Run tf_mnist.py in local conda environment.

$ pip install tensorflow
$ az ml experiment submit -c local tf_mnist.py

Run tf_mnist.py in a local Docker container.

$ az ml experiment submit -c docker tf_mnist.py

Run tf_mnist.py in a Docker container in a remote machine. Note you need to create/configure myvm.compute.

$ az ml experiment submit -c myvm tf_mnist.py

Run tf_mnist.py in a Docker container in a remote machine with GPU.

• Create a new compute context, nam it gpu (or any arbitary name)
• Use az ml computetarget attach to target the GPU equipped VM.
• In conda_dependencies.yml file, use tensorflow-gpu instead of tensorflow.
• In gpu.compute file, use microsoft/mmlspark:gpu as the base Docker image.
• In gpu.compute file, add a line nvidiaDocker: true
• In gpu.runconfig file, set Framework to Python
• Now run the script.

$ az ml experiment submit -c gpu tf_mnist.py

For more information on using GPU in Vienna execution, read this article.