def load_data(folder,
size,
first=0,
last=None,
read_flag=cv2.IMREAD_UNCHANGED):
return list(
map(lambda x: cv2.resize(cv2.imread(x, read_flag), size) / 255,
sorted(listdir_fullpath(folder))[first:last]))
def load_fixation_locs(folder, size=(224, 224)):
maps = []
for fix_map_name in listdir_fullpath(folder):
fix_map = np.array(scipy.io.loadmat(fix_map_name)['fixLocs'])
y_size, x_size = fix_map.shape
ones = np.where(fix_map > 0)
resized_map = np.zeros(size)
for y, x in zip(ones[0], ones[1]):
resized_map[int(y / y_size * size[0]),
int(x / x_size * size[1])] = 1
maps.append(resized_map)
return np.array(maps)
def predict(heatmap_type='jet'):
# creates model of neural network
prediction = neural_network_model(x, keep_prob)
cross_entropy = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=prediction, labels=y))
images_for_prediction = np.array(
load_data(args.images, (64, 64), first=5000))
images_original = sorted(listdir_fullpath(args.images))[5000:]
original_maps = np.array(load_data(args.maps, (64, 64), first=5000))
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, args.model) # restores saved model
predicted_heatmaps = make_prediction(prediction, {
x: images_for_prediction,
keep_prob: 1.0
})
binary_maps = np.array(
get_binary_fixation_maps(args.binary_maps, size=64, first=5000))
count_metrics(predicted_heatmaps, original_maps, binary_maps)
i = 0
for map, img in zip(predicted_heatmaps, images_original):
i += 1
#print ("working on: " + str(img.rsplit('/', 1)[1]))
# saving predicted heatmaps on image
visualize_heatmap(map, img, args.save_to, heatmap_type)
i += 1
p = map
print(str(i) + ". saving " + str(img.rsplit('/', 1)[1]))
plt.imshow(p, cmap='jet')
toimage(p).save("predicted_maps/" + str(img.rsplit('/', 1)[1]))
plt.savefig("predicted_maps/plot_" + str(img.rsplit('/', 1)[1]))
if i == 50:
break
def get_binary_fixation_maps(folder, size=256, first=0, last=None):
ret_maps = []
extracted_data = list(
map(
lambda x: {
'file': x.rsplit('/', 1)[1].replace(".mat", ".jpg"),
**get_size_and_fix(loadmat(x)['s'])
},
sorted(listdir_fullpath(folder))[first:last]))
for data in extracted_data:
#print("working on file: " + data['file'])
final_map = np.zeros([size, size])
for i in range(1, 6):
user_key = 'user_' + str(i)
user_fix = data[user_key]
final_map = apply_fixations(data['size'], data[user_key],
final_map)
ret_maps.append(final_map)
return ret_maps
config = tf.ConfigProto(intra_op_parallelism_threads=12,\
inter_op_parallelism_threads=12, allow_soft_placement=True,\
device_count = {'CPU' : 1, 'GPU' : 0})
config.gpu_options.allow_growth = True
session = tf.Session(config=config)
K.set_session(session)
#model = load_model('adadelta+binary.save')
model = load_model(args.model)
model.compile(loss=args.loss, optimizer=args.optimizer, metrics=['accuracy'])
model.summary()
imgs = np.array(load_data(args.images, (n, n), read_flag=cv2.IMREAD_COLOR))
img_names = sorted(listdir_fullpath(args.images))
predicted = model.predict(imgs, verbose=1)
i = 0
if not os.path.exists("predicted_maps"):
os.makedirs("predicted_maps")
for p, img in zip(predicted, img_names):
i += 1
#print(str(i) + ". saving " + str(img.rsplit('/', 1)[1]))
plt.imshow(p, cmap='jet')
if not os.path.exists("predicted_maps/" + args.model.rsplit("/", 2)[-1]):
os.makedirs("predicted_maps/" + args.model.rsplit("/", 2)[-1])
toimage(p).save("predicted_maps/" + args.model.rsplit("/", 2)[-1] + "/" +
str(img.rsplit('/', 1)[1]))
x = Conv2D(filters=1, kernel_size=5, activation='sigmoid', padding='same')(x)
print(K.int_shape(x))
x = Reshape((n, n))(x)
final = Model(input_layer, x)
small_model = None
autoencoder = None
final.compile(optimizer=args.optimizer, loss=args.loss)
final.summary()
split = int(0.85 * args.samples)
imgs_paths = sorted(listdir_fullpath(args.images))
maps_paths = sorted(listdir_fullpath(args.maps))
train_imgs_paths = imgs_paths[:split]
train_maps_paths = maps_paths[:split]
valid_imgs_paths = imgs_paths[split:]
valid_maps_paths = maps_paths[split:]
print("number of train samples: " + str(len(train_imgs_paths)))
print("number of valid samples: " + str(len(valid_imgs_paths)))
train_img_map_ds = load_images_and_maps(train_imgs_paths, train_maps_paths,
(n, n))
valid_img_map_ds = load_images_and_maps(valid_imgs_paths, valid_maps_paths,
(n, n))