TensorFlow on Debian/sid (including Keras via R)

I have been struggling with getting TensorFlow running on Debian/sid for quite some time. The main problem is that the CUDA libraries installed by Debian are CUDA 9.1 based, and the precompiled pip installable TensorFlow packages require CUDA 9.0 which resulted in an unusable installation. But finally I got around and found all the pieces.

Step 1: Install CUDA 9.0

The best way I found was going to the CUDA download page, select Linux, then x86_64, then Ubuntu, then 17.04, and finally deb (network>. In the text appearing click on the download button to obtain currently cuda-repo-ubuntu1704_9.0.176-1_amd64.deb.

After installing this package as root with

dpkg -i cuda-repo-ubuntu1704_9.0.176-1_amd64.deb

the nvidia repository signing key needs to be added

apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1704/x86_64/7fa2af80.pub

and finally install the CUDA 9.0 libraries (not all of cuda-9-0 because this would create problems with the normally installed nvidia libraries):

apt-get update
apt-get install cuda-libraries-9-0

This will install lots of libs into /usr/local/cuda-9.0 and add the respective directory to the ld.so path by creating a file /etc/ld.so.conf.d/cuda-9-0.conf.

Step 2: Install CUDA 9.0 CuDNN

One difficult to satisfy dependency are the CuDNN libraries. In our case we need the version 7 library for CUDA 9.0. To download these files one needs to have a NVIDIA developer account, which is quick and painless. After that go to the CuDNN page where one needs to select Download for CUDA 9.0 and then cuDNN v7.2.1 Runtime Library for Ubuntu 16.04 (Deb).

This will download a file libcudnn7_7.2.1.38-1+cuda9.0_amd64.deb which needs to be installed with dpkg -i libcudnn7_7.2.1.38-1+cuda9.0_amd64.deb.

Step 3: Install Tensorflow for GPU

This is the easiest one and can be done as explained on the TensorFlow installation page using

pip3 install --upgrade tensorflow-gpu

This will install several other dependencies, too.

Step 4: Check that everything works

Last but not least, make sure that TensorFlow can be loaded and find your GPU. This can be done with the following one-liner, and in my case gives the following output:

$ python3 -c "import tensorflow as tf; sess = tf.Session() ; print(tf.__version__)"
2018-09-11 16:30:27.075339: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-09-11 16:30:27.143265: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:897] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-09-11 16:30:27.143671: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1405] Found device 0 with properties: 
name: GeForce GTX 1050 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.4175
pciBusID: 0000:01:00.0
totalMemory: 3.94GiB freeMemory: 3.85GiB
2018-09-11 16:30:27.143702: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1484] Adding visible gpu devices: 0
2018-09-11 16:30:27.316389: I tensorflow/core/common_runtime/gpu/gpu_device.cc:965] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-09-11 16:30:27.316432: I tensorflow/core/common_runtime/gpu/gpu_device.cc:971]      0 
2018-09-11 16:30:27.316439: I tensorflow/core/common_runtime/gpu/gpu_device.cc:984] 0:   N 
2018-09-11 16:30:27.316595: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1097] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 3578 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1050 Ti, pci bus id: 0000:01:00.0, compute capability: 6.1)
1.10.1         
$

Addendum: Keras and R

With the above settled, the installation of Keras can be done via

apt-get install python3-keras

and this should pick up the TensorFlow backend automatically.

For R there is a Keras library that can be installed without

install.packages('keras')

on the R command line (as root).

After that running a simple MNIST code example should use your GPU from R (taken from Deep Learning with R from Manning Publications):

library(keras)
mnist <- dataset_mnist()
train_images <- mnist$train$x
train_labels <- mnist$train$y
test_images <- mnist$test$x
test_labels <- mnist$test$y
network <- keras_model_sequential() %>%
  layer_dense(units = 512, activation = "relu", input_shape = c(28 * 28)) %>%
  layer_dense(units = 10, activation = "softmax")
network %>% compile(
  optimizer = "rmsprop",
  loss = "categorical_crossentropy",
  metrics = c("accuracy")
)
train_images <- array_reshape(train_images, c(60000, 28 * 28))
train_images <- train_images / 255
test_images <- array_reshape(test_images, c(10000, 28 * 28))
test_images <- test_images / 255
train_labels <- to_categorical(train_labels)
test_labels <- to_categorical(test_labels)
network %>% fit(train_images, train_labels, epochs = 5, batch_size = 128)
metrics <- network %>% evaluate(test_images, test_labels)
metrics