Image Analogies on Ubuntu 14.04 with CUDA and Keras
This is a cleaned-up archival note from 2016. The commands below belong to the CUDA 7.5, cuDNN 5.0, Theano and early TensorFlow era, so treat them as historical context rather than current setup advice.
In 2016, I came across a tweet that mentioned the image-analogies repository.
It was an implementation of the Image Analogies paper and the Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis paper. It could produce some interesting visual effects, so I decided to try it and record the setup.
If you wanted to experiment with image analogies at the time, you needed either TensorFlow or Theano, plus Keras, which could work with both backends. I tried both TensorFlow and Theano, but used Theano in the end. I also decided to use CUDA because I had an NVIDIA GeForce GTX 960.
Setting all of this up on Ubuntu 14.04 was not a trivial process, at least not for me, so I wrote down the steps.
All installation steps below were tested on a fresh Ubuntu 14.04 LTS installation, available from the Ubuntu 14.04 release archive.
Installing CUDA Toolkit 7.5 and cuDNN 5.0
$ cd ~/Downloads
$ wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1404/x86_64/cuda-repo-ubuntu1404_7.5-18_amd64.deb
$ sudo dpkg -i cuda-repo-ubuntu1404_7.5-18_amd64.deb
$ sudo apt-get update
If you try to execute the next command:
$ sudo apt-get install cuda
you will probably get the following error:
The following packages have unmet dependencies.
cuda : Depends: cuda-7-5 (= 7.5-18) but it is not going to be installed
E: Unable to correct problems, you have held broken packages.
If this happens, add the additional repository and install the needed packages explicitly:
$ sudo add-apt-repository ppa:xorg-edgers/ppa
$ sudo apt-get update
$ sudo apt-get install cuda libcheese-gtk23 libcheese7 libclutter-1.0-0 libclutter-gtk-1.0-0 libcogl15
You also need to set LD_LIBRARY_PATH. To persist the change, add the same line to your ~/.bashrc profile.
$ export LD_LIBRARY_PATH=/usr/local/cuda-7.5/targets/x86_64-linux/lib/:$LD_LIBRARY_PATH
This is a little unfortunate, but without LD_LIBRARY_PATH you will get the following error when you try to compile something with CUDA:
Failed to compile cuda_ndarray.cu: libcublas.so.7.5: cannot open shared object file: No such file or directory
Reboot the system so it uses the NVIDIA driver installed with the CUDA package.
$ sudo reboot
After rebooting, check the NVIDIA driver status. You should get similar output.
$ nvidia-smi
+------------------------------------------------------+
| NVIDIA-SMI 352.99 Driver Version: 352.99 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
|===============================+======================+======================|
| 0 GeForce GTX 960 Off | 0000:01:00.0 On | N/A |
| 0% 33C P8 7W / 160W | 373MiB / 4095MiB | 0% Default |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: GPU Memory |
| GPU PID Type Process name Usage |
|=============================================================================|
| 0 1180 G /usr/bin/X 185MiB |
| 0 2105 G compiz 172MiB |
| 0 2659 G /usr/lib/firefox/firefox 1MiB |
+-----------------------------------------------------------------------------+
You might also want to check the NVIDIA kernel and client versions. They should match.
$ sudo dpkg --list | grep nvidia
Register on the NVIDIA cuDNN page, download cuDNN and install it using the following steps. I used cuDNN 5.0 for CUDA 7.5 because, at the time, Theano did not officially support cuDNN 5.1.
$ cd ~/Downloads
$ tar xvzf cudnn-7.5-linux-x64-v5.0-ga.tgz
$ sudo cp cuda/include/cudnn.h /usr/local/cuda/include
$ sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
$ sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
Installing TensorFlow
You can skip this section if you want to use Theano. To proceed with TensorFlow, first install JDK 8.
$ sudo add-apt-repository ppa:webupd8team/java
$ sudo apt-get update
$ sudo apt-get install oracle-java8-installer
To compile TensorFlow from source, install Bazel.
$ echo "deb [arch=amd64] http://storage.googleapis.com/bazel-apt stable jdk1.8" | sudo tee /etc/apt/sources.list.d/bazel.list
$ curl https://storage.googleapis.com/bazel-apt/doc/apt-key.pub.gpg | sudo apt-key add -
$ sudo apt-get update && sudo apt-get install bazel
$ sudo apt-get upgrade bazel
Install the additional dependencies.
$ sudo apt-get install python-numpy swig python-dev python-wheel git
Clone the TensorFlow repository from GitHub.
$ cd ~
$ git clone https://github.com/tensorflow/tensorflow.git
$ cd tensorflow
Configure the build as recommended in the old TensorFlow 0.10 setup guide and build TensorFlow.
$ ./configure
$ ./bazel build -c opt --config=cuda
Check whether the build was successful by running an example.
$ bazel-bin/tensorflow/cc/tutorials_example_trainer --use_gpu
Build a pip package and install it.
$ bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
$ sudo pip install /tmp/tensorflow_pkg/tensorflow-0.10.0rc0-cp27-none-linux_x86_64.whl
Check that the TensorFlow package is discoverable.
$ cd ~
$ python -c "import tensorflow;print(tensorflow.__version__)"
Installing Theano
$ sudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev git
$ export THEANO_FLAGS='cuda.root=/usr/local/cuda/,device=gpu,floatX=float32'
THEANO_FLAGS will not be persisted, so you may need to modify .theanorc as recommended in the Theano configuration documentation.
Installing Keras
$ sudo apt-get install libblas-dev liblapack-dev libyaml-cpp-dev gfortran
$ sudo pip install keras
Now configure Keras to use either Theano:
$ echo '{"epsilon": 1e-07, "floatx": "float32", "backend": "theano"}' >> ~/.keras/keras.json
or TensorFlow:
$ echo '{"epsilon": 1e-07, "floatx": "float32", "backend": "tensorflow"}' >> ~/.keras/keras.json
Installing Image Analogies
$ sudo apt-get install libhdf5-dev
$ sudo pip install neural-image-analogies
Clone the image-analogies repository.
$ cd ~
$ git clone https://github.com/awentzonline/image-analogies.git
Download the VGG16 weights.
$ cd image-analogies/examples
$ wget https://github.com/awentzonline/image-analogies/releases/download/v0.0.5/vgg16_weights.h5
Try to run an example.
$ ./sugar-skull.sh ./images/sugarskull-B.jpg example
To check whether the script uses the GPU, run the following command in another terminal window and watch the NVIDIA graphics card stats.
$ watch -n 0.5 nvidia-smi
RSS - Posts