Jason (Joo Seong) Jeong
Jason (Joo Seong) Jeong
Some of our sample files don't contain any information on how/where they were generated. Each file should state whether it was taken from an external source or not and how...
The methods `Logger.entering()` and `Logger.exiting()` are in some classes, mostly for debugging purposes. REEF also has these methods in various classes, mainly the group communication classes. However, I do feel...
Besides the online Optimizer, we'd like to build a tool that can find the optimal system configuration from collected metrics offline, for both testing and demonstration purposes.
The methods `sliceColumn`, `sliceRow`, `sliceRows`, and `sliceColumns` are currently not implemented for sparse matrices (`CSCMatrix`), but after more investigation it seems that they are in fact implementable (in a rather...
For lasso regression (and potentially other algorithms), we normally perform feature scaling on the input data before feeding it to the algorithm. However, feature scaling is not trivial in distributed...
Although we are planning to execute optimization periodically, we want to be able to implement other trigger conditions such as waiting for a change in the amount of available resources,...
In `dolphin-dnn` of the [open-source version dolphin](https://github.com/cmssnu/dolphin), accuracy stats for the training dataset and validation dataset are being gathered at the parameter server alongside the training model. This not only...
The current optimizer is strongly coupled to Dolphin: `OptimizationOrchestrator`, `MetricsReceiver`, `CostCalculator`, and `DefaultPlanExecutor`. However the optimizer should be as general as possible since our ultimate goal is to support all...
Similar to https://github.com/cmssnu/dolphin/issues/141, the dolphin framework should be able to process data in batches and not just single units. The best scenario would be that we can choose between batch...
The optimizer should be cautious when adding new evaluators or deleting existing ones. This can be done by adding a penalty term to the optimization formula that rises as evaluators...