Project name of "deepkit" changes nested directory behavior

[planetceres]

A logical name to keep DeepKit configuration files within a project would be nested in a subdirectory called deepkit. However, using this convention on a DeepKit server causes inconsistency with deepkit.yml config values.

Server Example

For example, say we're training a model in a parent repo of https://github.com/username/MyNetwork.git

If we create a directory ./deepkit and put deepkit.yml there, DeepKit can't seem to find the configuration. Notice the [object Object] placeholder value.

./MyNetwork
└──  deepkit
    └── deepkit.yml

But if we duplicate the same deepkit.yml file and nest it under ./deepkit/train, DeepKit is able to find the config

./MyNetwork
└──  deepkit
    └── train
        └── deepkit.yml

GUI Example

This also happens when using the GUI on localhost to manage a Project - but it is affected one level higher in the directory structure (so that the [object Object] placeholder value appears if ./deepkit.yml is in the root of the project:

./MyNetwork
└──  deepkit.yml

Summary

One solution is to make sure that DeepKit configuration files are not organized under a directory named deepkit . However, unless the directory name deepkit needs to be reserved, it might cause less confusion to correct this behavior so that experiments are not affected by this.

[marcj]

@planetceres thanks for the detailed description. Could you also please post your deepkit.yml content? I wonder if training is an actual config value defined in the file, what content it has, and if training is correctly displayed when the bug is not seen.

[planetceres]

While the example I showed above is a private project, this same behavior happens with the default examples in this repo and the deepkit python-sdk. The only difference is that I've had to replace parameters with config.

distributed-simple/deepkit.yml

image: python:3.7.0-alpine3.8

docker:
    binds:
        - /home/user/deepkit/deepkit-python-sdk/:/deepkit-sdk
    env:
        - PYTHONPATH=/deepkit-sdk

config:
    lr: 50
    optimizer: sgd
    training:
        epochs: 50
        samples: 20

tasks:
    ...

[marcj]

    training:
        epochs: 50
        samples: 20

this results always in [object Object] as nested objects are not supported in the create-experiment view. I tried your example with deepkit/deepkit.yml and it correctly displays the configuration in the create-experiment view (except the training entry config option).

I don't understand your example to be honest. You have shown two screenshots of different configurations. num-machines as of the second screenshot is not in your posted deepkit.yml, so I don't understand the reproduction case.

[marcj]

If we create a directory ./deepkit and put deepkit.yml there, DeepKit can't seem to find the configuration

But you showed a screenshot of this configuration, so Deepkit found it. I really don't get that bug. You said Deepkit doesnt find the configuration, yet you created a screenshot of this very configuration in the GUI. What exactly do you mean with "Deepkit can't seem to find it"?