def __call__(self, total_loss, learning_rate):
flags_solver = self.flags
var_list = get_variables_with_name(
name='ocnn', without='seg_header', verbose=flags_solver.verbose)
optim_backbone = Optimizer(var_list=var_list, mul=0.1)
solver1, lr1 = optim_backbone(total_loss, learning_rate)
var_list = get_variables_with_name(
name='seg_header', verbose=flags_solver.verbose)
optim_header = Optimizer(var_list=var_list, mul=1.0)
solver2, lr2 = optim_header(total_loss, learning_rate)
solver = tf.group([solver1, solver2])
return solver, lr2
def __call__(self, total_loss, learning_rate):
flags_solver = self.flags
var_list = get_variables_with_name(name='cls_header',
verbose=flags_solver.verbose)
optim_header = Optimizer(var_list=var_list, mul=1.0)
solver2, lr2 = optim_header(total_loss, learning_rate)
return solver2, lr2
def restore(self, sess, ckpt):
# !!! Restore the trainable/untrainable variables under the name scope `ocnn`
# Note the variables added by solvers are filtered out since they are not
# under the scope of `ocnn`
print('Restore from: ' + ckpt)
var_restore = get_variables_with_name(
'ocnn', without='predict_6/conv2', verbose=self.flags.verbose, train_only=False)
tf_saver = tf.train.Saver(var_list=var_restore)
tf_saver.restore(sess, ckpt)