def get_val_mIoU():
print("Start to get mIoU on val set...")
#prediction_files = glob.glob(os.path.join(FLAGS.saved_prediction_val_gray, '*.png'))
#f = open(VAL_LIST)
#lines = f.readlines()
annotation_files = glob.glob(
os.path.join(CITYSCAPE_ANNO_DIR, 'val/*/*_gtFine_labelTrainIds.png'))
prediction_files = [
os.path.join(FLAGS.saved_prediction_val_gray,
os.path.join(os.path.basename(filename)))
for filename in annotation_files
]
prediction_files = [
filename.replace('_gtFine_labelTrainIds.png', '_leftImg8bit.png')
for filename in prediction_files
]
for i, annotation_file in enumerate(annotation_files):
annotation_data = cv2.imread(annotation_file, cv2.IMREAD_GRAYSCALE)
annotation_data = annotation_data.reshape(-1)
if i == 0:
annotations_data = annotation_data
else:
annotations_data = np.concatenate(
(annotations_data, annotation_data))
print(annotations_data.shape)
for i, prediction_file in enumerate(prediction_files):
prediction_data = cv2.imread(prediction_file, cv2.IMREAD_GRAYSCALE)
prediction_data = prediction_data.reshape(-1)
if i == 0:
predictions_data = prediction_data
else:
predictions_data = np.concatenate(
(predictions_data, prediction_data))
print(predictions_data.shape)
mIoU, IoU = Utils.cal_batch_mIoU(predictions_data, annotations_data,
FLAGS.classes)
print(mIoU)
print(IoU)
def get_val_mIoU():
print("Start to get mIoU on val set...")
f = open(VAL_LIST)
lines = f.readlines()
annotation_files = [
os.path.join(ANNOTATION_PATH,
line.strip() + '.png') for line in lines
]
prediction_files = [
os.path.join(saved_prediction_val_gray,
line.strip() + '.png') for line in lines
]
for i, annotation_file in enumerate(annotation_files):
annotation_data = cv2.imread(annotation_file, cv2.IMREAD_GRAYSCALE)
annotation_data = annotation_data.reshape(-1)
if i == 0:
annotations_data = annotation_data
else:
annotations_data = np.concatenate(
(annotations_data, annotation_data))
print(annotations_data.shape)
for i, prediction_file in enumerate(prediction_files):
prediction_data = cv2.imread(prediction_file, cv2.IMREAD_GRAYSCALE)
prediction_data = prediction_data.reshape(-1)
if i == 0:
predictions_data = prediction_data
else:
predictions_data = np.concatenate(
(predictions_data, prediction_data))
print(predictions_data.shape)
mIoU, IoU = Utils.cal_batch_mIoU(predictions_data, annotations_data,
CLASSES)
print(mIoU)
print(IoU)
keep_prob: 1.0
})
learning_rate = sess.run(lr)
if i % FLAGS.print_steps == 0 and FLAGS.dataset == 'train':
print(datetime.datetime.now().strftime("%Y.%m.%d-%H:%M:%S"),
" | Step: %d | Train loss all: %f" % (i, train_loss_val_all))
if i % FLAGS.print_steps == 0 and FLAGS.dataset == 'trainval':
print(datetime.datetime.now().strftime("%Y.%m.%d-%H:%M:%S"),
" | Step: %d | Train loss all: %f" % (i, train_loss_val_all))
if i % FLAGS.iou_steps == 0 and FLAGS.dataset == 'train':
train_mIoU_val, train_IoU_val = Utils.cal_batch_mIoU(
pred_train, b_anno, FLAGS.classes)
test_mIoU_val, test_IoU_val = Utils.cal_batch_mIoU(
pred_test, b_anno_test, FLAGS.classes)
sess.run(tf.assign(train_mIoU, train_mIoU_val))
sess.run(tf.assign(test_mIoU, test_mIoU_val))
print('------------------------------')
print(
"Step: %d | Lr: %f | Train loss all: %f | Train loss: %f | Train mIoU: %f | Test loss all: %f | Test loss: %f | Test mIoU: %f"
% (i, learning_rate, train_loss_val_all, train_loss_val,
train_mIoU_val, test_loss_val_all, test_loss_val,
test_mIoU_val))
print('------------------------------')
print(train_IoU_val)
mIoU_val_all = []
for i in range(NUMS):
b_image, b_anno, b_filename = sess.run(
[image_batch, anno_batch, filename])
b_image_0, b_anno = input_data.aug_std(b_image, b_anno, type='test')
pred = sess.run(prediction,
feed_dict={
x: b_image_0,
y: b_anno,
keep_prob: 1.0
})
mIoU_val, IoU_val = Utils.cal_batch_mIoU(pred, b_anno, CLASSES)
#mIoU_val = sess.run(mIoU)
#print(mIoU_val)
mIoU_val_all.append(mIoU_val)
basename = b_filename[0].split('.')[0]
'''
# save raw image, annotation, and prediction
pred = pred.astype(np.uint8)
b_anno = b_anno.astype(np.uint8)
pred_color = color_gray(pred[0, :, :])
b_anno_color = color_gray(b_anno[0, :, :])
img = Image.fromarray(b_image[0])
anno = Image.fromarray(b_anno_color)
test_summary = sess.run(merged, feed_dict={x: image_batch_val, y: anno_batch_val})
test_summary_writer.add_summary(test_summary, i)
if i % PRINT_STEPS == 0:
train_loss_val_all = sess.run(loss_all, feed_dict={x: image_batch, y: anno_batch})
print(datetime.datetime.now().strftime("%Y.%m.%d-%H:%M:%S"), " | Step: %d, | Train loss all: %f" % (i, train_loss_val_all))
if i % SAVE_SUMMARY_STEPS == 0:
learning_rate = sess.run(lr)
pred_train, train_loss_val_all, train_loss_val = sess.run([predictions, loss_all, loss],
feed_dict={x: image_batch, y: anno_batch})
pred_test, test_loss_val_all, test_loss_val = sess.run([predictions, loss_all, loss],
feed_dict={x: image_batch_val, y: anno_batch_val})
train_mIoU_val, train_IoU_val = Utils.cal_batch_mIoU(pred_train, anno_batch, CLASSES)
test_mIoU_val, test_IoU_val = Utils.cal_batch_mIoU(pred_test, anno_batch_val, CLASSES)
sess.run(tf.assign(train_mIoU, train_mIoU_val))
sess.run(tf.assign(test_mIoU, test_mIoU_val))
print('------------------------------')
print(datetime.datetime.now().strftime("%Y.%m.%d-%H:%M:%S"), " | Step: %d, | Lr: %f, | train loss all: %f, | train loss: %f, | train mIoU: %f, | test loss all: %f, | test loss: %f, | test mIoU: %f" % (
i, learning_rate, train_loss_val_all, train_loss_val, train_mIoU_val, test_loss_val_all, test_loss_val, test_mIoU_val))
print('------------------------------')
print(train_IoU_val)
print(test_IoU_val)
print('------------------------------')
#prediction = tf.argmax(logits, axis=-1, name='predictions')
if i % SAVE_CHECKPOINT_STEPS == 0:
if not os.path.exists('images'):
x: b_image_test,
y: b_anno_test,
keep_prob: 1.0
})
learning_rate = sess.run(lr)
if i % 10 == 0:
print(
"train step: %d, learning rate: %f, train loss all: %f, train loss: %f, test loss all: %f, test loss: %f,"
% (i, learning_rate, train_loss_val_all, train_loss_val,
test_loss_val_all, test_loss_val))
if i % 200 == 0:
train_mIoU_val, train_IoU_val = Utils.cal_batch_mIoU(
pred_train, b_anno, CLASSES)
test_mIoU_val, test_IoU_val = Utils.cal_batch_mIoU(
pred_test, b_anno_test, CLASSES)
sess.run(tf.assign(train_mIoU, train_mIoU_val))
sess.run(tf.assign(test_mIoU, test_mIoU_val))
print(
"train step: %d, learning rate: %f, train loss all: %f, train loss: %f, train mIoU: %f, test loss all: %f, test loss: %f, test mIoU: %f"
% (i, learning_rate, train_loss_val_all, train_loss_val,
train_mIoU_val, test_loss_val_all, test_loss_val,
test_mIoU_val))
print(train_IoU_val)
print(test_IoU_val)
if i % 6000 == 0:
