def build_loss(self):
response = self.response
response_size = response.get_shape().as_list()[1:3] # [height, width]
gt = _construct_gt_response(response_size,
self.config['batch_size'],
self.stride,
self.config['gt_config'])
self.gt=gt
with tf.name_scope('Loss'):
loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=response,
labels=gt)
with tf.name_scope('Balance_weights'):
n_pos = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 1)))
n_neg = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 0)))
w_pos = 0.5 / n_pos
w_neg = 0.5 / n_neg
class_weights = tf.where(tf.equal(gt, 1),
w_pos * tf.ones_like(gt),
tf.ones_like(gt))
class_weights = tf.where(tf.equal(gt, 0),
w_neg * tf.ones_like(gt),
class_weights)
loss = loss * class_weights
# Note that we use reduce_sum instead of reduce_mean since the loss has
# already been normalized by class_weights in spatial dimension.
loss = tf.reduce_sum(loss, [1, 2])
batch_loss = tf.reduce_mean(loss, name='batch_loss')
tf.losses.add_loss(batch_loss)
total_loss = tf.losses.get_total_loss()
self.batch_loss = batch_loss
self.total_loss = total_loss
tf.summary.image(self.mode+'exemplar', self.examplars)
tf.summary.image(self.mode+'instance', self.instances)
mean_batch_loss, update_op1 = tf.metrics.mean(batch_loss)
mean_total_loss, update_op2 = tf.metrics.mean(total_loss)
with tf.control_dependencies([update_op1, update_op2]):
tf.summary.scalar(self.mode+'batch_loss', mean_batch_loss)
tf.summary.scalar(self.mode+'total_loss', mean_total_loss)
if self.mode == 'train':
with tf.name_scope("GT"):
tf.summary.image('GT', tf.reshape(gt[0], [1] + response_size + [1]))
tf.summary.image(self.mode+'Response', tf.expand_dims(tf.sigmoid(response), -1))
tf.summary.histogram(self.mode+'Response', self.response)
# Two more metrics to monitor the performance of training
tf.summary.scalar(self.mode+'center_score_error', center_score_error(response))
tf.summary.scalar(self.mode+'center_dist_error', center_dist_error(response))
def build_loss(self):
response = self.response
response_size = response.get_shape().as_list()[1:3] # [height, width]
gt = construct_gt_score_maps(response_size,
self.data_config['batch_size'],
self.model_config['embed_config']['stride'],
self.train_config['gt_config'])
with tf.name_scope('Loss'):
loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=response,
labels=gt)
with tf.name_scope('Balance_weights'):
n_pos = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 1)))
n_neg = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 0)))
w_pos = 0.5 / n_pos
w_neg = 0.5 / n_neg
class_weights = tf.where(tf.equal(gt, 1),
w_pos * tf.ones_like(gt),
tf.ones_like(gt))
class_weights = tf.where(tf.equal(gt, 0),
w_neg * tf.ones_like(gt),
class_weights)
loss = loss * class_weights
# Note that we use reduce_sum instead of reduce_mean since the loss has
# already been normalized by class_weights in spatial dimension.
loss = tf.reduce_sum(loss, [1, 2])
batch_loss = tf.reduce_mean(loss, name='batch_loss')
tf.losses.add_loss(batch_loss)
total_loss = tf.losses.get_total_loss()
self.batch_loss = batch_loss
self.total_loss = total_loss
tf.summary.image('exemplar', self.exemplars, family=self.mode)
tf.summary.image('instance', self.instances, family=self.mode)
mean_batch_loss, update_op1 = tf.metrics.mean(batch_loss)
mean_total_loss, update_op2 = tf.metrics.mean(total_loss)
with tf.control_dependencies([update_op1, update_op2]):
tf.summary.scalar('batch_loss', mean_batch_loss, family=self.mode)
tf.summary.scalar('total_loss', mean_total_loss, family=self.mode)
if self.mode == 'train':
tf.summary.image('GT', tf.reshape(gt[0], [1] + response_size + [1]), family='GT')
tf.summary.image('Response', tf.expand_dims(tf.sigmoid(response), -1), family=self.mode)
tf.summary.histogram('Response', self.response, family=self.mode)
# Two more metrics to monitor the performance of training
tf.summary.scalar('center_score_error', center_score_error(response), family=self.mode)
tf.summary.scalar('center_dist_error', center_dist_error(response), family=self.mode)
def build_loss(self):
response = self.response
response_size = response.get_shape().as_list()[1:3] # [height, width]
gt = construct_gt_score_maps(
response_size, self.data_config['batch_size'],
self.model_config['embed_config']['stride'],
self.train_config['gt_config'])
# loss: https://www.renom.jp/ja/notebooks/tutorial/basic_algorithm/lossfunction/notebook.html
with tf.name_scope('Loss'):
loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=response,
labels=gt)
with tf.name_scope('Balance_weights'):
n_pos = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 1)))
n_neg = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 0)))
w_pos = 0.5 / n_pos
w_neg = 0.5 / n_neg
class_weights = tf.where(tf.equal(gt, 1),
w_pos * tf.ones_like(gt),
tf.ones_like(gt))
class_weights = tf.where(tf.equal(gt, 0),
w_neg * tf.ones_like(gt),
class_weights)
loss = loss * class_weights
# Note that we use reduce_sum instead of reduce_mean since the loss has
# already been normalized by class_weights in spatial dimension.
loss = tf.reduce_sum(loss, [1, 2])
batch_loss = tf.reduce_mean(loss, name='batch_loss')
tf.losses.add_loss(batch_loss)
total_loss = tf.losses.get_total_loss()
self.batch_loss = batch_loss
self.total_loss = total_loss
# quantization
# good note: https://www.tensorflowers.cn/t/7136
if self.model_config['embed_config']['quantization']:
if self.train_config["export"]:
contrib_quantize.create_eval_graph()
else:
contrib_quantize.create_training_graph(quant_delay=200000)
tf.summary.image('exemplar', self.exemplars, family=self.mode)
tf.summary.image('instance', self.instances, family=self.mode)
mean_batch_loss, update_op1 = tf.metrics.mean(batch_loss)
mean_total_loss, update_op2 = tf.metrics.mean(total_loss)
with tf.control_dependencies([update_op1, update_op2]):
tf.summary.scalar('batch_loss',
mean_batch_loss,
family=self.mode)
tf.summary.scalar('total_loss',
mean_total_loss,
family=self.mode)
if self.mode == 'train':
tf.summary.image('GT',
tf.reshape(gt[0], [1] + response_size + [1]),
family='GT')
tf.summary.image('Response',
tf.expand_dims(tf.sigmoid(response), -1),
family=self.mode)
tf.summary.histogram('Response', self.response, family=self.mode)
# Two more metrics to monitor the performance of training
tf.summary.scalar('center_score_error',
center_score_error(response),
family=self.mode)
tf.summary.scalar('center_dist_error',
center_dist_error(response),
family=self.mode)
def build_loss(self):
response = self.response
response_size = response.get_shape().as_list()[1:3]
gt = construct_gt_score_maps(
response_size, self.data_config['batch_size'] * 2,
self.model_config['embed_config']['stride'],
self.train_config['gt_config'])
with tf.name_scope('Loss'):
loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=response,
labels=gt)
with tf.name_scope('Balance_weights'):
n_pos = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 1)))
n_neg = tf.reduce_sum(tf.to_float(tf.equal(gt[0], 0)))
w_pos = 0.5 / n_pos
w_neg = 0.5 / n_neg
class_weights = tf.where(tf.equal(gt, 1),
w_pos * tf.ones_like(gt),
tf.ones_like(gt))
class_weights = tf.where(tf.equal(gt, 0),
w_neg * tf.ones_like(gt),
class_weights)
loss = loss * class_weights
loss = tf.reduce_sum(loss, [1, 2])
batch_loss = tf.reduce_mean(loss, name='batch_loss')
tf.losses.add_loss(batch_loss)
# Get total loss
total_loss = tf.losses.get_total_loss()
self.batch_loss = batch_loss
self.total_loss = total_loss
tf.summary.image('exemplar', self.exemplars_show, family=self.mode)
tf.summary.image('instance', self.instances, family=self.mode)
mean_total_loss, update_op2 = tf.metrics.mean(total_loss)
with tf.control_dependencies([update_op2]):
tf.summary.scalar('total_loss',
mean_total_loss,
family=self.mode)
if self.mode == 'train':
tf.summary.image('GT',
tf.reshape(gt[0], [1] + response_size + [1]),
family='GT')
tf.summary.image('Response',
tf.expand_dims(tf.sigmoid(self.total_response),
-1),
family=self.mode)
tf.summary.histogram('Response',
self.total_response,
family=self.mode)
# Two more metrics to monitor the performance of training
tf.summary.scalar('center_score_error',
center_score_error(self.total_response),
family=self.mode)
tf.summary.scalar('center_dist_error',
center_dist_error(self.total_response),
family=self.mode)
