def testFlops(self):
def stats_graph(graph):
flops = tf.profiler.profile(
graph,
options=tf.profiler.ProfileOptionBuilder.float_operation())
params = tf.profiler.profile(
graph,
options=tf.profiler.ProfileOptionBuilder.
trainable_variables_parameter())
print('FLOPs: {}; Trainable params: {}'.format(
flops.total_float_ops, params.total_parameters))
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
with slim.arg_scope(shufflenet_v1.shufflenet_v1_arg_scope()):
shufflenet_v1.shufflenet_v1_base(inputs, depth_multiplier=2.0)
graph = tf.get_default_graph()
stats_graph(graph)
self.assertTrue(False)
def testOutputStride16BuildAndCheckAllEndPointsUptoStage_2_Sub_Unit_3(
self):
batch_size = 5
height, width = 224, 224
output_stride = 16
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_v1.shufflenet_v1_base(
inputs,
output_stride=output_stride,
final_endpoint='Stage_2/Unit_3')
endpoints_shapes = {
# regular conv and pool
'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, 240],
'Stage_0/Unit_1': [batch_size, 28, 28, 240],
'Stage_0/Unit_2': [batch_size, 28, 28, 240],
'Stage_0/Unit_3': [batch_size, 28, 28, 240],
# Stage 2 with 8 units
'Stage_1/Unit_0': [batch_size, 14, 14, 480],
'Stage_1/Unit_1': [batch_size, 14, 14, 480],
'Stage_1/Unit_2': [batch_size, 14, 14, 480],
'Stage_1/Unit_3': [batch_size, 14, 14, 480],
'Stage_1/Unit_4': [batch_size, 14, 14, 480],
'Stage_1/Unit_5': [batch_size, 14, 14, 480],
'Stage_1/Unit_6': [batch_size, 14, 14, 480],
'Stage_1/Unit_7': [batch_size, 14, 14, 480],
# Stage 3 with 4 units
'Stage_2/Unit_0': [batch_size, 14, 14, 960],
'Stage_2/Unit_1': [batch_size, 14, 14, 960],
'Stage_2/Unit_2': [batch_size, 14, 14, 960],
'Stage_2/Unit_3': [batch_size, 14, 14, 960],
}
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 testBuildAndCheckAllEndPointsApproximateFaceNet(self):
batch_size = 5
height, width = 128, 128
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_v1.shufflenet_v1_base(
inputs, final_endpoint='Stage_2/Unit_3', depth_multiplier=0.75)
# For the Conv2d_0 layer FaceNet has depth=16
endpoints_shapes = {
# regular conv and pool
'Conv2d_0': [batch_size, 64, 64, 24],
'MaxPool2d_0': [batch_size, 32, 32, 24],
# Stage 1 with 4 units
'Stage_0/Unit_0': [batch_size, 16, 16, 180],
'Stage_0/Unit_1': [batch_size, 16, 16, 180],
'Stage_0/Unit_2': [batch_size, 16, 16, 180],
'Stage_0/Unit_3': [batch_size, 16, 16, 180],
# Stage 2 with 8 units
'Stage_1/Unit_0': [batch_size, 8, 8, 360],
'Stage_1/Unit_1': [batch_size, 8, 8, 360],
'Stage_1/Unit_2': [batch_size, 8, 8, 360],
'Stage_1/Unit_3': [batch_size, 8, 8, 360],
'Stage_1/Unit_4': [batch_size, 8, 8, 360],
'Stage_1/Unit_5': [batch_size, 8, 8, 360],
'Stage_1/Unit_6': [batch_size, 8, 8, 360],
'Stage_1/Unit_7': [batch_size, 8, 8, 360],
# Stage 3 with 4 units
'Stage_2/Unit_0': [batch_size, 4, 4, 720],
'Stage_2/Unit_1': [batch_size, 4, 4, 720],
'Stage_2/Unit_2': [batch_size, 4, 4, 720],
'Stage_2/Unit_3': [batch_size, 4, 4, 720],
}
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 testBuildCustomNetworkUsingDepthChannelsDefs(self):
batch_size = 5
height, width = 224, 224
depth_channels_defs = {
'1': [144, 288, 576],
'2': [200, 400, 800],
'3': [480, 960, 1440],
'4': [272, 544, 1088],
'8': [384, 768, 1536],
}
inputs = tf.random_uniform((batch_size, height, width, 3))
net, end_points = shufflenet_v1.shufflenet_v1_base(
inputs,
final_endpoint='Stage_1/Unit_7',
depth_channels_defs=depth_channels_defs)
self.assertTrue(net.op.name.startswith('ShufflenetV1/Stage_1/Unit_7'))
self.assertListEqual(net.get_shape().as_list(),
[batch_size, 14, 14, 960])
expected_endpoints = [
# regular conv and pool
'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',
]
self.assertItemsEqual(end_points.keys(), expected_endpoints)
def testBuildBaseNetwork(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
net, endpoints = shufflenet_v1.shufflenet_v1_base(inputs)
print(net.op.name)
# self.assertTrue(net.op.name.startswith('ShufflenetV1/stage3/shufflenet_unit_3'))
self.assertListEqual(net.get_shape().as_list(),
[batch_size, 7, 7, 960])
expected_endpoints = [
# regular conv and pool
'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',
]
self.maxDiff = None
self.assertItemsEqual(endpoints.keys(), expected_endpoints)
def testBuildOnlyUptoFinalEndpoint(self):
batch_size = 5
height, width = 224, 224
endpoints = [
# regular conv and pool
'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',
]
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_v1.shufflenet_v1_base(
inputs, final_endpoint=endpoint)
print(out_tensor.op.name)
self.assertTrue(
out_tensor.op.name.startswith('ShufflenetV1/' + endpoint))
self.assertItemsEqual(endpoints[:index + 1], end_points.keys())
def testModelHasExpectedNumberOfParametersWithRate2_0(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_groups_2_0 = {
'1': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 23328,
'Stage_0/Unit_1': 66744,
'Stage_0/Unit_2': 110160,
'Stage_0/Unit_3': 153576,
'Stage_1/Unit_0': 239544,
'Stage_1/Unit_1': 409320,
'Stage_1/Unit_2': 579096,
'Stage_1/Unit_3': 748872,
'Stage_1/Unit_4': 918648,
'Stage_1/Unit_5': 1088424,
'Stage_1/Unit_6': 1258200,
'Stage_1/Unit_7': 1427976,
'Stage_2/Unit_0': 1765800,
'Stage_2/Unit_1': 2437128,
'Stage_2/Unit_2': 3108456,
'Stage_2/Unit_3': 3779784
},
'2': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 24548,
'Stage_0/Unit_1': 67248,
'Stage_0/Unit_2': 109948,
'Stage_0/Unit_3': 152648,
'Stage_1/Unit_0': 236848,
'Stage_1/Unit_1': 402248,
'Stage_1/Unit_2': 567648,
'Stage_1/Unit_3': 733048,
'Stage_1/Unit_4': 898448,
'Stage_1/Unit_5': 1063848,
'Stage_1/Unit_6': 1229248,
'Stage_1/Unit_7': 1394648,
'Stage_2/Unit_0': 1723048,
'Stage_2/Unit_1': 2373848,
'Stage_2/Unit_2': 3024648,
'Stage_2/Unit_3': 3675448
},
'3': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 25008,
'Stage_0/Unit_1': 66648,
'Stage_0/Unit_2': 108288,
'Stage_0/Unit_3': 149928,
'Stage_1/Unit_0': 231768,
'Stage_1/Unit_1': 391848,
'Stage_1/Unit_2': 551928,
'Stage_1/Unit_3': 712008,
'Stage_1/Unit_4': 872088,
'Stage_1/Unit_5': 1032168,
'Stage_1/Unit_6': 1192248,
'Stage_1/Unit_7': 1352328,
'Stage_2/Unit_0': 1669608,
'Stage_2/Unit_1': 2296968,
'Stage_2/Unit_2': 2924328,
'Stage_2/Unit_3': 3551688
},
'4': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 25264,
'Stage_0/Unit_1': 65928,
'Stage_0/Unit_2': 106592,
'Stage_0/Unit_3': 147256,
'Stage_1/Unit_0': 226952,
'Stage_1/Unit_1': 382264,
'Stage_1/Unit_2': 537576,
'Stage_1/Unit_3': 692888,
'Stage_1/Unit_4': 848200,
'Stage_1/Unit_5': 1003512,
'Stage_1/Unit_6': 1158824,
'Stage_1/Unit_7': 1314136,
'Stage_2/Unit_0': 1621496,
'Stage_2/Unit_1': 2228056,
'Stage_2/Unit_2': 2834616,
'Stage_2/Unit_3': 3441176
},
'8': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 28296,
'Stage_0/Unit_1': 70344,
'Stage_0/Unit_2': 112392,
'Stage_0/Unit_3': 154440,
'Stage_1/Unit_0': 236232,
'Stage_1/Unit_1': 394056,
'Stage_1/Unit_2': 551880,
'Stage_1/Unit_3': 709704,
'Stage_1/Unit_4': 867528,
'Stage_1/Unit_5': 1025352,
'Stage_1/Unit_6': 1183176,
'Stage_1/Unit_7': 1341000,
'Stage_2/Unit_0': 1652040,
'Stage_2/Unit_1': 2262600,
'Stage_2/Unit_2': 2873160,
'Stage_2/Unit_3': 3483720
},
}
for num_groups in endpoints_groups_2_0:
for scope, end_point in enumerate(
endpoints_groups_2_0[num_groups]):
print(num_groups, end_point)
with slim.arg_scope(
[slim.conv2d, slim.separable_conv2d, group_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v1.shufflenet_v1_base(
inputs,
final_endpoint=end_point,
num_groups=int(num_groups),
depth_multiplier=2.0,
scope=num_groups + '/' + str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=num_groups + '/' +
str(scope)))
self.assertAlmostEqual(
endpoints_groups_2_0[num_groups][end_point],
total_params)
def testModelHasExpectedNumberOfParametersWithRate1_5(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_groups_1_5 = {
'1': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 13770,
'Stage_0/Unit_1': 38556,
'Stage_0/Unit_2': 63342,
'Stage_0/Unit_3': 88128,
'Stage_1/Unit_0': 137052,
'Stage_1/Unit_1': 233280,
'Stage_1/Unit_2': 329508,
'Stage_1/Unit_3': 425736,
'Stage_1/Unit_4': 521964,
'Stage_1/Unit_5': 618192,
'Stage_1/Unit_6': 714420,
'Stage_1/Unit_7': 810648,
'Stage_2/Unit_0': 1001808,
'Stage_2/Unit_1': 1380888,
'Stage_2/Unit_2': 1759968,
'Stage_2/Unit_3': 2139048
},
'2': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 14646,
'Stage_0/Unit_1': 38856,
'Stage_0/Unit_2': 63066,
'Stage_0/Unit_3': 87276,
'Stage_1/Unit_0': 135426,
'Stage_1/Unit_1': 229476,
'Stage_1/Unit_2': 323526,
'Stage_1/Unit_3': 417576,
'Stage_1/Unit_4': 511626,
'Stage_1/Unit_5': 605676,
'Stage_1/Unit_6': 699726,
'Stage_1/Unit_7': 793776,
'Stage_2/Unit_0': 980076,
'Stage_2/Unit_1': 1348176,
'Stage_2/Unit_2': 1716276,
'Stage_2/Unit_3': 2084376
},
'3': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 15318,
'Stage_0/Unit_1': 39348,
'Stage_0/Unit_2': 63378,
'Stage_0/Unit_3': 87408,
'Stage_1/Unit_0': 134388,
'Stage_1/Unit_1': 225648,
'Stage_1/Unit_2': 316908,
'Stage_1/Unit_3': 408168,
'Stage_1/Unit_4': 499428,
'Stage_1/Unit_5': 590688,
'Stage_1/Unit_6': 681948,
'Stage_1/Unit_7': 773208,
'Stage_2/Unit_0': 953568,
'Stage_2/Unit_1': 1308888,
'Stage_2/Unit_2': 1664208,
'Stage_2/Unit_3': 2019528
},
'4': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 15372,
'Stage_0/Unit_1': 38496,
'Stage_0/Unit_2': 61620,
'Stage_0/Unit_3': 84744,
'Stage_1/Unit_0': 130644,
'Stage_1/Unit_1': 219384,
'Stage_1/Unit_2': 308124,
'Stage_1/Unit_3': 396864,
'Stage_1/Unit_4': 485604,
'Stage_1/Unit_5': 574344,
'Stage_1/Unit_6': 663084,
'Stage_1/Unit_7': 751824,
'Stage_2/Unit_0': 926856,
'Stage_2/Unit_1': 1270800,
'Stage_2/Unit_2': 1614744,
'Stage_2/Unit_3': 1958688
},
'8': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 17928,
'Stage_0/Unit_1': 42552,
'Stage_0/Unit_2': 67176,
'Stage_0/Unit_3': 91800,
'Stage_1/Unit_0': 139320,
'Stage_1/Unit_1': 230040,
'Stage_1/Unit_2': 320760,
'Stage_1/Unit_3': 411480,
'Stage_1/Unit_4': 502200,
'Stage_1/Unit_5': 592920,
'Stage_1/Unit_6': 683640,
'Stage_1/Unit_7': 774360,
'Stage_2/Unit_0': 952344,
'Stage_2/Unit_1': 1299672,
'Stage_2/Unit_2': 1647000,
'Stage_2/Unit_3': 1994328
},
}
for num_groups in endpoints_groups_1_5:
for scope, end_point in enumerate(
endpoints_groups_1_5[num_groups]):
print(num_groups, end_point)
with slim.arg_scope(
[slim.conv2d, slim.separable_conv2d, group_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v1.shufflenet_v1_base(
inputs,
final_endpoint=end_point,
depth_multiplier=1.5,
num_groups=int(num_groups),
scope=num_groups + '/' + str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=num_groups + '/' +
str(scope)))
self.assertAlmostEqual(
endpoints_groups_1_5[num_groups][end_point],
total_params)
def testModelHasExpectedNumberOfParametersWithRate0_5(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_groups_0_5 = {
'1': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 2430,
'Stage_0/Unit_1': 5508,
'Stage_0/Unit_2': 8586,
'Stage_0/Unit_3': 11664,
'Stage_1/Unit_0': 17604,
'Stage_1/Unit_1': 28944,
'Stage_1/Unit_2': 40284,
'Stage_1/Unit_3': 51624,
'Stage_1/Unit_4': 62964,
'Stage_1/Unit_5': 74304,
'Stage_1/Unit_6': 85644,
'Stage_1/Unit_7': 96984,
'Stage_2/Unit_0': 119232,
'Stage_2/Unit_1': 162648,
'Stage_2/Unit_2': 206064,
'Stage_2/Unit_3': 249480
},
'2': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 2796,
'Stage_0/Unit_1': 5856,
'Stage_0/Unit_2': 8916,
'Stage_0/Unit_3': 11976,
'Stage_1/Unit_0': 18026,
'Stage_1/Unit_1': 29376,
'Stage_1/Unit_2': 40726,
'Stage_1/Unit_3': 52076,
'Stage_1/Unit_4': 63426,
'Stage_1/Unit_5': 74776,
'Stage_1/Unit_6': 86126,
'Stage_1/Unit_7': 97476,
'Stage_2/Unit_0': 119576,
'Stage_2/Unit_1': 162276,
'Stage_2/Unit_2': 204976,
'Stage_2/Unit_3': 247676
},
'3': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 3138,
'Stage_0/Unit_1': 6348,
'Stage_0/Unit_2': 9558,
'Stage_0/Unit_3': 12768,
'Stage_1/Unit_0': 18828,
'Stage_1/Unit_1': 30048,
'Stage_1/Unit_2': 41268,
'Stage_1/Unit_3': 52488,
'Stage_1/Unit_4': 63708,
'Stage_1/Unit_5': 74928,
'Stage_1/Unit_6': 86148,
'Stage_1/Unit_7': 97368,
'Stage_2/Unit_0': 119088,
'Stage_2/Unit_1': 160728,
'Stage_2/Unit_2': 202368,
'Stage_2/Unit_3': 244008
},
'4': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 3200,
'Stage_0/Unit_1': 6264,
'Stage_0/Unit_2': 9328,
'Stage_0/Unit_3': 12392,
'Stage_1/Unit_0': 18444,
'Stage_1/Unit_1': 29528,
'Stage_1/Unit_2': 40612,
'Stage_1/Unit_3': 51696,
'Stage_1/Unit_4': 62780,
'Stage_1/Unit_5': 73864,
'Stage_1/Unit_6': 84948,
'Stage_1/Unit_7': 96032,
'Stage_2/Unit_0': 117384,
'Stage_2/Unit_1': 158048,
'Stage_2/Unit_2': 198712,
'Stage_2/Unit_3': 239376
},
'8': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 4104,
'Stage_0/Unit_1': 7704,
'Stage_0/Unit_2': 11304,
'Stage_0/Unit_3': 14904,
'Stage_1/Unit_0': 21528,
'Stage_1/Unit_1': 33336,
'Stage_1/Unit_2': 45144,
'Stage_1/Unit_3': 56952,
'Stage_1/Unit_4': 68760,
'Stage_1/Unit_5': 80568,
'Stage_1/Unit_6': 92376,
'Stage_1/Unit_7': 104184,
'Stage_2/Unit_0': 126648,
'Stage_2/Unit_1': 168696,
'Stage_2/Unit_2': 210744,
'Stage_2/Unit_3': 252792
},
}
for num_groups in endpoints_groups_0_5:
for scope, end_point in enumerate(
endpoints_groups_0_5[num_groups]):
print(num_groups, end_point)
with slim.arg_scope(
[slim.conv2d, slim.separable_conv2d, group_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v1.shufflenet_v1_base(
inputs,
final_endpoint=end_point,
depth_multiplier=0.5,
num_groups=int(num_groups),
scope=num_groups + '/' + str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=num_groups + '/' +
str(scope)))
self.assertAlmostEqual(
endpoints_groups_0_5[num_groups][end_point],
total_params)
def testModelHasExpectedNumberOfParametersWithRate0_25(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_groups_0_25 = {
'1': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 1215,
'Stage_0/Unit_1': 2106,
'Stage_0/Unit_2': 2997,
'Stage_0/Unit_3': 3888,
'Stage_1/Unit_0': 5562,
'Stage_1/Unit_1': 8640,
'Stage_1/Unit_2': 11718,
'Stage_1/Unit_3': 14796,
'Stage_1/Unit_4': 17874,
'Stage_1/Unit_5': 20952,
'Stage_1/Unit_6': 24030,
'Stage_1/Unit_7': 27108,
'Stage_2/Unit_0': 33048,
'Stage_2/Unit_1': 44388,
'Stage_2/Unit_2': 55728,
'Stage_2/Unit_3': 67068
},
'2': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 1422,
'Stage_0/Unit_1': 2352,
'Stage_0/Unit_2': 3282,
'Stage_0/Unit_3': 4212,
'Stage_1/Unit_0': 5922,
'Stage_1/Unit_1': 8982,
'Stage_1/Unit_2': 12042,
'Stage_1/Unit_3': 15102,
'Stage_1/Unit_4': 18162,
'Stage_1/Unit_5': 21222,
'Stage_1/Unit_6': 24282,
'Stage_1/Unit_7': 27342,
'Stage_2/Unit_0': 33392,
'Stage_2/Unit_1': 44742,
'Stage_2/Unit_2': 56092,
'Stage_2/Unit_3': 67442
},
'3': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 1593,
'Stage_0/Unit_1': 2598,
'Stage_0/Unit_2': 3603,
'Stage_0/Unit_3': 4608,
'Stage_1/Unit_0': 6438,
'Stage_1/Unit_1': 9648,
'Stage_1/Unit_2': 12858,
'Stage_1/Unit_3': 16068,
'Stage_1/Unit_4': 19278,
'Stage_1/Unit_5': 22488,
'Stage_1/Unit_6': 25698,
'Stage_1/Unit_7': 28908,
'Stage_2/Unit_0': 34968,
'Stage_2/Unit_1': 46188,
'Stage_2/Unit_2': 57408,
'Stage_2/Unit_3': 68628
},
'4': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 1652,
'Stage_0/Unit_1': 2640,
'Stage_0/Unit_2': 3628,
'Stage_0/Unit_3': 4616,
'Stage_1/Unit_0': 6388,
'Stage_1/Unit_1': 9452,
'Stage_1/Unit_2': 12516,
'Stage_1/Unit_3': 15580,
'Stage_1/Unit_4': 18644,
'Stage_1/Unit_5': 21708,
'Stage_1/Unit_6': 24772,
'Stage_1/Unit_7': 27836,
'Stage_2/Unit_0': 33888,
'Stage_2/Unit_1': 44972,
'Stage_2/Unit_2': 56056,
'Stage_2/Unit_3': 67140
},
'8': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 2088,
'Stage_0/Unit_1': 3312,
'Stage_0/Unit_2': 4536,
'Stage_0/Unit_3': 5760,
'Stage_1/Unit_0': 7920,
'Stage_1/Unit_1': 11520,
'Stage_1/Unit_2': 15120,
'Stage_1/Unit_3': 18720,
'Stage_1/Unit_4': 22320,
'Stage_1/Unit_5': 25920,
'Stage_1/Unit_6': 29520,
'Stage_1/Unit_7': 33120,
'Stage_2/Unit_0': 39744,
'Stage_2/Unit_1': 51552,
'Stage_2/Unit_2': 63360,
'Stage_2/Unit_3': 75168
},
}
for num_groups in endpoints_groups_0_25:
for scope, end_point in enumerate(
endpoints_groups_0_25[num_groups]):
print(num_groups, end_point)
with slim.arg_scope(
[slim.conv2d, slim.separable_conv2d, group_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v1.shufflenet_v1_base(
inputs,
final_endpoint=end_point,
depth_multiplier=0.25,
num_groups=int(num_groups),
scope=num_groups + '/' + str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=num_groups + '/' +
str(scope)))
self.assertAlmostEqual(
endpoints_groups_0_25[num_groups][end_point],
total_params)
def testModelHasExpectedNumberOfParameters(self):
batch_size = 5
height, width = 224, 224
inputs = tf.random_uniform((batch_size, height, width, 3))
endpoints_groups = {
'1': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 6804,
'Stage_0/Unit_1': 18144,
'Stage_0/Unit_2': 29484,
'Stage_0/Unit_3': 40824,
'Stage_1/Unit_0': 63072,
'Stage_1/Unit_1': 106488,
'Stage_1/Unit_2': 149904,
'Stage_1/Unit_3': 193320,
'Stage_1/Unit_4': 236736,
'Stage_1/Unit_5': 280152,
'Stage_1/Unit_6': 323568,
'Stage_1/Unit_7': 366984,
'Stage_2/Unit_0': 452952,
'Stage_2/Unit_1': 622728,
'Stage_2/Unit_2': 792504,
'Stage_2/Unit_3': 962280
},
'2': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 7598,
'Stage_0/Unit_1': 18948,
'Stage_0/Unit_2': 30298,
'Stage_0/Unit_3': 41648,
'Stage_1/Unit_0': 63748,
'Stage_1/Unit_1': 106448,
'Stage_1/Unit_2': 149148,
'Stage_1/Unit_3': 191848,
'Stage_1/Unit_4': 234548,
'Stage_1/Unit_5': 277248,
'Stage_1/Unit_6': 319948,
'Stage_1/Unit_7': 362648,
'Stage_2/Unit_0': 446848,
'Stage_2/Unit_1': 612248,
'Stage_2/Unit_2': 777648,
'Stage_2/Unit_3': 943048
},
'3': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 8028,
'Stage_0/Unit_1': 19248,
'Stage_0/Unit_2': 30468,
'Stage_0/Unit_3': 41688,
'Stage_1/Unit_0': 63408,
'Stage_1/Unit_1': 105048,
'Stage_1/Unit_2': 146688,
'Stage_1/Unit_3': 188328,
'Stage_1/Unit_4': 229968,
'Stage_1/Unit_5': 271608,
'Stage_1/Unit_6': 313248,
'Stage_1/Unit_7': 354888,
'Stage_2/Unit_0': 436728,
'Stage_2/Unit_1': 596808,
'Stage_2/Unit_2': 756888,
'Stage_2/Unit_3': 916968
},
'4': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 8332,
'Stage_0/Unit_1': 19416,
'Stage_0/Unit_2': 30500,
'Stage_0/Unit_3': 41584,
'Stage_1/Unit_0': 62936,
'Stage_1/Unit_1': 103600,
'Stage_1/Unit_2': 144264,
'Stage_1/Unit_3': 184928,
'Stage_1/Unit_4': 225592,
'Stage_1/Unit_5': 266256,
'Stage_1/Unit_6': 306920,
'Stage_1/Unit_7': 347584,
'Stage_2/Unit_0': 427280,
'Stage_2/Unit_1': 582592,
'Stage_2/Unit_2': 737904,
'Stage_2/Unit_3': 893216
},
'8': {
'Conv2d_0': 720,
'MaxPool2d_0': 720,
'Stage_0/Unit_0': 9864,
'Stage_0/Unit_1': 21672,
'Stage_0/Unit_2': 33480,
'Stage_0/Unit_3': 45288,
'Stage_1/Unit_0': 67752,
'Stage_1/Unit_1': 109800,
'Stage_1/Unit_2': 151848,
'Stage_1/Unit_3': 193896,
'Stage_1/Unit_4': 235944,
'Stage_1/Unit_5': 277992,
'Stage_1/Unit_6': 320040,
'Stage_1/Unit_7': 362088,
'Stage_2/Unit_0': 443880,
'Stage_2/Unit_1': 601704,
'Stage_2/Unit_2': 759528,
'Stage_2/Unit_3': 917352
},
}
for num_groups in endpoints_groups:
for scope, end_point in enumerate(endpoints_groups[num_groups]):
print(num_groups, end_point)
with slim.arg_scope(
[slim.conv2d, slim.separable_conv2d, group_conv2d],
normalizer_fn=slim.batch_norm):
shufflenet_v1.shufflenet_v1_base(
inputs,
final_endpoint=end_point,
num_groups=int(num_groups),
scope=num_groups + '/' + str(scope))
total_params, _ = slim.model_analyzer.analyze_vars(
slim.get_model_variables(scope=num_groups + '/' +
str(scope)))
self.assertAlmostEqual(
endpoints_groups[num_groups][end_point], total_params)