def auto_cropping(origin_image):
batch_size = len(origin_image)
terminals = np.zeros(batch_size)
ratios = np.repeat([[0, 0, 20, 20]], batch_size, axis=0)
img = crop_input(origin_image, generate_bbox(origin_image, ratios))
global_feature = sess.run(global_feature_placeholder,
feed_dict={image_placeholder: img})
h_np = np.zeros([batch_size, 1024])
c_np = np.zeros([batch_size, 1024])
while True:
action_np, h_np, c_np = sess.run(
(action, h, c),
feed_dict={
image_placeholder: img,
global_feature_placeholder: global_feature,
h_placeholder: h_np,
c_placeholder: c_np
})
ratios, terminals = command2action(action_np, ratios, terminals)
bbox = generate_bbox(origin_image, ratios)
if np.sum(terminals) == batch_size:
return bbox
img = crop_input(origin_image, bbox)
def step(self, action, steps, last_score, ratios, dones):
ratios, dones = command2action([action], ratios, dones)
bbox = generate_bbox(np.expand_dims(self.origin_img, axis=0), ratios)
# import pdb; pdb.set_trace();
next_img = crop_input(np.expand_dims(self.origin_img, axis=0), bbox)
cur_score, next_ob = self.score_feature(next_img[0])
rew = 1 if cur_score>last_score else -1
rew -= 0.001*(steps+1)
rew -= 5 if (bbox[0][2] - bbox[0][0]) > 2*(bbox[0][3] - bbox[0][1]) \
or (bbox[0][2] - bbox[0][0]) < 0.5*(bbox[0][3] - bbox[0][1]) else 0
return next_ob, rew, cur_score, ratios, dones