd client/python
# for the basic-example
client@host $ python client_rest.py
client@host $ python client_grpc.py
# for the image-example
client@host $ python client_rest.py /path/to/img.[png,jpg]
client@host $ python client_grpc.py /path/to/img.[png,jpg]
Create a model in Python, save the graph to disk and load it in C/C+/Go/Python to perform inference. As these examples are based on the TensorFlow C-API they require the libtensorflow_cc.so
library which is not shipped in the pip-package (tensorfow-gpu). Hence, you will need to build TensorFlow from source beforehand, e.g.,
user@host $ ls ${TENSORFLOW_SOURCE_DIR}
ACKNOWLEDGMENTS bazel-genfiles configure pip
ADOPTERS.md bazel-out configure.py py.pynano
ANDROID_NDK_HOME bazel-tensorflow configure.py.bkp README.md
...
user@host $ cd ${TENSORFLOW_SOURCE_DIR}
user@host $ ./configure
user@host $ # ... or whatever options you used here
user@host $ bazel build -c opt --copt=-mfpmath=both --copt=-msse4.2 --config=cuda //tensorflow:libtensorflow.so
user@host $ bazel build -c opt --copt=-mfpmath=both --copt=-msse4.2 --config=cuda //tensorflow:libtensorflow_cc.so
user@host $ export TENSORFLOW_BUILD_DIR=/tensorflow_dist
user@host $ mkdir ${TENSORFLOW_BUILD_DIR}
user@host $ cp ${TENSORFLOW_SOURCE_DIR}/bazel-bin/tensorflow/*.so ${TENSORFLOW_BUILD_DIR}/
user@host $ cp ${TENSORFLOW_SOURCE_DIR}/bazel-genfiles/tensorflow/cc/ops/*.h ${TENSORFLOW_BUILD_DIR}/includes/tensorflow/cc/ops/
We just run a very basic model
x = tf.placeholder(tf.float32, shape=[1, 2], name='input')
output = tf.identity(tf.layers.dense(x, 1), name='output')
Therefore, save the model like you regularly do. This is done in example.py
besides some outputs
user@host $ python example.py
[<tf.Variable 'dense/kernel:0' shape=(2, 1) dtype=float32_ref>, <tf.Variable 'dense/bias:0' shape=(1,) dtype=float32_ref>]
input [[1. 1.]]
output [[2.1909506]]
dense/kernel:0 [[0.9070684]
[1.2838823]]
dense/bias:0 [0.]
user@host $ python python/inference.py
[<tf.Variable 'dense/kernel:0' shape=(2, 1) dtype=float32_ref>, <tf.Variable 'dense/bias:0' shape=(1,) dtype=float32_ref>]
input [[1. 1.]]
output [[2.1909506]]
dense/kernel:0 [[0.9070684]
[1.2838823]]
dense/bias:0 [0.]
user@host $ cd cc
user@host $ cmake .
user@host $ make
user@host $ cd ..
user@host $ ./cc/inference_cc
input Tensor<type: float shape: [1,2] values: [1 1]>
output Tensor<type: float shape: [1,1] values: [2.19095063]>
dense/kernel:0 Tensor<type: float shape: [2,1] values: [0.907068372][1.28388226]>
dense/bias:0 Tensor<type: float shape: [1] values: 0>
user@host $ cd c
user@host $ cmake .
user@host $ make
user@host $ cd ..
user@host $ ./c/inference_c
2.190951
user@host $ go get github.com/tensorflow/tensorflow/tensorflow/go
user@host $ cd go
user@host $ ./build.sh
user@host $ cd ../
user@host $ ./inference_go
input [[1 1]]
output [[2.1909506]]
dense/kernel:0 [[0.9070684] [1.2838823]]
dense/bias:0 [0]