print('%d/100' % i)
# get batch of test samples
images, _, segs, txt_data, txt_len, captions, *_ = next(iter(dataloader))
images = images.cuda()
segs = segs.cuda()
txt_data = txt_data.cuda()
bimages = images
netG.eval()
netEs.eval()
netEb.eval()
# alignment
if align == 'shape':
bimages = roll(images, 2, dim=0) # for text and seg mismatched backgrounds
bsegs = roll(segs, 2, dim=0) # for text and seg mismatched backgrounds
elif align == 'background':
segs = roll(segs, 1, dim=0) # for text mismatched segmentations
elif align == 'all':
bimages = images.clone()
bsegs = segs.clone()
elif align == 'none':
bimages = roll(images, 2, dim=0) # for text and seg mismatched backgrounds
segs = roll(segs, 1, dim=0) # for text mismatched segmentations
bsegs = roll(segs, 2, dim=0) # for text and seg mismatched backgrounds
np_segs = to_numpy(segs)
np_bsegs = to_numpy(bsegs)
np_images = to_numpy(images)
np_bimages = to_numpy(bimages)
images = images.cuda()
segs = segs.cuda()
txt_data = txt_data.cuda()
bimages = images
netG.eval()
netEs.eval()
netEb.eval()
vis_samples = [None for i in range(8)]
for i in range(6):
# alignment
# bimages = roll(images, i+2, dim=0) # for text and seg mismatched backgrounds
# segs = roll(segs, i+1, dim=0) # for text mismatched segmentations
txt_data = roll(txt_data, i + 1)
txt_len = roll(txt_len, i + 1)
# np to save
np_segs = np.repeat(to_numpy(segs), 3, 1) * 2 - 1
np_images = to_numpy(images)
np_bimages = to_numpy(bimages)
# generate testing results
vis_samples[0] = np_images
vis_samples[1] = np_bimages
segs_code = netEs(segs)
bkgs_code = netEb(bimages)
*_, f_images, z_list = netG(txt_data, txt_len, segs_code, bkgs_code)
images, _, segs, txt_data, txt_len, captions, *_ = next(iter(dataloader))
images = images.cuda()
segs = segs.cuda()
txt_data = txt_data.cuda()
bimages = images
netG.eval()
netEs.eval()
netEb.eval()
vis_samples = [None for i in range(4 + 5 * 2)]
for i in range(6):
# alignment
# bimages = roll(images, i+2, dim=0) # for text and seg mismatched backgrounds
segs = roll(segs, i + 1, dim=0) # for text mismatched segmentations
if i == 5:
segs = torch.zeros(segs.shape).cuda()
# np to save
np_segs = np.repeat(to_numpy(segs), 3, 1) * 2 - 1
np_images = to_numpy(images)
np_bimages = to_numpy(bimages)
# generate testing results
vis_samples[0] = np_images
vis_samples[1] = np_bimages
vis_samples[2 + i * 2] = np_segs
segs_code = netEs(segs)
bkgs_code = netEb(bimages)