phdselfhelp / mlmono Goto Github PK

Ending metric will provide one time metric in the last step (or some step).

We'd like to have a context release such that the codebase (from the cli for now) can be evaluated from any env that can run docker.

See this for more info. https://cloud.google.com/ml-engine/docs/tensorflow/technical-overview

This project will be great for people who want to learn a bit about the basics of parallel / distributed training. I suggest looking into uber's Horovod for a good solution for this.

As of now we need to import the class definition manually in cli so that all class definition is called before cli logic is run, how do we do this automatically

1. unittesting
2. linting
3. build icon in readme
4. update readme on how to use MLIO

@jdaaph as you are already working on the sqeezeseg model, can you help us make an example config based on yours?

Estimator has quite a few issues that could only be solved with enough actual integration tests.

It'd be nice to have this function:

JX do you think this is a good starting point for you to get familiar with Python OOP? What I propose is definitely not the only way to do this tho...

two steps:

1. original config is a modifiable class instance that can be updated from passed args, see Issue #4 , you can implement this override by using arg-parse to read in the later string contains all the multi-level flags.

2. you can probably make a subclass of config that basically programmatically iterate over all field recursively and set them as python property (maybe see https://www.python-course.eu/python3_properties.php)

@FXuZ @jxwulittlebean If you guys have some time can also add a devtool dir to work on this, should be easy and a cool SDE exercise.

We might want to implement the feature to support the command line arguments overriding the config settings, sometime in the future. It's not first priority.