def testModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
with slim.arg_scope(inception.inception_v2_arg_scope()):
inception.inception_v2_base(inputs)
total_params, _ = slim.model_analyzer.analyze_vars(slim.get_model_variables())
self.assertAlmostEqual(10173112, total_params)
def testModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
with slim.arg_scope(inception.inception_v2_arg_scope()):
inception.inception_v2_base(inputs)
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables())
self.assertAlmostEqual(10173112, total_params)
def testBuildAndCheckAllEndPointsUptoMixed5c(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
_, end_points = inception.inception_v2_base(inputs,
final_endpoint='Mixed_5c')
endpoints_shapes = {
'Mixed_3b': [batch_size, 28, 28, 256],
'Mixed_3c': [batch_size, 28, 28, 320],
'Mixed_4a': [batch_size, 14, 14, 576],
'Mixed_4b': [batch_size, 14, 14, 576],
'Mixed_4c': [batch_size, 14, 14, 576],
'Mixed_4d': [batch_size, 14, 14, 576],
'Mixed_4e': [batch_size, 14, 14, 576],
'Mixed_5a': [batch_size, 7, 7, 1024],
'Mixed_5b': [batch_size, 7, 7, 1024],
'Mixed_5c': [batch_size, 7, 7, 1024],
'Conv2d_1a_7x7': [batch_size, 112, 112, 64],
'MaxPool_2a_3x3': [batch_size, 56, 56, 64],
'Conv2d_2b_1x1': [batch_size, 56, 56, 64],
'Conv2d_2c_3x3': [batch_size, 56, 56, 192],
'MaxPool_3a_3x3': [batch_size, 28, 28, 192]
}
self.assertItemsEqual(endpoints_shapes.keys(), end_points.keys())
for endpoint_name in endpoints_shapes:
expected_shape = endpoints_shapes[endpoint_name]
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(
end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def build_model(model_name, inputs, num_classes, is_training, dropout_keep_prob):
use_fcn = False
if model_name.find('fcn') >= 0:
use_fcn = True
model_base_name = model_name[0:-4]
else:
model_base_name = model_name
if model_base_name == 'vgg16':
net = vgg16_base(inputs)
elif model_base_name == 'inception_v1':
with slim.arg_scope(inception.inception_v1_arg_scope()):
net, _ = inception.inception_v1_base(inputs)
elif model_base_name == 'inception_v2':
with slim.arg_scope(inception.inception_v2_arg_scope()):
net, _ = inception.inception_v2_base(inputs)
elif model_base_name == 'inception_v3':
with slim.arg_scope(inception.inception_v3_arg_scope()):
net, _ = inception.inception_v3_base(inputs)
else:
raise Exception('model {} is not existed'.format(model_name))
with tf.variable_scope('not_pretrained'):
if use_fcn:
net = fully_convolutional_networks(net, num_classes, is_training, dropout_keep_prob)
else:
net = fully_connected_networks(net, num_classes, is_training, dropout_keep_prob)
return net
def testBuildAndCheckAllEndPointsUptoMixed5c(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
_, end_points = inception.inception_v2_base(inputs,
final_endpoint='Mixed_5c')
endpoints_shapes = {'Mixed_3b': [batch_size, 28, 28, 256],
'Mixed_3c': [batch_size, 28, 28, 320],
'Mixed_4a': [batch_size, 14, 14, 576],
'Mixed_4b': [batch_size, 14, 14, 576],
'Mixed_4c': [batch_size, 14, 14, 576],
'Mixed_4d': [batch_size, 14, 14, 576],
'Mixed_4e': [batch_size, 14, 14, 576],
'Mixed_5a': [batch_size, 7, 7, 1024],
'Mixed_5b': [batch_size, 7, 7, 1024],
'Mixed_5c': [batch_size, 7, 7, 1024],
'Conv2d_1a_7x7': [batch_size, 112, 112, 64],
'MaxPool_2a_3x3': [batch_size, 56, 56, 64],
'Conv2d_2b_1x1': [batch_size, 56, 56, 64],
'Conv2d_2c_3x3': [batch_size, 56, 56, 192],
'MaxPool_3a_3x3': [batch_size, 28, 28, 192]}
self.assertItemsEqual(endpoints_shapes.keys(), end_points.keys())
for endpoint_name in endpoints_shapes:
expected_shape = endpoints_shapes[endpoint_name]
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def testBuildOnlyUptoFinalEndpoint(self):
batch_size = 5
height, width = 224, 224
endpoints = [
"Conv2d_1a_7x7",
"MaxPool_2a_3x3",
"Conv2d_2b_1x1",
"Conv2d_2c_3x3",
"MaxPool_3a_3x3",
"Mixed_3b",
"Mixed_3c",
"Mixed_4a",
"Mixed_4b",
"Mixed_4c",
"Mixed_4d",
"Mixed_4e",
"Mixed_5a",
"Mixed_5b",
"Mixed_5c",
]
for index, endpoint in enumerate(endpoints):
with tf.Graph().as_default():
inputs = tf.random_uniform((batch_size, height, width, 3))
out_tensor, end_points = inception.inception_v2_base(inputs, final_endpoint=endpoint)
self.assertTrue(out_tensor.op.name.startswith("InceptionV2/" + endpoint))
self.assertItemsEqual(endpoints[: index + 1], end_points)
def testBuildBaseNetwork(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
mixed_5c, end_points = inception.inception_v2_base(inputs)
self.assertTrue(mixed_5c.op.name.startswith("InceptionV2/Mixed_5c"))
self.assertListEqual(mixed_5c.get_shape().as_list(), [batch_size, 7, 7, 1024])
expected_endpoints = [
"Mixed_3b",
"Mixed_3c",
"Mixed_4a",
"Mixed_4b",
"Mixed_4c",
"Mixed_4d",
"Mixed_4e",
"Mixed_5a",
"Mixed_5b",
"Mixed_5c",
"Conv2d_1a_7x7",
"MaxPool_2a_3x3",
"Conv2d_2b_1x1",
"Conv2d_2c_3x3",
"MaxPool_3a_3x3",
]
self.assertItemsEqual(end_points.keys(), expected_endpoints)
def build(self,
inputs,
input_pixel_size,
is_training,
scope='img_inception'):
"""Inception for BEV feature extraction
Args:
inputs: a tensor of size [batch_size, height, width, channels].
input_pixel_size: size of the input (H x W)
is_training: True for training, False fo validation/testing.
scope: Optional scope for the variables.
Returns:
The net, a rank-4 tensor of size [batch, height_out, width_out,
channels_out] and end_points dict.
"""
inception_config = self.config
with tf.variable_scope(scope, 'img_inception', [inputs]) as scope:
with slim.arg_scope([slim.batch_norm, slim.dropout],
is_training=is_training):
if inception_config.inception_v == 'inception_v1':
with slim.arg_scope(inception.inception_v1_arg_scope()):
net, end_points = inception.inception_v1_base(
inputs, scope=scope)
elif inception_config.inception_v == 'inception_v2':
with slim.arg_scope(inception.inception_v2_arg_scope()):
net, end_points = inception.inception_v2_base(
inputs, scope=scope)
elif inception_config.inception_v == 'inception_v3':
with slim.arg_scope(inception.inception_v3_arg_scope()):
net, end_points = inception.inception_v3_base(
inputs, scope=scope)
else:
raise ValueError('Invalid Inception version {},'.format(
inception_config.inception_v))
with tf.variable_scope('upsampling'):
# This feature extractor downsamples the input by a factor
# of 32
downsampling_factor = 32
downsampled_shape = input_pixel_size / downsampling_factor
upsampled_shape = downsampled_shape * \
inception_config.upsampling_multiplier
feature_maps_out = tf.image.resize_bilinear(
net, upsampled_shape)
return feature_maps_out, end_points
def testBuildOnlyUptoFinalEndpoint(self):
batch_size = 5
height, width = 224, 224
endpoints = ['Conv2d_1a_7x7', 'MaxPool_2a_3x3', 'Conv2d_2b_1x1',
'Conv2d_2c_3x3', 'MaxPool_3a_3x3', 'Mixed_3b', 'Mixed_3c',
'Mixed_4a', 'Mixed_4b', 'Mixed_4c', 'Mixed_4d', 'Mixed_4e',
'Mixed_5a', 'Mixed_5b', 'Mixed_5c']
for index, endpoint in enumerate(endpoints):
with tf.Graph().as_default():
inputs = tf.random_uniform((batch_size, height, width, 3))
out_tensor, end_points = inception.inception_v2_base(
inputs, final_endpoint=endpoint)
self.assertTrue(out_tensor.op.name.startswith(
'InceptionV2/' + endpoint))
self.assertItemsEqual(endpoints[:index+1], end_points)
def testBuildBaseNetwork(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
mixed_5c, end_points = inception.inception_v2_base(inputs)
self.assertTrue(mixed_5c.op.name.startswith('InceptionV2/Mixed_5c'))
self.assertListEqual(mixed_5c.get_shape().as_list(),
[batch_size, 7, 7, 1024])
expected_endpoints = ['Mixed_3b', 'Mixed_3c', 'Mixed_4a', 'Mixed_4b',
'Mixed_4c', 'Mixed_4d', 'Mixed_4e', 'Mixed_5a',
'Mixed_5b', 'Mixed_5c', 'Conv2d_1a_7x7',
'MaxPool_2a_3x3', 'Conv2d_2b_1x1', 'Conv2d_2c_3x3',
'MaxPool_3a_3x3']
self.assertItemsEqual(end_points.keys(), expected_endpoints)
def testInceptionV2_TotalCost(self):
conv_params = {
'activation_fn': tf.nn.relu6,
'weights_regularizer': contrib_layers.l2_regularizer(0.00004),
'weights_initializer': tf.random_normal_initializer(stddev=0.03),
'trainable': True,
'biases_initializer': tf.constant_initializer(0.0),
'normalizer_fn': contrib_layers.batch_norm,
'normalizer_params': {
'is_training': False,
'decay': 0.9997,
'scale': True,
'epsilon': 0.001,
}
}
tf.reset_default_graph()
with slim.arg_scope([slim.layers.conv2d, slim.layers.separable_conv2d],
**conv_params):
# Build model.
image = tf.zeros([1, 224, 224, 3])
net, _ = inception.inception_v2_base(image)
logits = slim.layers.fully_connected(
net,
1001,
activation_fn=None,
scope='logits',
weights_initializer=tf.random_normal_initializer(stddev=1e-3),
biases_initializer=tf.constant_initializer(0.0))
# Instantiate regularizers.
flop_reg = flop_regularizer.GammaFlopsRegularizer(
[logits.op], gamma_threshold=0.5)
p100_reg = latency_regularizer.GammaLatencyRegularizer(
[logits.op], gamma_threshold=0.5, hardware='P100')
v100_reg = latency_regularizer.GammaLatencyRegularizer(
[logits.op], gamma_threshold=0.5, hardware='V100')
model_size_reg = model_size_regularizer.GammaModelSizeRegularizer(
[logits.op], gamma_threshold=0.5)
with self.cached_session():
tf.global_variables_initializer().run()
# Verify costs are expected.
self.assertAllClose(3.86972e+09, flop_reg.get_cost())
self.assertAllClose(517536.0, p100_reg.get_cost())
self.assertAllClose(173330.453125, v100_reg.get_cost())
self.assertAllClose(1.11684e+07, model_size_reg.get_cost())
def testInceptionV2(self, hardware):
image = tf.zeros([1, 224, 224, 3])
net, _ = inception.inception_v2_base(image)
g = tf.get_default_graph()
self.regularizer = latency_regularizer.GammaLatencyRegularizer(
[net.op], gamma_threshold=0.5, hardware=hardware)
# Compute-bound convolution.
op = g.get_operation_by_name(
'InceptionV2/Mixed_3c/Branch_2/Conv2d_0c_3x3/Conv2D')
# FLOP cost = 2 * NHWRSCK
expected_cost = (2 * 28 * 28 * 3 * 3 * 96 * 96
/ resource_function.PEAK_COMPUTE[hardware])
self.assertAllClose(expected_cost, self.get_cost([op]))
# Memory-bound convolution.
op = g.get_operation_by_name(
'InceptionV2/Conv2d_1a_7x7/separable_conv2d')
# Memory cost = input_tensor + weight_tensor + output_tensor
# = NHWC + RSCK + NHWK
# Note that this is a pointwise convolution with kernel 1x1.
expected_cost = ((112 * 112 * 24 + 24 * 64 + 112 * 112 * 64) * 4
/ resource_function.MEMORY_BANDWIDTH[hardware])
self.assertAllClose(expected_cost, self.get_cost([op]))