4

I'm trying to install Keras in anaconda from the instructions given here.

I ran the command conda install -c conda-forge keras and then after that, I pressed y when the prompt appeared asking me to press y/n. Then some libraries were downloaded and immediately after that the anaconda prompt crashed.

I looked at this thread and I couldn't try installing it using pip because every time I activate the environment, it crashes in a few seconds after automatically running the following commands:

C:\Users\ashut>python C:\Users\ashut\Anaconda3\envs\py36\etc\keras\load_config.py  1>temp.txt

C:\Users\ashut>set /p KERAS_BACKEND= 0<temp.txt

C:\Users\ashut>del temp.txt

C:\Users\ashut>python -c "import keras"  1>nul 2>&1

I could barely manage to get a screenshot of the error that is generated after the commands given above are executed. I don't know what's happening here:

enter image description here

And then it crashes.

I expected at least the instructions on anaconda's own site to be reliable, but sadly this doesn't seems to be the case. If there's any other method to install Keras in anaconda without losing this environment, I'd be more than grateful to know.

Ashutosh Pathak
  • 1,619
  • 1
  • 8
  • 12

2 Answers2

1

You are missing build tools. Anaconda depends on their presence and it cannot supply them by itself, as these are proprietary Microsoft tools. To fix the problem install them from here. After you reboot your computer, Anaconda should work just fine.

pip works for you since it takes a binary, it does not compile it. It's suboptimal and in case of deep learning performance is important. Check e.g. this article to learn about differences between conda and pypi tensorflow.

Lukasz Tracewski
  • 10,794
  • 3
  • 34
  • 53
0

Nothing worked actually. I tried deleting the environment in which I installed Keras, but even deleting it and creating an environment with another name, the same four statements were automatically executed and crashed the anaconda prompt, as before.

So I reinstalled Anaconda, again made an environment for python 3.6, and installed tensorflow-gpu and keras using pip. I don't know why but it worked this time.

Edit 1: I found this after I'd performed the steps described above. This thread also concentrates on reinstalling Anaconda, as no solution to such an event is known yet.

Edit 2: Okay I think I figured out what the problem was. I had the wrong version of CUDA installed - CUDA 10, which is not yet supported by tensorflow and hence keras showed problems. So this time, I installed CUDA 9 and it's corresponding CuDNN version, uninstalled the pip versions of tensorflow & keras and installed them again using conda. It worked, without a hitch.

TL;DR:

The following works on CUDA 9, CuDNN 7.4.2

pip uninstall tensorflow-gpu
pip uninstall keras

conda install tensorflow-gpu
conda install keras-gpu

Read Lucasz's answer for why I uninstalled the pip versions.

Ashutosh Pathak
  • 1,619
  • 1
  • 8
  • 12
  • I reinstalled Anaconda, as I said above, and installed them using `pip` afterwards. I found your answer quite convincing but I'll have to try it myself before saying anything. Will update when I do so. – Ashutosh Pathak Jan 16 '19 at 15:42
  • 1
    Sure thing. This isn't just an alternative way; tensorflow from Anaconda or compiled (with specialised math libraries) performs much better on average. – Lukasz Tracewski Jan 16 '19 at 16:00