dkoutsou/dl-atlaskaggle

dl-atlaskaggle's Issues

Use predict_probas and get_probas_from_pred in baseline

change the baseline prediction and CV evaluation so that it predicts the labels from the probas as in the new neural implementation (and not direct prediction)

Runtime training DeepYeast

I ran the DeepYeast on GPU (files have been modified for GPU run in my mel branch - not send a pull request for now). So i checked that the system really used GPU but ... after 24hours not a single epochs was done... so i added save training metrics every 10 steps and val set also to see what is happedning and actually after 24hours only 10 steps were done.... the problem is the validation step (i.e. predict on the whole validation set which was set to 20% of the training set i.e. rgouhly 5000 images) took 10,000s.... and otherwise EACH single training step took 1000s with batch_size 32. these are figures on GPU run. On my computer with batch size 32 it is not even managing to do one single training step.

Random Forest baseline

In the sklearn predict function for multi class basically it checks if output proba is greater than 0,5 or not. However due to the highly imbalanced dataset it is either predicting classes 0, 25 or 0 and 25 or None. For most of the images it predicts no classes at all. We should probably handle this case differently like predicting the probabilities and then defining some custom rules to predict a least one class. Also we should probably try some "class_weights"="balanced" as rf parameter.