if (t == (int(db_helen['testset']['img'].shape[0]/testing_batch)+1)):
t_batch_x = db_helen['testset']['img'][(t*testing_batch):]
t_batch_y = db_helen['testset']['pts'][(t*testing_batch):]
batch_map=sess.run(pred_annotation,
feed_dict={image: t_batch_x/255,
keep_probability: 1.0,
train_phase:conf.training})
if (t == 0):
infered_pts = utils.map2pts(batch_map , gau = True)
else:
infered_pts = np.concatenate((infered_pts, utils.map2pts(batch_map, gau = True)), axis=0)
used_time = time.time()-start_time
print('Avg. inference time: %.4f' % (used_time/db_helen['testset']['img'].shape[0]))
for idx, content in enumerate(zip(db_helen['testset']['img'],infered_pts)):
img = content[0].copy()
for kp_idx, keypoint in enumerate(content[1]):
cv2.circle(img,(keypoint[0],keypoint[1]), 2, (0,255,0), -1)
cv2.imwrite(pred_dir + str(idx)+ '.png', img)
with open(model_dir + 'pts_'+ str(conf.testing)+ '.pickle', 'wb') as handle:
pickle.dump(infered_pts, handle)
norm_error_image, norm_error_image_eye = utils.eval_norm_error_image(infered_pts, db_helen['testset']['pts'])
pandas.DataFrame({'loss':norm_error_image,'loss_eye':norm_error_image_eye}).to_csv(model_dir + 'norm_error_image_' + str(conf.testing)+ '.csv')
sess.close()
infered_pts, acc_valid= sess.run([y_out_point, avg_losses],
feed_dict=feed_dict)
if (t == 0):
inferred_map = infered_pts
else:
inferred_map = np.concatenate((inferred_map, infered_pts), axis=0)
used_time = time.time()-start_time
print('Avg. inference time: %.4f' % (used_time/db_helen['testset']['img'].shape[0]))
inferred_map = np.asarray(inferred_map)
pts_maps = np.reshape(inferred_map, newshape=(-1,inferred_map.shape[1],inferred_map.shape[2]))
norm_error_image, norm_error_image_eye= utils.eval_norm_error_image(pts_maps, db_helen['testset']['pts'])
with open(model_dir + 'pts_'+ str(conf.testing)+ '.pickle', 'wb') as handle:
pickle.dump(pts_maps, handle)
pandas.DataFrame({'loss':norm_error_image,'loss_eye':norm_error_image_eye}).to_csv(model_dir + 'norm_error_image_' + str(conf.testing)+ '.csv')
for idx, content in enumerate(zip(db_helen['testset']['img'],pts_maps)):
img = content[0].copy()
for kp_idx, keypoint in enumerate(content[1]):
cv2.circle(img,(keypoint[0],keypoint[1]), 2, (0,255,0), -1)
cv2.imwrite(pred_dir + str(idx)+ '.png', img)
sess.close()
