make it possible to train with Pascal VOC dataset images,
using the .csv representing the
Visual Search Difficulty dataset,
by converting that .csv to the same format used by
'searchstims'.
(See paper repo for script that does this.)
Key change is to add 'CE-VSD' loss,
that causes training with 'searchstims'
to use a different set of transforms
add 'CE-VSD' loss to transforms.get_transforms
this returns the padding transform for the image,
plus ImageNet normalization,
instead of the searchstims image transform.
then add 'CE-VSD' loss everywhere else to avoid crashes.
probably a more elegant / general way to handle this,
but just doing this for now
add 'CE-VSD' to VALID_LOSS_FUNCTIONS in searchnets/config/train.py
make it possible to train with Pascal VOC dataset images, using the .csv representing the Visual Search Difficulty dataset, by converting that .csv to the same format used by 'searchstims'. (See paper repo for script that does this.)
Key change is to add 'CE-VSD' loss, that causes training with 'searchstims' to use a different set of transforms
add 'CE-VSD' loss to transforms.get_transforms
then add 'CE-VSD' loss everywhere else to avoid crashes.
probably a more elegant / general way to handle this, but just doing this for now
add 'CE-VSD' to VALID_LOSS_FUNCTIONS in searchnets/config/train.py
add 'CE-VSD' to losses in searchnets/train.py
add 'CE-VSD' in several places in AbstractTrainer
add 'CE-VSD' loss in engine/tester.py