def forwordgradients(img_list, batch_posimgids, batch_possamples, GBP, vector):
uniqueimgid = np.unique(batch_posimgids)
theuniI = 0
for theuniqid in uniqueimgid:
theuniexps = []
for ind in range(batch_possamples.shape[0]):
if batch_posimgids[ind] == theuniqid:
theuniexps.append(batch_possamples[ind])
theuniexps = torch.from_numpy(np.stack(theuniexps, 0)).view(-1,
4).numpy()
theimg = Image.open(img_list[theuniqid]).convert('RGB')
extractor = RegionExtractor(theimg, theuniexps, opts)
for i, regions in enumerate(extractor):
regions = Variable(regions)
if opts['use_gpu']:
regions = regions.cuda()
if i == 0:
feats = regions
else:
feats = torch.cat((feats, regions), 0)
if theuniI == 0:
posregions = feats
else:
posregions = torch.cat((posregions, feats), 0)
theuniI = theuniI + 1
guided_grads = GBP.generate_gradients(posregions, vector)
return guided_grads
def forward_samples(model, image, samples, out_layer='conv3'):
model.eval()
extractor = RegionExtractor(image, samples, opts)
for i, regions in enumerate(extractor):
if opts['use_gpu']:
regions = regions.cuda()
with torch.no_grad():
feat = model(regions, out_layer=out_layer)
if i == 0:
feats = feat.detach().clone()
else:
feats = torch.cat((feats, feat.detach().clone()), 0)
return feats
def forward_samples(model, image, samples, out_layer='conv3'):
assert len(samples) > 0
model.eval()
extractor = RegionExtractor(image, samples, opts['img_size'],
opts['padding'], opts['batch_test'])
for i, regions in enumerate(extractor):
regions = Variable(regions)
if opts['use_gpu']:
regions = regions.cuda()
feat = model(regions, out_layer=out_layer)
if i == 0:
feats = feat.data.clone()
else:
feats = torch.cat((feats, feat.data.clone()), 0)
return feats