Monoxide-Chen/uncertainty_retrieval

About Fashion200K dataset

Closed this issue · 3 comments

HaoliangZhou commented

Thanks for your time!
The training set of Fashion200k dataset is to get 'source_img', 'target_img', 'mod_text' for training, but the test sets of many projects are separate, for example, using def 'get_test_queries' to get 29,789 query, which contains 'source_img' and 'mod_text', as well as the 's_img_caption' and 't_img_caption', but I would like to ask how to get the 'target_img', so as to get the complete ternary for test.

Monoxide-Chen commented

Yes, the Fashion200k dataset does not directly provide the query triples. The target images and reference images are separated at this line in the code:

uncertainty_retrieval/data/fashion200k.py

Line 166 in f1fd828

def generate_random_train_queries_(self, n_modifications_per_image=3):
. The dataset processing code is inherited from CosMo, and we did not modify it.

HaoliangZhou commented

Thanks a lot! May I have your e-mail, please?