def test_mobilenet_v3_like_search_gradients(self):
model_spec = mobile_search_space_v3.mobilenet_v3_like_search()
model_spec = schema.map_oneofs(_with_largest_possible_masks, model_spec)
model = mobile_model_v3.get_model(
model_spec, num_classes=1001, force_stateless_batch_norm=True)
inputs = tf.random_normal(shape=[8, 224, 224, 3], dtype=tf.float32)
model.build(inputs.shape)
logits, unused_endpoints = model.apply(inputs, training=True)
labels = tf.one_hot([0]*8, 1001, dtype=tf.float32)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
variables = model.trainable_variables()
grads = tf.gradients(loss, variables)
for variable, grad in zip(variables, grads):
self.assertIsNotNone(
grad, msg='Gradient for {} is None'.format(variable.name))
self.evaluate(tf.global_variables_initializer())
for variable, array in zip(variables, self.evaluate(grads)):
self.assertFalse(
np.all(np.equal(array, 0)),
msg='Gradient for {} is identically zero'.format(variable.name))
def test_search_space_construction(self, ssd):
model_spec = mobile_search_space_v3.get_search_space_spec(ssd)
model_spec = test_utils.with_random_masks(model_spec)
model = mobile_model_v3.get_model(
model_spec, num_classes=1001, force_stateless_batch_norm=True)
inputs = tf.random_normal(shape=[128, 224, 224, 3])
model.build(inputs.shape)
output, unused_endpoints = model.apply(inputs, training=True)
self.assertEqual(output.shape, [128, 1001])
def test_dynamic_input_size(self, ssd):
model_spec = mobile_search_space_v3.get_search_space_spec(ssd)
model_spec = test_utils.with_random_pruning(model_spec)
model = mobile_model_v3.get_model(
model_spec, num_classes=1001, force_stateless_batch_norm=False)
# Input height/width are not known at graph construction time.
inputs = tf.placeholder(shape=[128, None, None, 3], dtype=tf.float32)
model.build(inputs.shape)
output, unused_endpoints = model.apply(inputs, training=True)
self.assertEqual(output.shape, [128, 1001])
def test_output_shapes(self):
model_spec = mobile_search_space_v3.mobilenet_v3_large()
model = mobile_model_v3.get_model(model_spec, num_classes=50)
features = tf.ones([8, 224, 224, 3])
model.build(features.shape)
logits, endpoints = model.apply(features, training=True)
self.assertEqual(logits.shape, tf.TensorShape([8, 50]))
self.assertEqual([x.shape for x in endpoints], [
[8, 112, 112, 16],
[8, 56, 56, 24],
[8, 28, 28, 40],
[8, 14, 14, 112],
[8, 7, 7, 160],
])
def test_output_shapes_with_variable_kernel_sizes(self):
model_spec = mobile_search_space_v3.mobilenet_v3_like_search()
model_spec = test_utils.with_random_masks(model_spec)
model = mobile_model_v3.get_model(
model_spec, num_classes=50, force_stateless_batch_norm=True)
features = tf.ones([8, 224, 224, 3])
model.build(features.shape)
logits, endpoints = model.apply(features, training=True)
self.assertEqual(logits.shape, tf.TensorShape([8, 50]))
self.assertLen(endpoints, 5)
endpoints[0].shape.assert_is_compatible_with([8, 112, 112, None])
endpoints[1].shape.assert_is_compatible_with([8, 56, 56, None])
endpoints[2].shape.assert_is_compatible_with([8, 28, 28, None])
endpoints[3].shape.assert_is_compatible_with([8, 14, 14, None])
endpoints[4].shape.assert_is_compatible_with([8, 7, 7, None])
def get_standalone_model(model_spec, **kwargs):
"""Create a standalone model based on `model_spec`.
Args:
model_spec: basic_specs.ConvTowerSpec namedtuple.
**kwargs: Additional keyword arguments to pass to the model builder.
Returns:
An instance of rematlib/layers.Layer. The layer's apply() function will
return a tuple `(output, endpoints)`, where `output` is a Tensor and
`endpoints` is a list of Tensors for object detection.
Raises:
ValueError: If `ssd` is not recognized.
"""
return mobile_model_v3.get_model(model_spec=model_spec,
force_stateless_batch_norm=False,
**kwargs)
def get_model_for_search(model_spec, **kwargs):
"""Create a one-shot/weight-shared model based on `model_spec`.
NOTE: We will always force the selected algorithm to run with stateless batch
normalization. This makes the method suitable for architecture searches but
not stand-alone model training.
Args:
model_spec: basic_specs.ConvTowerSpec namedtuple.
**kwargs: Additional keyword arguments to pass to the model builder.
Returns:
An instance of rematlib/layers.Layer. The layer's apply() function will
return a tuple `(output, endpoints)`, where `output` is a Tensor and
`endpoints` is a list of Tensors for object detection.
"""
return mobile_model_v3.get_model(model_spec=model_spec,
force_stateless_batch_norm=True,
**kwargs)
def test_output_shapes_with_variable_filter_sizes(self):
filter_multipliers = (0.5, 1.0, 2.0)
model_spec = mobile_search_space_v3.mobilenet_v3_large()
model_spec = search_space_utils.scale_conv_tower_spec(
model_spec, multipliers=filter_multipliers)
model_spec = test_utils.with_random_masks(model_spec)
model = mobile_model_v3.get_model(model_spec, num_classes=50)
features = tf.ones([8, 224, 224, 3])
model.build(features.shape)
logits, endpoints = model.apply(features, training=True)
self.assertEqual(logits.shape, tf.TensorShape([8, 50]))
self.assertEqual([x.shape for x in endpoints], [
[8, 112, 112, int(16 * max(filter_multipliers))],
[8, 56, 56, int(24 * max(filter_multipliers))],
[8, 28, 28, int(40 * max(filter_multipliers))],
[8, 14, 14, int(112 * max(filter_multipliers))],
[8, 7, 7, int(160 * max(filter_multipliers))],
])
