# tensor(0., device='cuda:0')
# (Pdb) pp data['mask_B'].size()
# torch.Size([1, 1, 256, 256])
# (Pdb) pp data['mask_B'].min()
# tensor(-0.5000, device='cuda:0')
# (Pdb) pp data['mask_B'].max()
# tensor(-0.5000, device='cuda:0')
model.set_input(data)
model.test(True) # True means that losses will be computed
visuals = util.get_subset_dict(model.get_current_visuals(),
to_visualize)
psnrs[i,
pp] = util.calculate_psnr_np(util.tensor2im(visuals['real']),
util.tensor2im(visuals['fake']))
save_images(webpage,
visuals,
img_path,
aspect_ratio=opt.aspect_ratio,
width=opt.display_winsize)
if i % 5 == 0:
print('processing (%04d)-th image... %s' % (i, img_path))
if i == opt.how_many - 1:
break
webpage.save()
for (pp, sample_p) in enumerate(sample_ps):
img_path = [('%08d_%.3f' % (i, sample_p)).replace('.', 'p')]
# data = util.get_colorization_data(data_raw[0], opt, ab_thresh=0., p=sample_p)
model.set_input(data)
model.test(True) # True means that losses will be computed
visuals = util.get_subset_dict(model.get_current_visuals(),
to_visualize)
vid = []
for frames in range(0, visuals['fake_reg'].shape[1], 3):
gray = util.tensor2im(visuals['gray'][:, frames:frames + 3])
real = util.tensor2im(visuals['real'][:, frames:frames + 3])
generated = util.tensor2im(
visuals['fake_reg'][:, frames:frames + 3])
vid.append(np.vstack((gray, real, generated)))
psnrs[i, pp] += util.calculate_psnr_np(
real, generated) / (visuals['fake_reg'].shape[1] / 3)
#calculate_smoothness_here
vid = np.array(vid)
arr2vid(vid, "vid_{}".format(i))
# psnrs[i, pp] = util.calculate_psnr_np(util.tensor2im(visuals['real']), util.tensor2im(visuals['fake_reg']))
entrs[i, pp] = model.get_current_losses()['G_entr']
# save_images(webpage, visuals, img_path, aspect_ratio=opt.aspect_ratio, width=opt.display_winsize)
print("")
if i % 5 == 0:
print('processing (%04d)-th image... %s' % (i, img_path))
if i == opt.how_many - 1:
break
webpage.save()
data_raw[0] = data_raw[0].cuda()
data_raw[0] = util.crop_mult(data_raw[0], mult=8)
for nn in range(N):
# embed()
data = util.get_colorization_data(data_raw,
opt,
ab_thresh=0.,
num_points=num_points[nn])
model.set_input(data)
model.test()
visuals = model.get_current_visuals()
psnrs[i, nn] = util.calculate_psnr_np(
util.tensor2im(visuals['real']),
util.tensor2im(visuals['fake_reg']))
if i == opt.how_many - 1:
break
bar.update(i)
# Save results
psnrs_mean = np.mean(psnrs, axis=0)
psnrs_std = np.std(psnrs, axis=0) / np.sqrt(opt.how_many)
np.save('./checkpoints/%s/psnrs_mean_%s' % (opt.name, str_now), psnrs_mean)
np.save('./checkpoints/%s/psnrs_std_%s' % (opt.name, str_now), psnrs_std)
np.save('./checkpoints/%s/psnrs_%s' % (opt.name, str_now), psnrs)
print(', ').join(['%.2f' % psnr for psnr in psnrs_mean])
def select_photos():
to_visualize = ['gray', 'real', 'fake_reg', 'mask', 'hint']
opt.load_model = True
opt.num_threads = 1
opt.batch_size = 1
opt.display_id = -1
opt.phase = 'val'
opt.serial_batches = True
opt.aspect_ratio = 1.
model = create_model(opt)
model.setup(opt)
# tensor_image = tensor_image.cuda()
sample_ps = 0.03125
data_raw = [None] * 4
# loader = transforms.Compose(transforms.ToTensor())
# raw_image = Image.open(args.input)
# # tensor_image = TF.to_tensor(raw_image)
# tensor_image = loader(raw_image.float())
# tensor_image = raw_image.unsqueeze(0)
# data_raw[0] = tensor_image.cuda()
# print(opt.input)
tensor_image = Image.open(opt.input).convert('RGB')
tensor_image = tensor_image.resize((opt.loadSize, opt.loadSize))
tensor_image = ToTensor()(tensor_image).unsqueeze(0)
data_raw[0] = tensor_image.cuda()
# data_raw[0] = util.crop_mult(data_raw[0], mult=8)
data = util.get_colorization_data(data_raw, opt, ab_thresh=0., p=sample_ps)
model.set_input(data)
# model.eval()
model.test(True)
# gets the visuals from the model
global visuals
visuals = util.get_subset_dict(model.get_current_visuals(), to_visualize)
# output images
raw_image = Image.fromarray(util.tensor2im(visuals['real']))
image = raw_image.resize((512, 512), Image.ANTIALIAS)
image = ImageTk.PhotoImage(image)
label = tk.Label(text="Original Image", compound='top', image=image)
label.photo = image # assign to class variable to resolve problem with bug in `PhotoImage`
label.grid(row=1, column=1)
all_labels.append(label)
raw_image = Image.fromarray(util.tensor2im(visuals['hint']))
# lab_raw_image = util.rgb2lab(raw_image, opt)
image = raw_image.resize((512, 512), Image.ANTIALIAS)
image = ImageTk.PhotoImage(image)
label = tk.Label(image=image, text="Hint Image", compound='top')
label.bind("<Button-1>", lambda e: choose_colour())
label.photo = image
label.grid(row=1, column=3)
all_labels.append(label)
raw_image = Image.fromarray(util.tensor2im(visuals['fake_reg']))
image = raw_image.resize((512, 512), Image.ANTIALIAS)
image = ImageTk.PhotoImage(image)
label = tk.Label(image=image, text="Colourised Image", compound='top')
label.bind("<Button-1>", lambda e: choose_colour())
label.photo = image
label.grid(row=1, column=5)
all_labels.append(label)
original = Image.open(opt.input)
original = original.resize((opt.loadSize, opt.loadSize))
original = np.asarray(original)
label = tk.Label(text="PSNR Numpy: " + str(
util.calculate_psnr_np(original, util.tensor2im(visuals['fake_reg']))))
label.grid(row=4, column=1)
all_labels.append(label)
label = tk.Label(
text="MSE : " +
str(util.calculate_mse(original, util.tensor2im(visuals['fake_reg']))))
label.grid(row=4, column=3)
all_labels.append(label)
label = tk.Label(
text="MSE : " +
str(util.calculate_mae(original, util.tensor2im(visuals['fake_reg']))))
label.grid(row=4, column=5)
all_labels.append(label)
print("data_raw[0].shape")
print(data_raw[0].shape)
for nn in range(N):
# embed()
data = util.get_colorization_data(data_raw,
opt,
ab_thresh=0.,
num_points=num_points[nn])
model.set_input(data)
model.test()
visuals = model.get_current_visuals()
real = util.tensor2im(visuals['real'])
generated = util.tensor2im(visuals['fake_reg'])
psnrs[i, nn] = util.calculate_psnr_np(real, generated)
plt.subplot(121)
plt.imshow(real)
plt.subplot(122)
plt.imshow(generated)
plt.title("PSNR: " + str(psnrs[i, nn]))
plt.savefig("real_vs_generated" + str(i) + ".png")
if i == opt.how_many - 1:
break
bar.update(i)
print(psnrs)
# Save results
psnrs_mean = np.mean(psnrs, axis=0)
data_raw[0] = util.crop_mult(data_raw[0], mult=8)
mask = util.build_mask(data_raw[0].shape, R=R, G=G).cuda()
data = util.reveal_rgb_values(data_raw, mask, opt)
model.set_input(data)
model.test()
visuals = model.get_current_visuals()
if not (R,G) in psnrs:
psnrs[(R,G)] = np.zeros(opt.how_many)
if not (R,G) in mses:
mses[(R,G)] = np.zeros(opt.how_many)
mse_value = util.calculate_mse_np(util.tensor2im(visuals['real_rgb']), util.tensor2im(visuals['fake_rgb']))
psnr_value = util.calculate_psnr_np(util.tensor2im(visuals['real_rgb']), util.tensor2im(visuals['fake_rgb']))
mses[(R,G)][i] = mse_value
psnrs[(R,G)][i] = psnr_value
if i == opt.how_many - 1:
break
k = k + 1
bar.update(k)
# Save results
psnrs_mean = np.zeros((values_num, values_num))
mses_mean = np.zeros((values_num, values_num))
for i in range(psnrs_mean.shape[0]):
for j in range(psnrs_mean.shape[1]):
r,g = i*jump_number, j*jump_number
