def testRaiseValueErrorWithInvalidDepthMultiplier(self):
batch_size = 5
height, width = 224, 224
num_classes = 1000
inputs = tf.random_uniform((batch_size, height, width, 3))
with self.assertRaises(ValueError):
_ = shufflenet_v2.shufflenet_v2_base(inputs,
num_classes,
depth_multiplier=-0.1)
with self.assertRaises(ValueError):
_ = shufflenet_v2.shufflenet_v2_base(inputs,
num_classes,
depth_multiplier=0.75)
def test2_0xBasicNetworkModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_num_params = {
# Conv1 & MaxPool
'Conv2d_0': 720,
'MaxPool2d_0': 720,
# Stage 1 with 4 units
'Stage_0/Unit_0': 24310,
'Stage_0/Unit_1': 56274,
'Stage_0/Unit_2': 88238,
'Stage_0/Unit_3': 120202,
# Stage 2 with 8 units
'Stage_1/Unit_0': 306862,
'Stage_1/Unit_1': 430326,
'Stage_1/Unit_2': 553790,
'Stage_1/Unit_3': 677254,
'Stage_1/Unit_4': 800718,
'Stage_1/Unit_5': 924182,
'Stage_1/Unit_6': 1047646,
'Stage_1/Unit_7': 1171110,
# Stage 3 with 4 units
'Stage_2/Unit_0': 1901646,
'Stage_2/Unit_1': 2386718,
'Stage_2/Unit_2': 2871790,
'Stage_2/Unit_3': 3356862,
# Conv2d_1
'Conv2d_1': 5361854,
}
for scope, end_point in enumerate(endpoints_num_params):
with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v2.shufflenet_v2_base(inputs,
final_endpoint=end_point,
depth_multiplier=2.0,
scope=str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=str(scope)))
self.assertAlmostEqual(endpoints_num_params[end_point],
total_params)
def test1_5xBasicNetworkModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_num_params = {
# Conv1 & MaxPool
'Conv2d_0': 720,
'MaxPool2d_0': 720,
# Stage 1 with 4 units
'Stage_0/Unit_0': 14824,
'Stage_0/Unit_1': 31896,
'Stage_0/Unit_2': 48968,
'Stage_0/Unit_3': 66040,
# Stage 2 with 8 units
'Stage_1/Unit_0': 164776,
'Stage_1/Unit_1': 229896,
'Stage_1/Unit_2': 295016,
'Stage_1/Unit_3': 360136,
'Stage_1/Unit_4': 425256,
'Stage_1/Unit_5': 490376,
'Stage_1/Unit_6': 555496,
'Stage_1/Unit_7': 620616,
# Stage 3 with 4 units
'Stage_2/Unit_0': 1003944,
'Stage_2/Unit_1': 1258088,
'Stage_2/Unit_2': 1512232,
'Stage_2/Unit_3': 1766376,
# Conv2d_1
'Conv2d_1': 2490344,
}
for scope, end_point in enumerate(endpoints_num_params):
with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v2.shufflenet_v2_base(inputs,
final_endpoint=end_point,
depth_multiplier=1.5,
scope=str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=str(scope)))
self.assertAlmostEqual(endpoints_num_params[end_point],
total_params)
def test0_5xBasicNetworkModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_num_params = {
# Conv1 & MaxPool
'Conv2d_0': 720,
'MaxPool2d_0': 720,
# Stage 1 with 4 units
'Stage_0/Unit_0': 3240,
'Stage_0/Unit_1': 4824,
'Stage_0/Unit_2': 6408,
'Stage_0/Unit_3': 7992,
# Stage 2 with 8 units
'Stage_1/Unit_0': 16488,
'Stage_1/Unit_1': 21960,
'Stage_1/Unit_2': 27432,
'Stage_1/Unit_3': 32904,
'Stage_1/Unit_4': 38376,
'Stage_1/Unit_5': 43848,
'Stage_1/Unit_6': 49320,
'Stage_1/Unit_7': 54792,
# Stage 3 with 4 units
'Stage_2/Unit_0': 85608,
'Stage_2/Unit_1': 105768,
'Stage_2/Unit_2': 125928,
'Stage_2/Unit_3': 146088,
# Conv2d_1
'Conv2d_1': 345768,
}
for scope, end_point in enumerate(endpoints_num_params):
with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v2.shufflenet_v2_base(inputs,
final_endpoint=end_point,
depth_multiplier=0.5,
scope=str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=str(scope)))
self.assertAlmostEqual(endpoints_num_params[end_point],
total_params)
def testBasicNetworkModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_num_params = {
# Conv1 & MaxPool
'Conv2d_0': 720,
'MaxPool2d_0': 720,
# Stage 1 with 4 units
'Stage_0/Unit_0': 8374,
'Stage_0/Unit_1': 16146,
'Stage_0/Unit_2': 23918,
'Stage_0/Unit_3': 31690,
# Stage 2 with 8 units
'Stage_1/Unit_0': 75886,
'Stage_1/Unit_1': 104886,
'Stage_1/Unit_2': 133886,
'Stage_1/Unit_3': 162886,
'Stage_1/Unit_4': 191886,
'Stage_1/Unit_5': 220886,
'Stage_1/Unit_6': 249886,
'Stage_1/Unit_7': 278886,
# Stage 3 with 4 units
'Stage_2/Unit_0': 448014,
'Stage_2/Unit_1': 559838,
'Stage_2/Unit_2': 671662,
'Stage_2/Unit_3': 783486,
# Conv2d_1
'Conv2d_1': 1261694,
}
for scope, end_point in enumerate(endpoints_num_params):
with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v2.shufflenet_v2_base(inputs,
final_endpoint=end_point,
scope=str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=str(scope)))
self.assertAlmostEqual(endpoints_num_params[end_point],
total_params)
def testOutputStride8BuildAndCheckAllEndPointsUptoConv2d_1(self):
batch_size = 5
height, width = 224, 224
output_stride = 8
inputs = tf.random_uniform((batch_size, height, width, 3))
with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
normalizer_fn=slim.batch_norm):
_, end_points = shufflenet_v2.shufflenet_v2_base(
inputs, output_stride=output_stride, final_endpoint='Conv2d_1')
endpoints_shapes = {
# Conv1 & MaxPool
'Conv2d_0': [batch_size, 112, 112, 24],
'MaxPool2d_0': [batch_size, 56, 56, 24],
# Stage 1 with 4 units
'Stage_0/Unit_0': [batch_size, 28, 28, 116],
'Stage_0/Unit_1': [batch_size, 28, 28, 116],
'Stage_0/Unit_2': [batch_size, 28, 28, 116],
'Stage_0/Unit_3': [batch_size, 28, 28, 116],
# Stage 2 with 8 units
'Stage_1/Unit_0': [batch_size, 28, 28, 232],
'Stage_1/Unit_1': [batch_size, 28, 28, 232],
'Stage_1/Unit_2': [batch_size, 28, 28, 232],
'Stage_1/Unit_3': [batch_size, 28, 28, 232],
'Stage_1/Unit_4': [batch_size, 28, 28, 232],
'Stage_1/Unit_5': [batch_size, 28, 28, 232],
'Stage_1/Unit_6': [batch_size, 28, 28, 232],
'Stage_1/Unit_7': [batch_size, 28, 28, 232],
# Stage 3 with 4 units
'Stage_2/Unit_0': [batch_size, 28, 28, 464],
'Stage_2/Unit_1': [batch_size, 28, 28, 464],
'Stage_2/Unit_2': [batch_size, 28, 28, 464],
'Stage_2/Unit_3': [batch_size, 28, 28, 464],
# Conv2 to mix up features
'Conv2d_1': [batch_size, 28, 28, 1024]
}
self.assertItemsEqual(endpoints_shapes.keys(), end_points.keys())
for endpoint_name, expected_shape in endpoints_shapes.items():
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(
end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def testBuild2_0xBasicNetworkAndCheckAllEndPointsUptoConv2d_1(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
normalizer_fn=slim.batch_norm):
_, end_points = shufflenet_v2.shufflenet_v2_base(
inputs, depth_multiplier=2.0, final_endpoint='Conv2d_1')
endpoints_shapes = {
# Conv1 & MaxPool
'Conv2d_0': [batch_size, 112, 112, 24],
'MaxPool2d_0': [batch_size, 56, 56, 24],
# Stage 1 with 4 units
'Stage_0/Unit_0': [batch_size, 28, 28, 244],
'Stage_0/Unit_1': [batch_size, 28, 28, 244],
'Stage_0/Unit_2': [batch_size, 28, 28, 244],
'Stage_0/Unit_3': [batch_size, 28, 28, 244],
# Stage 2 with 8 units
'Stage_1/Unit_0': [batch_size, 14, 14, 488],
'Stage_1/Unit_1': [batch_size, 14, 14, 488],
'Stage_1/Unit_2': [batch_size, 14, 14, 488],
'Stage_1/Unit_3': [batch_size, 14, 14, 488],
'Stage_1/Unit_4': [batch_size, 14, 14, 488],
'Stage_1/Unit_5': [batch_size, 14, 14, 488],
'Stage_1/Unit_6': [batch_size, 14, 14, 488],
'Stage_1/Unit_7': [batch_size, 14, 14, 488],
# Stage 3 with 4 units
'Stage_2/Unit_0': [batch_size, 7, 7, 976],
'Stage_2/Unit_1': [batch_size, 7, 7, 976],
'Stage_2/Unit_2': [batch_size, 7, 7, 976],
'Stage_2/Unit_3': [batch_size, 7, 7, 976],
# Conv2 to mix up features
'Conv2d_1': [batch_size, 7, 7, 2048]
}
self.assertItemsEqual(endpoints_shapes.keys(), end_points.keys())
for endpoint_name, expected_shape in endpoints_shapes.items():
self.assertTrue(endpoint_name in end_points)
self.assertListEqual(
end_points[endpoint_name].get_shape().as_list(),
expected_shape)
def testBuildBaseNetwork(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
net, endpoints = shufflenet_v2.shufflenet_v2_base(inputs)
self.assertListEqual(net.get_shape().as_list(),
[batch_size, 7, 7, 1024])
self.maxDiff = None
expected_endpoints = [
# Conv1 & MaxPool
'Conv2d_0',
'MaxPool2d_0',
# Stage 1 with 4 units
'Stage_0/Unit_0',
'Stage_0/Unit_1',
'Stage_0/Unit_2',
'Stage_0/Unit_3',
# Stage 2 with 8 units
'Stage_1/Unit_0',
'Stage_1/Unit_1',
'Stage_1/Unit_2',
'Stage_1/Unit_3',
'Stage_1/Unit_4',
'Stage_1/Unit_5',
'Stage_1/Unit_6',
'Stage_1/Unit_7',
# Stage 3 with 4 units
'Stage_2/Unit_0',
'Stage_2/Unit_1',
'Stage_2/Unit_2',
'Stage_2/Unit_3',
# Conv2 to mix up features
'Conv2d_1',
]
self.assertItemsEqual(endpoints.keys(), expected_endpoints)
def testBuildOnlyUptoFinalEndpoint(self):
batch_size = 5
height, width = 224, 224
endpoints = [
# Conv1 & MaxPool
'Conv2d_0',
'MaxPool2d_0',
# Stage 1 with 4 units
'Stage_0/Unit_0',
'Stage_0/Unit_1',
'Stage_0/Unit_2',
'Stage_0/Unit_3',
# Stage 2 with 8 units
'Stage_1/Unit_0',
'Stage_1/Unit_1',
'Stage_1/Unit_2',
'Stage_1/Unit_3',
'Stage_1/Unit_4',
'Stage_1/Unit_5',
'Stage_1/Unit_6',
'Stage_1/Unit_7',
# Stage 3 with 4 units
'Stage_2/Unit_0',
'Stage_2/Unit_1',
'Stage_2/Unit_2',
'Stage_2/Unit_3',
# Conv2 to mix up features
'Conv2d_1',
]
for index, endpoint in enumerate(endpoints):
with tf.Graph().as_default():
inputs = tf.random_uniform((batch_size, height, width, 3))
out_tensor, end_points = shufflenet_v2.shufflenet_v2_base(
inputs, final_endpoint=endpoint)
print(out_tensor.op.name)
self.assertTrue(
out_tensor.op.name.startswith('ShufflenetV2/' + endpoint))
self.assertItemsEqual(endpoints[:index + 1], end_points.keys())
