test_predict = total_image_predict(
ori_image_path=args.test_image_path,
input_placeholder=image,
logits_prob_node=logits_prob,
is_training_placeholder=model._is_training,
sess=sess,
multi_scale=args.multi_scale)
test_label = cv2.imread(args.test_label_path, cv2.IMREAD_GRAYSCALE)
# 保存图片
cv2.imwrite(filename='%spredict_color_%d.png' % (log_path, step),
img=color_predicts(img=test_predict))
result = iou(y_pre=np.reshape(test_predict, -1),
y_true=np.reshape(test_label, -1))
print("======================%d======================" % step)
for key in result.keys():
offset = 40 - key.__len__()
print(key + ' ' * offset + '%.4f' % result[key])
test_summary = tf.Summary(value=[
tf.Summary.Value(tag=key, simple_value=result[key])
for key in result.keys()
])
# 记录summary
summary_writer.add_summary(test_summary, step)
summary_writer.flush()
saver.save(sess, model_path, write_meta_graph=True, global_step=100000)
(512, 512)).reshape((512, 512, 1))
img = np.concatenate((img_rgb, img_thermal), axis=2)
else:
img_rgb = cv2.imread(line[0], cv2.IMREAD_COLOR)
img_rgb = cv2.resize(img_rgb, (512, 512))
img = img_rgb
test_predict = multi_scale_predict(
image=img,
input_placeholder=image,
is_training_placeholder=model._is_training,
logits_prob_node=logits_prob,
sess=sess,
multi=False,
size=512)
g_truth = cv2.resize(np.array(Image.open(line[2])), (512, 512))
result = iou(y_pre=np.reshape(test_predict, -1),
y_true=np.reshape(g_truth, -1))
label_valid = np.unique(g_truth)
for key in sorted(result.keys()):
if key in key_map and not key_map[key] in label_valid:
continue
if np.isnan(result[key]) or result[key] != result[key]:
continue
if not key in all_tests_result:
all_tests_result[key] = [result[key]]
else:
all_tests_result[key].append(result[key])
test_label = color_predicts(img=g_truth)
test_predict = color_predicts(img=test_predict)
full_image = np.concatenate((img_rgb, test_label, test_predict),
axis=1)
out_name = line[0].split('/')[-1].split('.')[0] + '_predict.jpg'
