def inference_with_restored_model(net,
image,
label,
bbox_gt=None,
verbose=True,
times=1,
gt_prob=0.8):
if verbose:
print('Processing image with shape%s' % (image.shape, ))
bbox_pred = net.inference_box(image)
bbox_pred = my_utils.bbox_xy_to_tlwh(bbox_pred, size=image.shape[:2])
prediction = crf.crf_from_bbox(image, bbox=bbox_pred, gt_prob=gt_prob)
if times > 1:
for i in range(times - 1):
unary = crf.get_unary_term(prediction,
unary_from='label',
n_classes=2,
gt_prob=gt_prob)
prediction = crf.crf_post_process(image, unary)
result = my_utils.count_many(prediction, label)
prediction = {
'label': prediction,
'bbox': bbox_pred,
}
if bbox_gt is not None:
prediction['IoU'] = my_utils.calc_bbox_iou(bbox_pred, bbox_gt)
return result, prediction
def crf_eval_one(mm, image, label):
out, prob = mm.inference(image, ['mask', 'prob'])
out, prob = np.squeeze(out), np.squeeze(prob)
unary = crf.get_unary_term(prob)
crf_out = crf.crf_post_process(image, unary)
result = my_utils.count_many(predictions=crf_out, labels=label)
return result
def test_five():
config = utils.load_config()
dermis = inputs.SkinData(config['data_dir'], 'dermis')
with tf.Graph().as_default():
global_step = tf.train.get_or_create_global_step()
image_ph = tf.placeholder(dtype=tf.float32, shape=(None, None, 3))
images = tf.expand_dims(image_ph, axis=0)
net = FCN(images, net_params=config['net_params'])
h, w = tf.shape(image_ph)[0], tf.shape(image_ph)[1]
upscore = tf.image.resize_images(net.endpoints['conv4'], size=(h, w))
prob_one = tf.nn.sigmoid(upscore)
prob_zero = 1 - prob_one
probs = tf.concat([prob_zero, prob_one], axis=3)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess,
tf.train.latest_checkpoint(config['train_dir']))
logger.info('Model at step-%d restored successfully!' %
sess.run(global_step))
utils.create_if_not_exists(config['save_path'])
result_before = {'TP': 0, 'TN': 0, 'FP': 0, 'FN': 0}
result_after = result_before.copy()
def update_dict(d, to_update):
for key in to_update:
d[key] += to_update[key]
_, x_test, _, y_test = train_test_split(
dermis.images,
dermis.labels,
random_state=config['split_seed'])
for i, (image, label) in enumerate(zip(x_test, y_test)):
prep_image = image_prep_for_test(image)
probs_o = np.squeeze(
sess.run(probs, feed_dict={image_ph: prep_image}))
cnn_result = np.argmax(probs_o, axis=2)
cnn_crf_result = crf_post_process(image, probs_o)
result_bi = count_many(cnn_result, label)
result_ai = count_many(cnn_crf_result, label)
update_dict(result_before, result_bi)
update_dict(result_after, result_ai)
result_bi.update(metric_many_from_counter(result_bi))
result_ai.update(metric_many_from_counter(result_ai))
if i % 5 == 0:
logger.info(
'Image-{}\nresult before\n{}\nresult after\n{}\n'.
format(i, result_bi, result_ai))
result_before.update(metric_many_from_counter(result_before))
result_after.update(metric_many_from_counter(result_after))
logger.info('\nresult before\n{}\nresult after\n{}\n'.format(
result_before, result_after))
def test_four():
config = utils.load_config()
dermis = inputs.SkinData(config['data_dir'], 'dermis')
with tf.Graph().as_default():
global_step = tf.train.get_or_create_global_step()
image_ph = tf.placeholder(dtype=tf.float32, shape=(None, None, 3))
images = tf.expand_dims(image_ph, axis=0)
net = FCN(images, net_params=config['net_params'])
h, w = tf.shape(image_ph)[0], tf.shape(image_ph)[1]
upscore = tf.image.resize_images(net.endpoints['conv4'], size=(h, w))
prob_one = tf.nn.sigmoid(upscore)
prob_zero = 1 - prob_one
probs = tf.concat([prob_zero, prob_one], axis=3)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess,
tf.train.latest_checkpoint(config['train_dir']))
logger.info('Model at step-%d restored successfully!' %
sess.run(global_step))
utils.create_if_not_exists(config['save_path'])
total_count = 0
true_count = 0
for i, (image,
label) in enumerate(zip(dermis.images, dermis.labels)):
prep_image = image_prep_for_test(image)
probs_o = np.squeeze(
sess.run(probs, feed_dict={image_ph: prep_image}))
cnn_result = np.argmax(probs_o, axis=2)
accuracy_before, _, _ = metric_accuracy(cnn_result, label)
cnn_crf_result = crf_post_process(image, probs_o)
accuracy_i, true_count_i, total_count_i = metric_accuracy(
cnn_crf_result, label)
true_count += true_count_i
total_count += total_count_i
ss = 'DOWN' if accuracy_before > accuracy_i else 'UP'
path = os.path.join(config['save_path'],
dermis.listing[i].split('/')[-1])
path = '{:<6} ({:.3f}) ({:.3f}) {}.jpg'.format(
path, accuracy_before, accuracy_i, ss)
save_all_two(image, label, cnn_result, cnn_crf_result, path)
if i % 5 == 0:
logger.info(
'Image-%d accuracy before(%.3f) after(%.3f) %s' %
(i, accuracy_before, accuracy_i, ss))
accuracy = true_count * 1.0 / total_count
logger.info('Accuracy after crf: %.3f' % accuracy)
def get_cnn_crf_result(image, bbox, lesion_prob):
h, w, _ = image.shape
p0 = np.ones((h, w))
p1 = np.zeros((h, w))
top, left, height, width = bbox
lesion_prob = np.squeeze(lesion_prob)
assert (height, width
) == lesion_prob.shape[:-1], "prob shape%s not equal to bbox%s" % (
lesion_prob.shape[:-1], bbox)
p0[top:top + height, left:left + width] = lesion_prob[:, :, 0]
p1[top:top + height, left:left + width] = lesion_prob[:, :, 1]
prob = np.stack([p0, p1], axis=2)
result = crf_post_process(image, prob)
return result