def infer(self, source_obj, model_dir, save_dir):
source_provider = InjectDataProvider(source_obj)
source_iter = source_provider.get_iter(self.batch_size)
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=self.retrieve_generator_vars())
self.restore_model(saver, model_dir)
def save_imgs(imgs, count):
p = os.path.join(save_dir, "inferred_%04d.png" % count)
save_concat_images(imgs, img_path=p)
print("generated images saved at %s" % p)
count = 0
batch_buffer = list()
for source_imgs in source_iter:
fake_imgs = self.generate_fake_samples(source_imgs)[0]
merged_fake_images = merge(scale_back(fake_imgs),
[self.batch_size, 1])
batch_buffer.append(merged_fake_images)
if len(batch_buffer) == 10:
save_imgs(batch_buffer, count)
batch_buffer = list()
count += 1
if batch_buffer:
# last batch
save_imgs(batch_buffer, count)
def validate_model(self, images, epoch, step):
fake_imgs, real_imgs, d_loss, g_loss, l1_loss = self.generate_fake_samples(images)
print("Sample: d_loss: %.5f, g_loss: %.5f, l1_loss: %.5f" % (d_loss, g_loss, l1_loss))
merged_fake_images = merge(scale_back(fake_imgs), [self.batch_size, 1])
merged_real_images = merge(scale_back(real_imgs), [self.batch_size, 1])
merged_pair = np.concatenate([merged_real_images, merged_fake_images], axis=1)
model_id, _ = self.get_model_id_and_dir()
model_sample_dir = os.path.join(self.sample_dir, model_id)
if not os.path.exists(model_sample_dir):
os.makedirs(model_sample_dir)
sample_img_path = os.path.join(model_sample_dir, "sample_%02d_%04d.png" % (epoch, step))
misc.imsave(sample_img_path, merged_pair)
def test(self, source_provider, model_dir, save_dir):
source_len = len(source_provider.data.examples)
total_count = source_len
source_len = min(10, source_len)
source_iter = source_provider.get_iter(source_len)
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=self.retrieve_generator_vars())
self.restore_model(saver, model_dir)
def save_imgs(imgs, count, threshold):
p = os.path.join(save_dir,
"inferred_%04d_%.2f.png" % (count, threshold))
save_concat_images(imgs, img_path=p)
print("generated images saved at %s" % p)
count = 0
threshold = 0.1
batch_buffer = list()
accuracy = 0.0
for source_imgs in source_iter:
fake_imgs, real_imgs, d_loss, g_loss, l1_loss = self.generate_fake_samples(
source_imgs)
img_shape = fake_imgs.shape
fake_imgs_reshape = np.reshape(
np.array(fake_imgs),
[img_shape[0], img_shape[1] * img_shape[2] * img_shape[3]])
real_imgs_reshape = np.reshape(
np.array(real_imgs),
[img_shape[0], img_shape[1] * img_shape[2] * img_shape[3]])
# threshold
for bt in range(fake_imgs_reshape.shape[0]):
for it in range(fake_imgs_reshape.shape[1]):
if fake_imgs_reshape[bt][it] >= threshold:
fake_imgs_reshape[bt][it] = 1.0
else:
fake_imgs_reshape[bt][it] = -1.0
for bt in range(fake_imgs_reshape.shape[0]):
over = 0.0
less = 0.0
base = 0.0
for it in range(fake_imgs_reshape.shape[1]):
if real_imgs_reshape[bt][
it] == 1.0 and fake_imgs_reshape[bt][it] != 1.0:
over += 1
if real_imgs_reshape[bt][it] != 1.0 and fake_imgs_reshape[
bt][it] == -1.0:
less += 1
if real_imgs_reshape[bt][it] != 1.0:
base += 1
print("over:{} - under:{} - base:{}".format(over, less, base))
accuracy += 1 - ((over + less) / base)
print("avg acc:{}".format(1 - ((over + less) / base)))
fake_imgs_reshape = np.reshape(fake_imgs_reshape, fake_imgs.shape)
real_imgs_reshape = np.reshape(real_imgs_reshape, real_imgs.shape)
merged_fake_images = merge(scale_back(fake_imgs_reshape),
[source_len, 1])
merged_real_images = merge(scale_back(real_imgs_reshape),
[source_len, 1])
merged_pair = np.concatenate(
[merged_real_images, merged_fake_images], axis=1)
batch_buffer.append(merged_pair)
count += 1
if batch_buffer:
# last batch
save_imgs(batch_buffer, count, threshold)
accuracy = accuracy / total_count
print("Average accruacy: %.5f" % accuracy)
def test(self, source_provider, model_dir, save_dir):
source_len = len(source_provider.data.examples)
source_len = min(16, source_len)
source_iter = source_provider.get_iter(source_len)
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=self.retrieve_generator_vars())
self.restore_model(saver, model_dir)
def save_imgs(imgs, count, threshold):
p = os.path.join(save_dir, "inferred_%04d_%.2f.png" % (count, threshold))
save_concat_images(imgs, img_path=p)
print("generated images saved at %s" % p)
def save_img(img, mse_diff, nrmse_diff, ssim_diff, psnr_diff):
p = os.path.join(save_dir, "cgan_patch%.4f-%.4f-%.4f-%.4f.png" % (ssim_diff, mse_diff, nrmse_diff,
psnr_diff))
save_image(img, img_path=p)
# print("generated ssim: %.4f images saved at %s" % (ssim_diff, p) )
def save_batch_samples(imgs, count, threshold):
p = os.path.join(save_dir, "cgan_test_sample id:%04d_count:%04d_%.2f.png" % (self.experiment_id, count,
threshold))
try:
save_concat_images(imgs, img_path=p)
# print("test batch samples saved!")
except Exception as e:
print(e)
def save_single_img(img, count, bt):
p = os.path.join(save_dir, "cgan_single_%d_%d.png" % (count, bt))
save_image(img, img_path=p)
print("cgan single sample id: %d _ %d saved" % (count, bt))
count = 0
threshold = 0.1
# batch_buffer = list()
for source_imgs in source_iter:
fake_imgs, real_imgs, d_loss, g_loss, l1_loss = self.generate_fake_samples(source_imgs)
img_shape = fake_imgs.shape
fake_imgs_reshape = np.reshape(np.array(fake_imgs),
[img_shape[0], img_shape[1] * img_shape[2] * img_shape[3]])
real_imgs_reshape = np.reshape(np.array(real_imgs),
[img_shape[0], img_shape[1] * img_shape[2] * img_shape[3]])
fake_imgs_reshape_saved = fake_imgs_reshape
real_imgs_reshape_saved = real_imgs_reshape
# threshold -- fixed
for bt in range(fake_imgs_reshape.shape[0]):
# statistics mean of generator output
g_mean = np.mean(np.array(fake_imgs_reshape))
g_min = np.min(np.array(fake_imgs_reshape))
g_max = np.max(np.array(fake_imgs_reshape))
print("g_mean : %.05f g_min: %.05f g_max:%.05f" % (g_mean, g_min, g_max))
for it in range(fake_imgs_reshape.shape[1]):
if fake_imgs_reshape[bt][it] >= threshold:
fake_imgs_reshape[bt][it] = 1.0
else:
fake_imgs_reshape[bt][it] = -1.0
# otsu threshold
# radius = 15
# selem = disk(radius)
#
# local_otsu = rank.otsu(fake_imgs_reshape, selem)
# fake_imgs_reshape >= local_otsu
# valid pixels
for bt in range(fake_imgs_reshape.shape[0]):
p_over = 0
p_less = 0
p_valid = 0
for it in range(fake_imgs_reshape.shape[1]):
if fake_imgs_reshape[bt][it] == 1.0 and real_imgs_reshape[bt][it] != 1.0:
p_over += 1
if fake_imgs_reshape[bt][it] != 1.0 and real_imgs_reshape[bt][it] == 1.0:
p_less += 1
if real_imgs_reshape[bt][it] == 1.0:
p_valid += 1
p_accuracy = 1.0 * (p_valid - p_over - p_less) / p_valid
print("cgan count %d sample %d pixel accuracy: %.05f" % (count, bt, p_accuracy))
# save ave sample images
for bt in range(fake_imgs_reshape.shape[0]):
fk_reshape = np.reshape(fake_imgs_reshape_saved[bt], (fake_imgs.shape[1], fake_imgs.shape[2]))
save_single_img(fk_reshape, count, bt)
# mse, nrmse, ssim and psnr
for bt in range(fake_imgs_reshape.shape[0]):
mse_diff = compare_mse(real_imgs_reshape[bt], fake_imgs_reshape[bt])
nrmse_diff = compare_nrmse(real_imgs_reshape[bt], fake_imgs_reshape[bt], norm_type="Euclidean")
ssim_diff = compare_ssim(real_imgs_reshape[bt], fake_imgs_reshape[bt])
psnr_diff = compare_psnr(real_imgs_reshape[bt], fake_imgs_reshape[bt])
print("mse diff:{} | nrmse diff:{} | ssim:{} | psnr:{}".format(mse_diff, nrmse_diff,
ssim_diff, psnr_diff))
# kde
r_reshape = np.reshape(real_imgs_reshape[bt], [1, img_shape[1] * img_shape[2] * img_shape[3]])
f_reshape = np.reshape(fake_imgs_reshape[bt], [1, img_shape[1] * img_shape[2] * img_shape[3]])
# kde = KernelDensity(kernel="gaussian", bandwidth=1).fit(r_reshape)
# kde_score_fake = kde.score_samples(f_reshape)
# kde_score_real = kde.score_samples(r_reshape)
# print("fake: %0.3f real: %0.3f" % (kde_score_fake, kde_score_real))
# save the images with ssim > 0.8 and ssim < 0.5
if ssim_diff > 0.8 or ssim_diff < 0.5:
fk_reshape = np.reshape(fake_imgs_reshape_saved[bt], (1, fake_imgs.shape[1], fake_imgs.shape[2],
fake_imgs.shape[3]))
rl_reshape = np.reshape(real_imgs_reshape_saved[bt], (1, real_imgs.shape[1], real_imgs.shape[2],
real_imgs.shape[3]))
fk_reshape = merge(scale_back(fk_reshape), [1, 1])
rl_reshape = merge(scale_back(rl_reshape), [1, 1])
pair = np.concatenate([rl_reshape, fk_reshape], axis=1)
save_img(pair, mse_diff, nrmse_diff, ssim_diff, psnr_diff)
fake_imgs_reshape = np.reshape(fake_imgs_reshape, fake_imgs.shape)
real_imgs_reshape = np.reshape(real_imgs_reshape, real_imgs.shape)
merged_fake_images = merge(scale_back(fake_imgs_reshape), [source_len, 1])
merged_real_images = merge(scale_back(real_imgs_reshape), [source_len, 1])
merged_pair = np.concatenate([merged_real_images, merged_fake_images], axis=1)
merged_pair_splited = np.split(merged_pair, 4)
save_batch_samples(merged_pair_splited, count, threshold)
# batch_buffer.append(merged_pair)
count += 1
# if batch_buffer:
# # last batch
# save_imgs(batch_buffer, count, threshold)
def test(self, source_provider, model_dir, save_dir):
source_len = len(source_provider.data.examples)
source_len = min(16, source_len)
source_iter = source_provider.get_iter(source_len)
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=self.retrieve_generator_vars())
self.restore_model(saver, model_dir)
def save_imgs(imgs, count, threshold):
p = os.path.join(
save_dir, "wgan_inferred_%04d_%.2f.png" % (count, threshold))
save_concat_images(imgs, img_path=p)
print("wgan generated images saved at %s" % p)
def save_img(img, mse_diff, nrmse_diff, ssim_diff, psnr_diff):
p = os.path.join(
save_dir, "wgan-%.4f-%.4f-%.4f-%.4f.png" %
(ssim_diff, mse_diff, nrmse_diff, psnr_diff))
save_image(img, img_path=p)
print("wgan generated ssim: %.4f images saved at %s" %
(ssim_diff, p))
def save_single_img(img, count, bt):
p = os.path.join(save_dir, "wgan_single_%d_%d.png" % (count, bt))
save_image(img, img_path=p)
print("wgan single sample id: %d_%d saved" % (count, bt))
count = 0
threshold = 0.1
# batch_buffer = list()
for source_imgs in source_iter:
fake_imgs, real_imgs, d_loss, g_loss, l1_loss = self.generate_fake_samples(
source_imgs)
img_shape = fake_imgs.shape
fake_imgs_reshape = np.reshape(
np.array(fake_imgs),
[img_shape[0], img_shape[1] * img_shape[2] * img_shape[3]])
real_imgs_reshape = np.reshape(
np.array(real_imgs),
[img_shape[0], img_shape[1] * img_shape[2] * img_shape[3]])
fake_imgs_reshape_saved = fake_imgs_reshape
real_imgs_reshape_saved = real_imgs_reshape
# threshold -- fixed
for bt in range(fake_imgs_reshape.shape[0]):
for it in range(fake_imgs_reshape.shape[1]):
if fake_imgs_reshape[bt][it] >= threshold:
fake_imgs_reshape[bt][it] = 1.0
else:
fake_imgs_reshape[bt][it] = -1.0
# valid pixels
for bt in range(fake_imgs_reshape.shape[0]):
p_over = 0
p_less = 0
p_valid = 0
for it in range(fake_imgs_reshape.shape[1]):
if fake_imgs_reshape[bt][
it] == 1.0 and real_imgs_reshape[bt][it] != 1.0:
p_over += 1
if fake_imgs_reshape[bt][it] != 1.0 and real_imgs_reshape[
bt][it] == 1.0:
p_less += 1
if real_imgs_reshape[bt][it] == 1.0:
p_valid += 1
p_accuracy = 1.0 * (p_valid - p_over - p_less) / p_valid
print("wgan count: %d sample %d pixel accuracy: %.05f" %
(count, bt, p_accuracy))
# save ave sample images
for bt in range(fake_imgs_reshape.shape[0]):
fk_reshape = np.reshape(
fake_imgs_reshape_saved[bt],
(fake_imgs.shape[1], fake_imgs.shape[2]))
save_single_img(fk_reshape, count, bt)
# mse, nrmse, ssim and psnr
for bt in range(fake_imgs_reshape.shape[0]):
mse_diff = compare_mse(real_imgs_reshape[bt],
fake_imgs_reshape[bt])
nrmse_diff = compare_nrmse(real_imgs_reshape[bt],
fake_imgs_reshape[bt],
norm_type="Euclidean")
ssim_diff = compare_ssim(real_imgs_reshape[bt],
fake_imgs_reshape[bt])
psnr_diff = compare_psnr(real_imgs_reshape[bt],
fake_imgs_reshape[bt])
print("mse diff:{} | nrmse diff:{} | ssim:{} | psnr:{}".format(
mse_diff, nrmse_diff, ssim_diff, psnr_diff))
# save the images with ssim > 0.8 and ssim < 0.5
if ssim_diff > 0.8 or ssim_diff < 0.5:
fk_reshape = np.reshape(
fake_imgs_reshape_saved[bt],
(1, fake_imgs.shape[1], fake_imgs.shape[2],
fake_imgs.shape[3]))
rl_reshape = np.reshape(
real_imgs_reshape_saved[bt],
(1, real_imgs.shape[1], real_imgs.shape[2],
real_imgs.shape[3]))
fk_reshape = merge(scale_back(fk_reshape), [1, 1])
rl_reshape = merge(scale_back(rl_reshape), [1, 1])
pair = np.concatenate([rl_reshape, fk_reshape], axis=1)
save_img(pair, mse_diff, nrmse_diff, ssim_diff, psnr_diff)
fake_imgs_reshape = np.reshape(fake_imgs_reshape, fake_imgs.shape)
real_imgs_reshape = np.reshape(real_imgs_reshape, real_imgs.shape)
merged_fake_images = merge(scale_back(fake_imgs_reshape),
[source_len, 1])
merged_real_images = merge(scale_back(real_imgs_reshape),
[source_len, 1])
merged_pair = np.concatenate(
[merged_real_images, merged_fake_images], axis=1)
# batch_buffer.append(merged_pair)
count += 1
