def score_jnd_dataset(data_loader, func):
''' Function computes JND score using distance function 'func' in dataset 'data_loader'
INPUTS
data_loader - CustomDatasetDataLoader object - contains a JNDDataset inside
func - callable distance function - calling d=func(in0,in1) should take 2
pytorch tensors with shape Nx3xXxY, and return numpy array of length N
OUTPUTS
[0] - JND score in [0,1], mAP score (area under precision-recall curve)
[1] - dictionary with following elements
ds - N array containing distances between two patches shown to human evaluator
sames - N array containing fraction of people who thought the two patches were identical
CONSTS
N - number of test triplets in data_loader
'''
ds = []
gts = []
# bar = pb.ProgressBar(max_value=data_loader.load_data().__len__())
for (i, data) in enumerate(data_loader.load_data()):
ds += func(data['p0'], data['p1']).tolist()
gts += data['same'].cpu().numpy().flatten().tolist()
# bar.update(i)
sames = np.array(gts)
ds = np.array(ds)
sorted_inds = np.argsort(ds)
ds_sorted = ds[sorted_inds]
sames_sorted = sames[sorted_inds]
TPs = np.cumsum(sames_sorted)
FPs = np.cumsum(1 - sames_sorted)
FNs = np.sum(sames_sorted) - TPs
precs = TPs / (TPs + FPs)
recs = TPs / (TPs + FNs)
score = util.voc_ap(recs, precs)
return (score, dict(ds=ds, sames=sames))
def score_jnd_dataset(data_loader,func):
''' Function computes JND score using distance function 'func' in dataset 'data_loader'
INPUTS
data_loader - CustomDatasetDataLoader object - contains a JNDDataset inside
func - callable distance function - calling d=func(in0,in1) should take 2
pytorch tensors with shape Nx3xXxY, and return numpy array of length N
OUTPUTS
[0] - JND score in [0,1], mAP score (area under precision-recall curve)
[1] - dictionary with following elements
ds - N array containing distances between two patches shown to human evaluator
sames - N array containing fraction of people who thought the two patches were identical
CONSTS
N - number of test triplets in data_loader
'''
ds = []
gts = []
# bar = pb.ProgressBar(max_value=data_loader.load_data().__len__())
for (i,data) in enumerate(data_loader.load_data()):
ds+=func(data['p0'],data['p1']).tolist()
gts+=data['same'].cpu().numpy().flatten().tolist()
# bar.update(i)
sames = np.array(gts)
ds = np.array(ds)
sorted_inds = np.argsort(ds)
ds_sorted = ds[sorted_inds]
sames_sorted = sames[sorted_inds]
TPs = np.cumsum(sames_sorted)
FPs = np.cumsum(1-sames_sorted)
FNs = np.sum(sames_sorted)-TPs
precs = TPs/(TPs+FPs)
recs = TPs/(TPs+FNs)
score = util.voc_ap(recs,precs)
return(score, dict(ds=ds,sames=sames))