eytankats / simple_converge Goto Github PK

• add support for several folds for one experiment
• remove matplotlib plots automatic upload

Add capability to add custom loss/metric/callback/model to list of existing implementations without need to change infrastructure.

Add option to set prepared in advanced training/validation/test files.

Data splitter is not initialised in 'test' method of 'main.py' script

Epoch number parameter will help control inputs and labels depending on training process stage.

Currently unified balancing strategy is applied on both training and validation data.

• Move image, plots and metrics code to utils directory.
• Organize code in utils directory in the form of single functions and not classes.
• Add documentation to methods if missing.

When test_simulation is True and a few of n folds are chosen test files will be loaded incorrectly and this can lead to crush.

Create class that responsible for management of training/testing/inference.

In 'inference' method of 'main.py' script dataframe is divided to batches but the batches are not used during model inference

Currently data split is the same for all folds.

Divide Dataset class to two different classes: one will be responsible for data loading, other for predictions postprocessing.

Edit 'unique values' log message in 'choose_leave_out_values' method

Implement custom training loop that will allow more flexibility comparing to 'compile' and 'fit' methods.

'fig_size' argument have to be preceded with argument name 'figsize'.

Add option to oversample / subsample groups to specific number of examples