def test():
args = cmd.args()
if args.load_generated_images:
print ("Loading images...")
input_images, target_images, generated_images, names = load_generated_images(args.generated_images_dir)
else:
print ("Generate images...")
from keras import backend as K
if args.use_dropout_test:
K.set_learning_phase(1)
dataset = PoseHMDataset(test_phase=True, **vars(args))
generator = make_generator(args.image_size, args.use_input_pose, args.warp_skip, args.disc_type, args.warp_agg,
args.use_bg, args.pose_rep_type)
assert (args.generator_checkpoint is not None)
generator.load_weights(args.generator_checkpoint)
input_images, target_images, generated_images, names = generate_images(dataset, generator, args.use_input_pose)
print ("Save images to %s..." % (args.generated_images_dir, ))
save_images(input_images, target_images, generated_images, names,
args.generated_images_dir)
print ("Compute inception score...")
inception_score = get_inception_score(generated_images)
print ("Inception score %s" % inception_score[0])
print ("Compute structured similarity score (SSIM)...")
structured_score = ssim_score(generated_images, target_images)
print ("SSIM score %s" % structured_score)
print ("Compute l1 score...")
norm_score = l1_score(generated_images, target_images)
print ("L1 score %s" % norm_score)
print ("Compute masked inception score...")
generated_images_masked = create_masked_image(names, generated_images, args.annotations_file_test)
reference_images_masked = create_masked_image(names, target_images, args.annotations_file_test)
inception_score_masked = get_inception_score(generated_images_masked)
print ("Inception score masked %s" % inception_score_masked[0])
print ("Compute masked SSIM...")
structured_score_masked = ssim_score(generated_images_masked, reference_images_masked)
print ("SSIM score masked %s" % structured_score_masked)
print ("Inception score = %s, masked = %s; SSIM score = %s, masked = %s; l1 score = %s" %
(inception_score, inception_score_masked, structured_score, structured_score_masked, norm_score))
def compute_scores(epoch, image_shape, generator, dataset, images_inception=50000, images_fid=10000,
log_file=None, cache_file='mnist_fid.npz', additional_info=""):
compute_inception = images_inception != 0
compute_fid = images_fid != 0
number_of_images = max(images_inception, images_fid)
if not (compute_inception or compute_fid):
return
images = np.empty((number_of_images, ) + image_shape)
labels = np.empty((number_of_images, ))
generator_input = generator.get_input_at(0)
if type(generator_input) != list:
generator_input = [generator_input]
predict_fn = K.function(generator_input + [K.learning_phase()], [generator.get_output_at(0)])
bs = dataset._batch_size
conditional = False
for begin in tqdm(range(0, number_of_images, bs)):
end = min(number_of_images, begin + bs)
n_images = end - begin
g_s = dataset.next_generator_sample_test()
if len(g_s) == 2:
labels[begin:end] = np.squeeze(g_s[1], axis=1)[:n_images]
conditional = True
images[begin:end] = predict_fn(g_s + [False])[0][:n_images]
images *= 127.5
images += 127.5
def to_rgb(array):
if array.shape[-1] != 3:
#hack for grayscale mnist
return np.concatenate([array, array, array], axis=-1)
else:
return array
#save_images('baseline_16', to_rgb(images), labels)
draw_grid(os.path.join(os.path.dirname(log_file), "epoch_%s_imagegrid.png" % epoch),
to_rgb(images), labels if conditional else None)
if compute_inception:
str = "INCEPTION SCORE: %s, %s" % get_inception_score(to_rgb(images[:images_inception]))
print (str)
if log_file is not None:
with open(log_file, 'a') as f:
print >>f, ("Epoch %s " % (epoch, )) + str #+ " " + additional_info
if compute_fid:
true_images = 127.5 * dataset._X_test + 127.5
str = "FID SCORE: %s" % calculate_fid_given_arrays([to_rgb(true_images)[:images_fid],
to_rgb(images)[:images_fid]], cache_file=cache_file)
print (str)
if log_file is not None:
with open(log_file, 'a') as f:
print >>f, ("Epoch %s " % (epoch, )) + str #+ " " + additional_info
def compute_scores(epoch, image_shape, generator, dataset, number_of_images=50000, compute_inception=True, compute_fid=True,
log_file=None, cache_file='mnist_fid.npz', additional_info=""):
if not (compute_inception or compute_fid):
return
images = np.empty((number_of_images, ) + image_shape)
generator_input = generator.get_input_at(0)
if type(generator_input) != list:
generator_input = [generator_input]
generator_output = generator.get_output_at(0)
if type(generator_output) != list:
generator_output = [generator_output]
predict_fn = K.function(generator_input + [K.learning_phase()], [generator_output[0]])
for begin in tqdm(range(0, number_of_images, dataset._batch_size)):
end = min(number_of_images, begin + dataset._batch_size)
n_images = end - begin
g_s = dataset.next_generator_sample_test()
images[begin:end] = predict_fn(g_s + [False])[0][:n_images]
images *= 127.5
images += 127.5
def to_rgb(array):
if array.shape[-1] != 3:
#hack for grayscale mnist
return np.concatenate([array, array, array], axis=-1)
else:
return array
if compute_inception:
str = "INCEPTION SCORE: %s, %s" % get_inception_score(to_rgb(images))
print (str)
if log_file is not None:
with open(log_file, 'a') as f:
print >>f, ("Epoch %s " % (epoch, )) + str + " " + additional_info
if compute_fid:
true_images = 127.5 * dataset._X_test + 127.5
str = "FID SCORE: %s" % calculate_fid_given_arrays([to_rgb(true_images)[:number_of_images],
to_rgb(images)[:number_of_images]], cache_file=cache_file)
print (str)
if log_file is not None:
with open(log_file, 'a') as f:
print >>f, ("Epoch %s " % (epoch, )) + str + " " + additional_info
def test():
args = cmd.args()
if args.load_generated_images:
print("Loading images...")
input_images, target_images, generated_images, names = load_generated_images(
args.generated_images_dir)
else:
print("Generate images...")
from keras import backend as K
if args.use_dropout_test:
K.set_learning_phase(1)
dataset = PoseHMDataset(test_phase=True, **vars(args))
generator = make_generator(args.image_size, args.use_input_pose,
args.warp_skip, args.disc_type,
args.warp_agg)
assert (args.generator_checkpoint is not None)
generator.load_weights(args.generator_checkpoint)
input_images, target_images, generated_images, names = generate_images(
dataset, generator, args.use_input_pose)
print("Save images to %s..." % (args.generated_images_dir, ))
save_images(input_images, target_images, generated_images, names,
args.generated_images_dir)
if args.is_test_real_data:
generated_images = target_images
print('Compute FID score')
fid = FID_score()
fid_score = fid.calculate_fid_images(generated_images, target_images)
print("FID score %s" % fid_score)
print("Compute inception score...")
inception_score = get_inception_score(generated_images)
print("Inception score %s" % inception_score[0])
print("Compute structured similarity score (SSIM)...")
structured_score = ssim_score(generated_images, target_images)
print("SSIM score %s" % structured_score)
print("Compute l1 score...")
norm_score = l1_score(generated_images, target_images)
print("L1 score %s" % norm_score)
def test():
args = cmd.args()
if args.load_generated_images:
print("Loading images...")
input_images, target_images, generated_images, names = load_generated_images(
args.generated_images_dir)
else:
print("Generate images...")
from keras import backend as K
if args.use_dropout_test:
K.set_learning_phase(1)
dataset = PoseHMDataset(test_phase=True, **vars(args))
generator = make_generator(args.image_size, args.use_input_pose,
args.warp_agg, args.num_landmarks,
args.num_mask)
assert (args.generator_checkpoint is not None)
generator.load_weights(args.generator_checkpoint)
input_images, pose_inp_array, out_pose, pose_out_array, inp_pose, target_images, generated_images, names = generate_images(
dataset, generator, args.use_input_pose)
print("Save images to %s..." % (args.generated_images_dir, ))
save_images(input_images, pose_inp_array, out_pose, pose_out_array,
inp_pose, target_images, generated_images, names,
args.generated_images_dir)
print("Compute inception score...")
inception_score = get_inception_score(generated_images)
print("Inception score %s" % inception_score[0])
# print ("Compute Frechet distance...")
# fid.create_inception_graph('/tmp/imagenet/classify_image_graph_def.pb') # load the graph into the current TF graph
# with tf.Session() as sess:
# sess.run(tf.global_variables_initializer())
# mu_gen, sigma_gen = fid.calculate_activation_statistics(generated_images, sess, batch_size=100)
# mu_real, sigma_real = fid.calculate_activation_statistics(target_images, sess, batch_size=100)
#
# fid_value = fid.calculate_frechet_distance(mu_gen, sigma_gen, mu_real, sigma_real)
# print ("Frechet distance %s" % fid_value)
print("Compute structured similarity score (SSIM)...")
structured_score = ssim_score(generated_images, target_images)
print("SSIM score %s" % structured_score)
print("Compute l1 score...")
norm_score = l1_score(generated_images, target_images)
print("L1 score %s" % norm_score)
print("Compute masked inception score...")
generated_images_masked = create_masked_image(names, generated_images,
args.annotations_file_test)
reference_images_masked = create_masked_image(names, target_images,
args.annotations_file_test)
inception_score_masked = get_inception_score(generated_images_masked)
print("Inception score masked %s" % inception_score_masked[0])
# print ("Compute masked Frechet distance...")
# fid.create_inception_graph('/tmp/imagenet/classify_image_graph_def.pb') # load the graph into the current TF graph
# with tf.Session() as sess:
# sess.run(tf.global_variables_initializer())
# mu_gen, sigma_gen = fid.calculate_activation_statistics(generated_images_masked, sess, batch_size=100)
# mu_real, sigma_real = fid.calculate_activation_statistics(reference_images_masked, sess, batch_size=100)
#
# fid_value = fid.calculate_frechet_distance(mu_gen, sigma_gen, mu_real, sigma_real)
# print("Frechet distance masked %s" % fid_value)
print("Compute masked SSIM...")
structured_score_masked = ssim_score(generated_images_masked,
reference_images_masked)
print("SSIM score masked %s" % structured_score_masked)
print(
"Inception score = %s, masked = %s; SSIM score = %s, masked = %s; l1 score = %s"
% (inception_score, inception_score_masked, structured_score,
structured_score_masked, norm_score))
