Suboptimal dockerfile structure in gpulibs

[nerok]

I was looking into which parts of this image that mainly used space, as I felt the 17.9GB of uncompressed image (looked at cschranz/gpu-jupyter:v1.4_cuda-11.0_ubuntu-18.04). According to Docker desktop on my machine, these lines contributed around 6GB of space: https://github.com/iot-salzburg/gpu-jupyter/blob/d704653e3d8f816b583e1bb55fdcc2f2f1f9bdfb/src/Dockerfile.gpulibs#L11-L22

This differs a lot from how for instance the Jupyter Project uses Conda in their dockerfiles, such as in the base-notebook: https://github.com/jupyter/docker-stacks/blob/6d61708f1747d4fb15e3c0166805ebb5fba41ea1/base-notebook/Dockerfile#L134-L144

A better approach for the lines in gpulibs would be something like:

RUN conda install --quiet --yes \
    pyyaml mkl mkl-include setuptools cmake cffi typing && \
    pip install torch torchvision torchaudio torchviz && \
    conda clean --all -f -y && \
    fix-permissions $CONDA_DIR && \
    fix-permissions /home/$NB_USER

I tried using this diff on my machine and got a image size of 13.8GB (uncompressed) on the "v1.4_cuda-11.0_ubuntu-18.04" -branch

diff --git a/.build/Dockerfile b/.build/Dockerfile
index f20b066..596c051 100755
--- a/.build/Dockerfile
+++ b/.build/Dockerfile
@@ -307,15 +307,9 @@ RUN pip install --upgrade pip && \

 # Install PyTorch with dependencies
 RUN conda install --quiet --yes \
-    pyyaml mkl mkl-include setuptools cmake cffi typing
-
-# Check compatibility here:
-# https://pytorch.org/get-started/locally/
-# Installation via conda leads to errors installing cudatoolkit=10.1
-RUN pip install torch torchvision torchaudio torchviz
-
-# Clean installation
-RUN conda clean --all -f -y && \
+    pyyaml mkl mkl-include setuptools cmake cffi typing && \
+    pip install torch torchvision torchaudio torchviz && \
+    conda clean --all -f -y && \
     fix-permissions $CONDA_DIR && \
     fix-permissions /home/$NB_USER

Some similar fixes might be found for pip aswell.

If there is a reason why you have split this into multiple stages, there might be smart to add a comment about it.

[mathematicalmichael]

nice work! This seems a reasonable approach, feel free to open a PR. Doesn't look like it changes any functionality, just makes the image smaller, so I see nothing wrong with it.

[nerok]

Added a PR now, not sure who should be reviewing it.