Making run_docker.sh faster.

Describe the feature and the current behavior/state.

Using docker to make the life of contributors easier is a very good idea. But the current implementation has a drawback (that we can fix).

Currently running any command with run_docker.sh installs dependencies even if the docker image is already present locally. For example every time I call:

bash tools/run_docker.sh -c 'make unit-test'

It downloads and install tensorflow, wich takes several minutes. Same for bash tools/run_docker.sh -c 'make code-format' which installs clang format every time.

What I propose is to make a dockerfile for each of those commands and install the necessary dependencies in it. Therefore, running bash tools/run_docker.sh -c 'make code-format' would takes only a few seconds the second time and we would avoid wasting bandwidth.

Relevant information

• Are you willing to contribute it (yes/no): maybe
• Are you willing to maintain it going forward? (yes/no): maybe
• Is there already an implementation in another framework? (if so, where): In keras-tuner

Who will benefit with this feature?

Contributors

Any other info.

[gabrieldemarmiesse]

See #1022 and #1021 where there is no need to use the interactive mode to have fast formatting and sanity check.

[gabrieldemarmiesse]

See also #1030