Question regards to train/val/test split

[seanko29]

Hello. I have a question (maybe very newbie one T^T) regards to train/val/test split.
I read the question & answers in #539 and have extra questions on this.

The image below is my custom dataset.
So, I have a normal image folder which I ONLY want to use for training purpose.
I have a test normal/abnormal directory separately constructed.
If I put the split_ratio into null and specify abnormal_dir and normal_test_dir separately and make create_validation_set as true, will the library automatically get 50% of validation from abnormal_dir and normal_test_dir??

I do not want the model to get the split from the train data into the validation or test set.

So, in summary, train dataset for train dataset only and other abnormal / test dir for both val&test only.
Thank you in advanced :)

[djdameln]

If I put the split_ratio into null and specify abnormal_dir and normal_test_dir separately and make create_validation_set as true, will the library automatically get 50% of validation from abnormal_dir and normal_test_dir??

Yes, when create_validation_set is set to true, the evaluation images will be divided equally (50%) into a validation set and a test set which are mutually exclusive. When set to false, the test set and the validation set will both contain the same images.

I do not want the model to get the split from the train data into the validation or test set.

The settings you mentioned should achieve this. Whenever a normal_test_dir is specified, the normal images for evaluation (validation+testing) will all be taken from the normal_test_dir, and all normal images in normal_dir will be reserved exclusively for the training set. You do not even need to set split_ratio to null in this case, the parameter will just be ignored.

Please note that from v0.4.0, the naming of some of these parameters will change. create_validation_set will be deprecated in favor of validation_split_mode, and split_ratio will be renamed to normal_split_ratio. In addition, a val_split_ratio parameter will be added to change the val/test split to something other than 50%. Some documentation on this new configuration convention will follow.

[seanko29]

Thank you for the comment @djdameln !! Really appreciate it!
I see that the current version is v0.3.7. Will you upgrade it straight to v0.4.0 ?

Looking forward to the new version as well!! :D

[seanko29]

@djdameln Just a follow-up question, if you do not mind.
IF I leave the path as ' ' (empty string) like above and just directly mention normal_dir, will it automatically find the training path?

[djdameln]

Yes, that should work as well. If the path parameter is empty, you may pass the absolute path to the train and test images.