def generate_roi_feature_dataset(dataset,
netname,
model_path,
feature_save_path,
using_attribute_flag=True,
using_clstm_flag=True):
'''
对dataset下面的每一个slice生成测试的结果
:param dataset: dataset的路径
:return:
'''
nc_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='nc_roi_placeholder')
art_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='art_roi_placeholder')
pv_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='pv_roi_placeholder')
nc_patch_placeholder = tf.placeholder(
tf.float32,
[None, config.PATCH_IMAGE_HEIGHT, config.PATCH_IMAGE_WIDTH, 3],
name='nc_patch_placeholder')
art_patch_placeholder = tf.placeholder(
tf.float32,
[None, config.PATCH_IMAGE_HEIGHT, config.PATCH_IMAGE_WIDTH, 3],
name='art_patch_placeholder')
pv_patch_placeholder = tf.placeholder(
tf.float32,
[None, config.PATCH_IMAGE_HEIGHT, config.PATCH_IMAGE_WIDTH, 3],
name='pv_patch_placeholder')
batch_label_placeholder = tf.placeholder(tf.int32, [None, 1],
name='batch_label_input')
batch_attrs_placeholder = tf.placeholder(tf.float32, [None, 15],
name='batch_attrs_input')
batch_size_placeholder = tf.placeholder(tf.int32, [], name='batch_size')
# label_placeholder = tf.placeholder(tf.int32, [None], name='label_placeholder')
net = networks_with_attrs(nc_roi_placeholder,
art_roi_placeholder,
pv_roi_placeholder,
nc_patch_placeholder,
art_patch_placeholder,
pv_patch_placeholder,
batch_attrs_placeholder,
netname,
is_training=False,
num_classes=config.num_classes,
batch_size=batch_size_placeholder,
use_attribute_flag=using_attribute_flag,
clstm_flag=using_clstm_flag)
logits = net.logits
ce_loss, center_loss, gb_ce, lb_ce = net.build_loss(
batch_label_placeholder, add_to_collection=False)
predictions = []
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
roi_features = []
with tf.Session(config=gpu_config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# model_path = '/media/dl-box/HDD3/ld/PycharmProjects/GL_BD_LSTM/logs/model.ckpt-150809'
saver = tf.train.Saver()
print('restore from ', model_path)
saver.restore(sess, model_path)
batch_count = 0
print(dataset)
slice_names = os.listdir(dataset)
print(slice_names)
for idx, slice_name in enumerate(slice_names):
if slice_name.startswith('.DS'):
continue
# if not slice_name.endswith('0'):
# continue
print(slice_name, idx, ' / ', len(slice_names))
cur_data_dir = os.path.join(dataset, slice_name)
logits_values = generate_roi_feature_with_attributions(
dataset, slice_name, config.patch_size, sess, logits,
nc_roi_placeholder, art_roi_placeholder, pv_roi_placeholder,
nc_patch_placeholder, art_patch_placeholder,
pv_patch_placeholder, batch_attrs_placeholder,
batch_size_placeholder)
roi_feature = np.asarray([0., 0., 0., 0., 0.], np.float32)
patch_num = len(logits_values) * 1.0
for value in np.unique(logits_values):
roi_feature[value] = np.sum(logits_values == value) * 1.0
roi_feature /= patch_num
roi_features.append(roi_feature)
print(slice_name[-1], ' logits_values is ', logits_values)
print('roi_feature is ', roi_feature)
roi_features = np.asarray(roi_features, np.float32)
np.save(feature_save_path, roi_features)
def generate_roi_feature_dataset(cur_slice_dir,
netname,
model_path,
using_attribute_flag=True,
using_clstm_flag=True,
global_branch_flag=True,
local_branch_flag=True,
patch_size=5):
'''
对dataset下面的每一个slice生成测试的结果
:param dataset: dataset的路径
:return:
'''
nc_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='nc_roi_placeholder')
art_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='art_roi_placeholder')
pv_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='pv_roi_placeholder')
nc_patch_placeholder = tf.placeholder(
tf.float32,
[None, config.PATCH_IMAGE_HEIGHT, config.PATCH_IMAGE_WIDTH, 3],
name='nc_patch_placeholder')
art_patch_placeholder = tf.placeholder(
tf.float32,
[None, config.PATCH_IMAGE_HEIGHT, config.PATCH_IMAGE_WIDTH, 3],
name='art_patch_placeholder')
pv_patch_placeholder = tf.placeholder(
tf.float32,
[None, config.PATCH_IMAGE_HEIGHT, config.PATCH_IMAGE_WIDTH, 3],
name='pv_patch_placeholder')
batch_label_placeholder = tf.placeholder(tf.int32, [None, 1],
name='batch_label_input')
batch_attrs_placeholder = tf.placeholder(tf.float32, [None, 15],
name='batch_attrs_input')
batch_size_placeholder = tf.placeholder(tf.int32, [], name='batch_size')
# label_placeholder = tf.placeholder(tf.int32, [None], name='label_placeholder')
net = networks_with_attrs(nc_roi_placeholder,
art_roi_placeholder,
pv_roi_placeholder,
nc_patch_placeholder,
art_patch_placeholder,
pv_patch_placeholder,
batch_attrs_placeholder,
netname,
is_training=False,
num_classes=config.num_classes,
batch_size=batch_size_placeholder,
use_attribute_flag=using_attribute_flag,
clstm_flag=using_clstm_flag,
global_branch_flag=global_branch_flag,
local_branch_flag=local_branch_flag)
logits = net.logits
ce_loss, center_loss, gb_ce, lb_ce = net.build_loss(
batch_label_placeholder, add_to_collection=False)
predictions = []
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
roi_features = []
with tf.Session(config=gpu_config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# model_path = '/media/dl-box/HDD3/ld/PycharmProjects/GL_BD_LSTM/logs/model.ckpt-150809'
saver = tf.train.Saver()
print('restore from ', model_path)
saver.restore(sess, model_path)
batch_count = 0
label_map = generate_roi_feature_with_attributions(
os.path.dirname(cur_slice_dir), os.path.basename(cur_slice_dir),
patch_size, sess, logits, nc_roi_placeholder, art_roi_placeholder,
pv_roi_placeholder, nc_patch_placeholder, art_patch_placeholder,
pv_patch_placeholder, batch_attrs_placeholder,
batch_size_placeholder)
print(np.shape(label_map))
cv2.imwrite('./' + os.path.basename(cur_slice_dir) + '.PNG',
np.asarray(label_map, np.uint8))
def evulate_imgs_batch_with_attributions(nc_rois_paths, art_rois_paths,
pv_rois_paths, nc_patches_paths,
art_patches_paths, pv_patches_paths,
attrs, labels, netname, model_path):
batch_dataset = Generate_Batch_Data_with_attributions(
nc_rois_paths, art_rois_paths, pv_rois_paths, nc_patches_paths,
art_patches_paths, pv_patches_paths, attrs, config.val_batch_size,
labels)
generator = batch_dataset.generate_next_batch()
nc_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='nc_roi_placeholder')
art_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='art_roi_placeholder')
pv_roi_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='pv_roi_placeholder')
nc_patch_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='nc_patch_placeholder')
art_patch_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='art_patch_placeholder')
pv_patch_placeholder = tf.placeholder(
tf.float32, [None, config.ROI_IMAGE_HEIGHT, config.ROI_IMAGE_WIDTH, 3],
name='pv_patch_placeholder')
batch_label_placeholder = tf.placeholder(tf.int32, [None, 1],
name='batch_label_input')
batch_size_placeholder = tf.placeholder(tf.int32, [], name='batch_size')
attrs_placeholder = tf.placeholder(tf.float32, [None, 15],
'attributions_placeholder')
# label_placeholder = tf.placeholder(tf.int32, [None], name='label_placeholder')
net = networks_with_attrs(nc_roi_placeholder,
art_roi_placeholder,
pv_roi_placeholder,
nc_patch_placeholder,
art_patch_placeholder,
pv_patch_placeholder,
attrs_placeholder,
base_name=netname,
is_training=True,
num_classes=config.num_classes,
batch_size=batch_size_placeholder)
logits = net.logits
ce_loss, center_loss, gb_ce, lb_ce = net.build_loss(
batch_label_placeholder, add_to_collection=False)
predictions = []
gts = labels
gpu_config = tf.ConfigProto()
gpu_config.gpu_options.allow_growth = True
with tf.Session(config=gpu_config) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
# model_path = '/media/dl-box/HDD3/ld/PycharmProjects/GL_BD_LSTM/logs/model.ckpt-150809'
saver = tf.train.Saver()
print('restore from ', model_path)
saver.restore(sess, model_path)
batch_count = 0
ce_loss_values = []
center_loss_values = []
while True:
nc_batch_rois, art_batch_rois, pv_batch_rois, \
nc_batch_patches, art_batch_patches, pv_batch_patches, \
batch_attrs, batch_label, epoch_end = generator.next()
# print('batch_label is ', batch_label, np.shape(batch_label))
logits_v, ce_loss_v = sess.run(
[logits, ce_loss],
feed_dict={
nc_roi_placeholder: nc_batch_rois,
art_roi_placeholder: art_batch_rois,
pv_roi_placeholder: pv_batch_rois,
nc_patch_placeholder: nc_batch_patches,
art_patch_placeholder: art_batch_patches,
pv_patch_placeholder: pv_batch_patches,
batch_size_placeholder: len(pv_batch_patches),
attrs_placeholder: batch_attrs,
batch_label_placeholder: np.expand_dims(batch_label,
axis=1)
})
ce_loss_values.append(ce_loss_v)
print(np.mean(ce_loss_values))
predictions.extend(np.argmax(logits_v, axis=1))
# print(np.argmax(logits_v, axis=1), batch_label)
if batch_count % 100 == 0 and batch_count != 0:
print(
'%d/%d\n' %
(batch_count * config.val_batch_size, len(nc_rois_paths)))
batch_count += 1
if epoch_end:
break
# with open('./evulation.txt', 'w') as save_txt_file:
# lines = []
# for gt, prediction in zip(gts, predictions):
# line = '%d %d\n' % (prediction, gt)
# lines.append(line)
# save_txt_file.writelines(lines)
# save_txt_file.close()
# calculate the accuracy
gts = np.asarray(gts, np.uint8)
predictions = np.asarray(predictions, np.uint8)
total_acc = np.sum(gts == predictions) / (1.0 * len(gts))
print('the total acc is ', total_acc)
for class_id in range(5):
idx = np.where(gts == class_id)
cur_predictions = predictions[idx]
cur_gts = gts[idx]
cur_acc = np.sum(cur_predictions == cur_gts) / (1.0 * len(cur_gts))
print('the %d\'s acc is %.4f' % (class_id, cur_acc))
def create_clones_with_attrs(train_batch_queue, val_batch_queue):
with tf.device('/cpu:0'):
global_step = slim.create_global_step()
learning_rate = tf.constant(FLAGS.learning_rate, name='learning_rate')
optimizer = tf.train.MomentumOptimizer(learning_rate,
momentum=FLAGS.momentum, name='Momentum')
tf.summary.scalar('learning_rate', learning_rate)
# place clones
pixel_link_loss = 0 # for summary only
gradients = []
# for clone_idx, gpu in enumerate(config.gpus):
# do_summary = clone_idx == 0 # only summary on the first clone
# reuse = clone_idx > 0
with tf.variable_scope(tf.get_variable_scope()) as sc:
b_nc_roi, b_art_roi, b_pv_roi, b_nc_patch, b_art_patch, b_pv_patch, b_label, b_attrs = train_batch_queue.dequeue()
val_b_nc_roi, val_b_art_roi, val_b_pv_roi, val_b_nc_patch, val_b_art_patch, val_b_pv_patch, \
val_b_label, val_b_attrs = val_batch_queue.dequeue()
# build model and loss
train_net = networks_with_attrs(b_nc_roi, b_art_roi, b_pv_roi, b_nc_patch, b_art_patch, b_pv_patch, b_attrs,
FLAGS.netname, True, num_classes=config.num_classes, batch_size=FLAGS.batch_size,
use_attribute_flag=FLAGS.attribute_flag, clstm_flag=FLAGS.clstm_flag)
# ce_loss, center_loss, global_loss, local_loss = net.build_loss(b_label)
centerloss_lambda = 1.0
if not FLAGS.centerloss_flag:
print('do not use the center loss')
centerloss_lambda = 0.0
ce_loss, center_loss, global_loss, local_loss, center_update_op = train_net.build_loss(b_label,
lambda_center_loss=centerloss_lambda)
sc.reuse_variables()
val_net = networks_with_attrs(val_b_nc_roi, val_b_art_roi, val_b_pv_roi, val_b_nc_patch, val_b_art_patch,
val_b_pv_patch, val_b_attrs, FLAGS.netname, False, config.num_classes,
batch_size=FLAGS.batch_size, use_attribute_flag=FLAGS.attribute_flag,
clstm_flag=FLAGS.clstm_flag)
val_ce_loss, val_center_loss, val_global_loss, val_local_loss = val_net.build_loss(val_b_label,
lambda_center_loss=centerloss_lambda,
add_to_collection=False)
losses = tf.get_collection(tf.GraphKeys.LOSSES)
assert len(losses) == 4
total_clone_loss = tf.add_n(losses)
pixel_link_loss += total_clone_loss
regularization_loss = tf.add_n(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
total_clone_loss = total_clone_loss + regularization_loss
clone_gradients = optimizer.compute_gradients(total_clone_loss)
gradients.append(clone_gradients)
tf.summary.scalar('final cross entropy', ce_loss)
tf.summary.scalar('center loss', center_loss)
tf.summary.scalar('regularization_loss', regularization_loss)
tf.summary.scalar('global cross entropy', global_loss)
tf.summary.scalar('local cross entropy', local_loss)
tf.summary.scalar('val/final cross_entropy', val_ce_loss)
tf.summary.scalar('val/center loss', val_center_loss)
tf.summary.scalar('val/global cross entropy', val_global_loss)
tf.summary.scalar('val/local cross entropy', val_local_loss)
# add all gradients together
# note that the gradients do not need to be averaged, because the average operation has been done on loss.
averaged_gradients = sum_gradients(gradients)
apply_grad_op = optimizer.apply_gradients(averaged_gradients, global_step=global_step)
train_ops = [apply_grad_op, center_update_op]
bn_update_op = util.tf.get_update_op()
if bn_update_op is not None:
train_ops.append(bn_update_op)
# moving average
if FLAGS.using_moving_average:
tf.logging.info('using moving average in training, \
with decay = %f' % (FLAGS.moving_average_decay))
ema = tf.train.ExponentialMovingAverage(FLAGS.moving_average_decay)
ema_op = ema.apply(tf.trainable_variables())
with tf.control_dependencies([apply_grad_op]): # ema after updating
train_ops.append(tf.group(ema_op))
train_op = control_flow_ops.with_dependencies(train_ops, pixel_link_loss, name='train_op')
train_step_kwargs = {}
train_step_kwargs['val_loss'] = val_ce_loss
train_step_kwargs['val_gb_loss'] = val_global_loss
train_step_kwargs['val_lb_loss'] = val_local_loss
return train_op, None
