On my local machine I have it set up so that I install these dependencies in a virtual environment while doing ML testing
tensorflow
keras
h5py
requests
pillow
tensorflow-hub
I have scripts that will go through, make a venv, install the requirements, and perform the training. It would be nice if I could use these scripts anywhere, including on the Google Deep Learning Platform VMs, but when I try to install tensorflow in the venv it no longer uses the GPU, so I'm assuming it's no longer using the system install of tensorflow.
I also tried the --system-site-packages flag and that said that tensorflow was already installed but then it stopped using the GPU.
Let's say I do the following
$ virtualenv --sysem-site-packages venv/
$ source venv/bin/activate
$ pip install -r requirements.txt
Collecting tensorflow (from -r requirements.txt (line 1))
Using cached https://files.pythonhosted.org/packages/1a/c4/8cb95df0bf06089014259b25997c3921a87aa08e2cd981417d91ca92f7e9/tensorflow-1.10.1-cp27-cp27mu-manylinux1_x86_64.whl
Requirement already satisfied: keras in /usr/local/lib/python2.7/dist-packages (from -r requirements.txt (line 2)) (2.2.2)
Requirement already satisfied: h5py in /usr/lib/python2.7/dist-packages (from -r requirements.txt (line 3)) (2.7.0)
Requirement already satisfied: requests in /usr/lib/python2.7/dist-packages (from -r requirements.txt (line 4)) (2.12.4)
Requirement already satisfied: pillow in /usr/lib/python2.7/dist-packages (from -r requirements.txt (line 5)) (4.0.0)
Requirement already satisfied: tensorflow-hub in /home/john/.local/lib/python2.7/site-packages (from -r requirements.txt (line 6)) (0.1.1)
Collecting numpy<=1.14.5,>=1.13.3 (from tensorflow->-r requirements.txt (line 1))
Using cached https://files.pythonhosted.org/packages/6a/a9/c01a2d5f7b045f508c8cefef3b079fe8c413d05498ca0ae877cffa230564/numpy-1.14.5-cp27-cp27mu-manylinux1_x86_64.whl
Requirement already satisfied: grpcio>=1.8.6 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (1.14.1)
Requirement already satisfied: protobuf>=3.6.0 in /home/john/.local/lib/python2.7/site-packages (from tensorflow->-r requirements.txt (line 1)) (3.6.1)
Requirement already satisfied: termcolor>=1.1.0 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (1.1.0)
Requirement already satisfied: backports.weakref>=1.0rc1 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (1.0.post1)
Requirement already satisfied: absl-py>=0.1.6 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (0.3.0)
Requirement already satisfied: wheel in ./venv/lib/python2.7/site-packages (from tensorflow->-r requirements.txt (line 1)) (0.31.1)
Requirement already satisfied: tensorboard<1.11.0,>=1.10.0 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (1.10.0)
Requirement already satisfied: six>=1.10.0 in /home/john/.local/lib/python2.7/site-packages (from tensorflow->-r requirements.txt (line 1)) (1.11.0)
Requirement already satisfied: gast>=0.2.0 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (0.2.0)
Requirement already satisfied: mock>=2.0.0 in /usr/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (2.0.0)
Requirement already satisfied: enum34>=1.1.6 in /usr/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (1.1.6)
Requirement already satisfied: astor>=0.6.0 in /usr/local/lib/python2.7/dist-packages (from tensorflow->-r requirements.txt (line 1)) (0.7.1)
Collecting setuptools<=39.1.0 (from tensorflow->-r requirements.txt (line 1))
Using cached https://files.pythonhosted.org/packages/8c/10/79282747f9169f21c053c562a0baa21815a8c7879be97abd930dbcf862e8/setuptools-39.1.0-py2.py3-none-any.whl
Requirement already satisfied: pyyaml in /usr/lib/python2.7/dist-packages (from keras->-r requirements.txt (line 2)) (3.12)
Requirement already satisfied: scipy>=0.14 in /usr/lib/python2.7/dist-packages (from keras->-r requirements.txt (line 2)) (0.18.1)
Requirement already satisfied: keras-applications==1.0.4 in /usr/local/lib/python2.7/dist-packages (from keras->-r requirements.txt (line 2)) (1.0.4)
Requirement already satisfied: keras-preprocessing==1.0.2 in /usr/local/lib/python2.7/dist-packages (from keras->-r requirements.txt (line 2)) (1.0.2)
Requirement already satisfied: futures>=2.2.0 in /usr/local/lib/python2.7/dist-packages (from grpcio>=1.8.6->tensorflow->-r requirements.txt (line 1)) (3.2.0)
Requirement already satisfied: markdown>=2.6.8 in /usr/lib/python2.7/dist-packages (from tensorboard<1.11.0,>=1.10.0->tensorflow->-r requirements.txt (line 1)) (2.6.8)
Requirement already satisfied: werkzeug>=0.11.10 in /usr/lib/python2.7/dist-packages (from tensorboard<1.11.0,>=1.10.0->tensorflow->-r requirements.txt (line 1)) (0.11.15)
tensorflow-serving-api 1.10.0 has requirement protobuf==3.6.0, but you'll have protobuf 3.6.1 which is incompatible.
Installing collected packages: numpy, setuptools, tensorflow
Found existing installation: numpy 1.15.1
Not uninstalling numpy at /home/john/.local/lib/python2.7/site-packages, outside environment /home/john/retrain/venv
Can't uninstall 'numpy'. No files were found to uninstall.
Found existing installation: setuptools 40.2.0
Uninstalling setuptools-40.2.0:
Successfully uninstalled setuptools-40.2.0
Successfully installed numpy-1.14.5 setuptools-39.1.0 tensorflow-1.10.1
For all of the dependencies tensorflow it shows that as existing on the system so it doesn't install them, but it then goes and installed tensorflow anyways. Why is that?
Unfortunately, there is no "fat" binary of tensorflow that is optimized for both CPU and GPU. However it is possible to use tensorflow-gpu on both instances.
It is, in fact possible to use tensoflow-gpu binaries on the instance without the GPU. In order to use it you need to install CUDA and CuDNN on the instance (even though the instance does not have an Nvidia GPU). CUDA, inside has a mock(stub) Nvidia driver that will allow CUDA and CuDNN work on top of CPU, in order to use it on linux you need to run the following commands:
sudo ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/libcuda.so.1
sudo ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/libcuda.so
assuming that /usr/local/cuda is the path where CUDA is installed (might be different on different platforms). As soon as this is done one can actually install and use tensorflow-gpu on the CPU-only instance.
I know that this looks like a hack and might not even work on some platforms but at least in some way it makes possible to use same requirenemts.txt and even the same binary on GPU and non GPU instances.