def testBatchNormScopeDoesHasIsTrainingWhenItsNotNone(self):
sc = mobilenet.training_scope(is_training=False)
self.assertIn('is_training',
sc[slim.arg_scope_func_key(slim.batch_norm)])
sc = mobilenet.training_scope(is_training=True)
self.assertIn('is_training',
sc[slim.arg_scope_func_key(slim.batch_norm)])
sc = mobilenet.training_scope()
self.assertIn('is_training',
sc[slim.arg_scope_func_key(slim.batch_norm)])
def training_scope(**kwargs):
"""Defines MobilenetV2 training scope.
Usage:
with tf.contrib.slim.arg_scope(mobilenet_v2.training_scope()):
logits, endpoints = mobilenet_v2.mobilenet(input_tensor)
with slim.
Args:
**kwargs: Passed to mobilenet.training_scope. The following parameters
are supported:
weight_decay- The weight decay to use for regularizing the model.
stddev- Standard deviation for initialization, if negative uses xavier.
dropout_keep_prob- dropout keep probability
bn_decay- decay for the batch norm moving averages.
Returns:
An `arg_scope` to use for the mobilenet v2 model.
"""
return lib.training_scope(**kwargs)
def model_fn(self, is_training=True, *args, **kwargs):
x_input = None
y_input = None
original_image = None
if len(args) > 0:
if is_training:
x_input, y_input = args[0].dequeue()
else:
x_input, original_image = args[0]
x_input = tf.identity(x_input, 'input_node')
else:
x_input = tf.placeholder(
tf.float32,
shape=[1, ctx.params.output_size, ctx.params.output_size, 3],
name='input_node')
with slim.arg_scope(mobilenet.training_scope(is_training=is_training)):
_, end_points = self._mobilenet_v2(x_input,
depth_multiplier=1.0,
output_stride=16)
last_layer = end_points['layer_18/output'] # _ x 8 x 8 x _
middle_layer = end_points[
'layer_4/depthwise_output'] # _ x 32 x 32 x _
with tf.variable_scope(None,
'DECODER', [last_layer, middle_layer],
reuse=None):
logits = self.decoder(last_layer, middle_layer, is_training,
[2, 4, 8], 0.00004)
if is_training:
# resize to target size
logits = tf.image.resize_bilinear(
logits, [ctx.params.origin_height, ctx.params.origin_width])
# for labels
one_hot_labels = slim.one_hot_encoding(y_input,
ctx.params.num_classes)
cross_entroy = tf.nn.softmax_cross_entropy_with_logits(
labels=one_hot_labels, logits=logits)
softmax_logits = tf.nn.softmax(logits, dim=-1)
_, logits_1 = tf.split(softmax_logits,
num_or_size_splits=2,
axis=3)
logits_1 = tf.reshape(logits_1,
shape=[
ctx.params.batch_size,
ctx.params.origin_height,
ctx.params.origin_width
])
label_batch_float = tf.cast(y_input, tf.float32)
weights_pt = tf.pow(
1 - (logits_1 * label_batch_float + (1 - logits_1) *
(1 - label_batch_float)), 2)
weights_total = (logits_1 * 5.02 +
(1 - logits_1) * 4.98) * weights_pt
focal_loss = weights_total * cross_entroy
loss = tf.reduce_mean(focal_loss)
tf.losses.add_loss(loss)
return loss
else:
logits = tf.nn.softmax(logits)
logits = tf.identity(logits, 'output_node')
_, positive_s = tf.split(logits, 2, 3)
return positive_s, original_image
def testBatchNormScopeDoesNotHaveIsTrainingWhenItsSetToNone(self):
sc = mobilenet.training_scope(is_training=None)
self.assertNotIn('is_training',
sc[slim.arg_scope_func_key(slim.batch_norm)])
