sed-inf-u-szeged / deepwaterframework Goto Github PK

Create a new "Manual File Input with Fixed Test Set" strategy, where one can use fixed training and fixed test csv pairs. By default only one pair needs to be set up, but the user can add multiple csv pairs as well. In this case the training won't use 10-fold cross-validation but use the provided train-test csv pairs (if 5 pairs were given, 5-fold will be performed). In this case the models should be dumped as well (pickle for scikit and dump for tensorflow). Dump models in each fold. In a sense, this feature assembling will be the generalization of the automatic 10-fold cross-validation.

At the moment, DWF evaluates the models and saves their parameters and performance measures, but not the model itself. A new feature should be added that allows the users to use a trained model for prediction of new data (from csv or whatever that is the input of the feature assembling).

Each experiment should have a priority (HIGH, NORMAL, LOW) with an extra option, called IMMEDIATE that causes all other tasks to stop and tasks in this experiment will be executed before everything else. Otherwise tasks are executed with a probability proportional to the experiment priorities in a Round Robin fashion.
Tasks should be re-ordered within experiments and run according to this order.

The worker should be named on client start (default is machine name).

When I try to run run.bat file after entering username and password I face with this message shown below.

Can you help me with that?

The learning algorithms crash if the is no sandbox folder in the client working directory. In general, we should check if the folder is there before deltion and do not fail if there is not.

The rebuild_run_and_init script asks for a Samba username but not for password on the first run. After the second run, it asks for both. Make it ask for both or none of them.

It should be evident from the worker log that which task caused the owrker to crash.

When the user selects the path of an input file, a checksum should be calculated for that input and sent along with the path, so that the client can check if the file it founds on the path is the same what the user selected.

Implement a simple Round Robin scheduler of tasks accross experiments.

A simple, general search box that searches strings anywhere in the task list.

Add ordering capabilities and a filtering possibility to the Experiment view.