A question about how to test retrieval
gimpong opened this issue · comments
Hi, thank you for building such a useful and concise framework for image retrival task!
I just read the file inference.py
and got confused about the testing process.
PyTorch-Image-Retrieval/inference.py
Lines 20 to 23 in 8116251
My question is, why is the database for retrieval getting divided into batches? Shouldn't we keep it as a whole for every query, since we want to rank all the reference images in whole database according to the pair-wise similarity?
In the batch_process
function, all feature vectors are concatenated.
PyTorch-Image-Retrieval/inference.py
Lines 28 to 29 in 8116251
PyTorch-Image-Retrieval/inference.py
Lines 111 to 122 in 8116251