mccorby / federatedandroidtrainer Goto Github PK

Both models are sharing a (expected) number of methods and logic

A mechanism to select the example to run of those in the app (Iris, MNIST, etc...)

Each new trainer has a different set of data. This should be provided by the repository being reset with each training session

Right now the batch size is used by too many objects.
This could be better set in a trainingParams object

The serialisation of the gradients must be done in accordance to the structure of the gradients

Hi,

I'm from the deeplearning4j project. We are interested in using this and maybe even integrating this as a proper module in the framework. We would need to know the license though.

Any help would be great.

Every training process gets different training and tests datasets

Though the training might vary (by selecting a different slot every time), the test should remain the same to compare between executions

Add an abstract encryption layer/component. This should be executed by the corresponding use cases as encryption is part of the business rules

Use Docker to create a container of the server

They can be used for any model

The communication must be in the form of serialised INDArrays

Can someone please suggest how to get rid of this dependency injection error. The DaggerMainComponent is generated but still, Gradle cannot find the symbol.

Hi,

I want to run your code.
But there are no .json files. (config.json, currentRound.json)

Can you upload the files?

Inject members in the MainActivity. Use Dagger2

Each feature has to have its own configuration in the same package.
Right now they are in the app package because of the Android dependency.

Work to remove that dependency and bring the ModelConfiguration classes to their corresponding features

Federated server must be a separated service

Show a list of the models already trained with the loss after training or updating the gradients

The list should keep the previous and the current loss to see the differences after updating gradients

The model must be built by the server and passed to the clients when they start (if the model does not exist locally)

As described in https://arxiv.org/pdf/1610.02527.pdf

Abstract the way gradients are merged in the server. This should be defined by whoever is defining the model

Currently it's dividing the dataset following the Iris DataSet