def downsample_fn(x, scale=4):
h, w = x.shape[0:2]
hs, ws = h // scale, w // scale
x = imresize(x, size=[hs, ws], interp='bicubic')
x = imresize(x, size=[h, w], interp='bicubic')
return x
def _imresize(self, x, w, h=0, range=[-1., 1.]):
x = x.astype(np.float32)
x = (x - range[0]) * 255. / (range[1] - range[0])
h = w if h == 0 else h
if len(x.shape) == 2:
x = np.expand_dims(x, axis=2)
x = imresize(x, size=[w, h], interp='bicubic', mode=None)
x = (x / 255.) * (range[1] - range[0]) + range[0]
return x
def main(argv=None):
keep_probability = tf.placeholder(tf.float32, name="keep_probabilty")
image = tf.placeholder(tf.float32, shape=[None, IMAGE_SIZE, IMAGE_SIZE, 3], name="input_image")
pred_annotation, _ = inference(image, keep_probability)
sess = tf.Session()
print("Setting up Saver...")
saver = tf.train.Saver()
sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(FLAGS.logs_dir)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
print("Model restored...")
else:
print("Couldn't restore model.")
if FLAGS.mode == "train":
print("train")
cfg = config.TRAIN
elif FLAGS.mode == "valid":
print("valid")
cfg = config.VALID
if not os.path.exists(cfg.segment_preprocessed_path):
os.makedirs(cfg.segment_preprocessed_path)
img_list = sorted(tl.files.load_file_list(path=cfg.hr_img_path, regx='.*.png', printable=False))
rem = len(img_list) % FLAGS.batch_size
for idx in range(0, len(img_list) - rem, FLAGS.batch_size):
b_imgs_list = img_list[idx : idx + FLAGS.batch_size]
load_image = lambda fn, path: downsample_preserve_aspect_ratio_fn(
get_imgs_fn(fn, path), size=[256, 256], mean_center=False)
b_imgs = tl.prepro.threading_data(b_imgs_list, fn=load_image, path=cfg.hr_img_path)
pred = sess.run(pred_annotation,
feed_dict = {image: b_imgs, keep_probability: 1.0})
pred = np.squeeze(pred, axis=3)
# for i in range(FLAGS.batch_size):
# plt.subplot(1, 2, 1)
# plt.imshow(b_imgs[0])
# plt.subplot(1, 2, 2)
# plt.imshow(pred[0].astype(np.uint8), cmap=cm.Paired, vmin=0, vmax=151)
# plt.savefig('compare_{}.png'.format(i))
# plt.show()
for i in range(FLAGS.batch_size):
img = np.expand_dims(pred[i].astype(np.uint8), axis=2)
img = imresize(img, size=[96, 96], interp='nearest', mode=None)
img = np.squeeze(img, axis=2)
one_hot = segment_helper.label_to_one_hot(img)
segment_helper.save_one_hot(one_hot, cfg.segment_preprocessed_path,
get_frame_key(b_imgs_list[i]) + '_seg.npy')
print('saved %d from %s' % (FLAGS.batch_size + idx, cfg.hr_img_path))
def resize_fn(x, scaled_size=256):
# We obtained the LR images by downsampling the HR images using bicubic kernel with downsampling factor r = 4.
scale_ratio = max(scaled_size / x.shape[0],
scaled_size / x.shape[1]) + 0.05
height = int(x.shape[0] * scale_ratio)
width = int(x.shape[1] * scale_ratio)
x = imresize(x, size=[width, height], interp='bicubic')
return x
def upsample_fn(x, scale=4):
h, w = x.shape[0:2]
newh, neww = h * scale, w * scale
x = imresize(x, size=(newh, neww), interp='bicubic')
return x
def downsample_fn(x):
# We obtained the LR images by downsampling the HR images using bicubic kernel with downsampling factor r = 4.
x = imresize(x, size=[96, 96], interp='bicubic')
x /= 255. / 2.
x -= 1.
return x