Transfer Learning of your model

[karthik-sem]

@xtudbxk .The model is for 21 categories.Can u please suggest methods to convert it to a 2 category model?

[xtudbxk]

Sorry for the late reply. It seems there exist two things you have to do for changing it to a two categories model. The first thing is that you should change the 'category_num' in the code from 21 to 2. And the second is that you should generate the two categories' seeds by yourself. You can use CAM or DRFI to generate the seeds.

[karthik-sem]

Thank you @xtudbxk .It is very helpful.Can u please tell what i should include in SegmentationClasses Folder for my dataset?I cannot add the mask images directly as in a sense it is fully supervised data. The Augmented images should be combination of what ?

[xtudbxk]

You should put the input images in JPEGImages folder and put the pixel-level tags in the SegmentationClassAug folder for fully supervised data. And you can refer to my code for details of paths for dataset.

[karthik-sem]

But instead of pixel level tags can we include CAM images which highlights the foreground?I want it to be completely weakly supervised without using pixel level tags.

[xtudbxk]

Of course~ You can store the generated cam dataset with the same format of "localization_cues-cal.pickle", and put it on the data folder. For the details of the format, see "pythonlib/dataset_DSRG.py".

[karthik-sem]

Thank you @xtudbxk