Tensorflow – Keras tensorflow backend does not detect GPU

kerastensorflowtensorflow-gpu

I am running keras with tensorflow backend on linux.
First, I installed tensorflow GPU version by itself, and run the following code to check and found out that it's running on GPU and shows the GPU it's running on, device mapping, etc. The tensorflow I use was from https://storage.googleapis.com/tensorflow/linux/gpu/tensorflow-
0.11.0-cp27-none-linux_x86_64.whl

a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
c = tf.matmul(a, b)
# Creates a session with log_device_placement set to True.
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
# Runs the op.
print(sess.run(c))

Then, I installed keras using conda install keras. I checked conda list and now I have 2 versions of tensorflow (1.1.0 and 0.11.0). I tried import tensorflow as tf which results in: 

2017-07-18 16:35:59.569535: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.
2017-07-18 16:35:59.569629: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.2 instructions, but these are available on your machine and could speed up CPU computations.
2017-07-18 16:35:59.569690: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX instructions, but these are available on your machine and could speed up CPU computations.
2017-07-18 16:35:59.569707: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2017-07-18 16:35:59.569731: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
Device mapping: no known devices.
2017-07-18 16:35:59.579959: I tensorflow/core/common_runtime/direct_session.cc:257] Device mapping:

MatMul: (MatMul): /job:localhost/replica:0/task:0/cpu:0
2017-07-18 16:36:14.369948: I tensorflow/core/common_runtime/simple_placer.cc:841] MatMul: (MatMul)/job:localhost/replica:0/task:0/cpu:0
b: (Const): /job:localhost/replica:0/task:0/cpu:0
2017-07-18 16:36:14.370051: I tensorflow/core/common_runtime/simple_placer.cc:841] b: (Const)/job:localhost/replica:0/task:0/cpu:0
a: (Const): /job:localhost/replica:0/task:0/cpu:0
2017-07-18 16:36:14.370109: I tensorflow/core/common_runtime/simple_placer.cc:841] a: (Const)/job:localhost/replica:0/task:0/cpu:0

I already set CUDA_VISIBLE_DEVICES, which works before keras was installed.
Is this because of the tensorflow version? Can I choose to install 0.11.0 instead of 1.1.0 when installing keras?
If the problem is due to tensorflow not detecting a GPU, how can I solve this issue? I read in this link and it says that tensorflow will automatically run on GPU is it detects one.

Best Answer

Chances are that Keras, depending on a newer version of TensorFlow, has caused the installation of a CPU-only TensorFlow package (tensorflow) that is hiding the older, GPU-enabled version (tensorflow-gpu).

I would upgrade the GPU-enabled version first. Usually you can just do pip install --upgrade tensorflow-gpu, but you have Anaconda-specific instructions in the TensorFlow installation page. Then you can uninstall the CPU-only TensorFlow package with pip uninstall tensorflow. Now import tensorflow as tf should actually import the GPU-enabled package which, as you suggest, should in turn detect your GPU automatically.

Related Topic