GPU docker image does not find keras

[marketneutral]

I have used the CPU docker image extensively and am trying to migrate to the GPU one. I am having an issue with just the basics.

• I am running Ubuntu 20.04
• Docker version 20.10.7

I try the following sequence

• docker pull gcr.io/kaggle-gpu-images:v102
• docker run -it --gpus all gcr.io/kaggle-gpu-images/python:v102 /bin/bash
• nvidia-smi is fine... all looks good
• ipython
• import keras

this throws

---------------------------------------------------------------------------
ModuleNotFoundError                       Traceback (most recent call last)
<ipython-input-1-88d96843a926> in <module>
----> 1 import keras

ModuleNotFoundError: No module named 'keras'

on the Kaggle site, this works fine with GPU enabled

I should note that I can import other libraries locally (e.g., import torch works fine).

Do you have an idea of what is happening to cause this issue?

Thank you.

[rosbo]

Hi @marketneutral,

Thanks for the detailed repro steps.

We've removed multi-backend Keras from our base image since it has been discontinued. See #1045.

We will upgrade to TensorFlow 2.6 soon and I was hitting issue with multi-backend Keras + TF 2.6 and decided to remove it from the base image to keep it lean given that it was discontinued.

You should use tf.keras instead.

Thank you