Currently, the in-browser training does a single, cancellable run and saves the final model but no other data about the run. This also leads to some visual bugs, like changing the number of epochs on a second run will make the loading bar change size, or changing the number of cell types messing with the confusion matrix. Rerunning the training will overwrite the first model, and there is no way of exporting the model. The following list of features should eventually be implemented:
Save all hyperparameter data, the loss logs and confusion matrix, and final model after each run without being overwritten by future training runs
Allow one to select previous runs, use those models, and view the visualization data and such
Pause and resume training, where pausing saves the model
Continue training a model that was already saved
Export and import models from previous runs
More advanced and customizable features, such as asking to save model on every X epochs, early stopping, auto stopping, etc.
Currently, the in-browser training does a single, cancellable run and saves the final model but no other data about the run. This also leads to some visual bugs, like changing the number of epochs on a second run will make the loading bar change size, or changing the number of cell types messing with the confusion matrix. Rerunning the training will overwrite the first model, and there is no way of exporting the model. The following list of features should eventually be implemented: