ShapeNetV2 bug in generating splits?
berendbaas opened this issue · comments
Hi there,
Thanks for your interesting work and sharing your code.
I ran into a small issue into train/test split generation in ShapeNetV2. You might want to be aware, in case you're reusing the code in the future
When a shuffled train-test split for a shape category does not exist, the code seems to iterate over all categories listed in synsetoffset2category.txt
(they get loaded into self.numbercat2namecat
). It then creates a train/test split out of a concatenation of all these categories.
Line 788 in da4c07d
I suspect the intended behaviour was to create a train/test split for only a single category for each non-existing split-file? (As the filename contains the id of the category it's supposed to contain the split for).