def _test_downsample_unit():
x = tf.random.uniform((32, 56, 56, 24))
m1 = ShuffleNetV2DownsampleUnit(24, 116, se_reduction=None)
o1 = m1(x)
m2 = ShuffleNetV2DownsampleUnit(24, 116, se_reduction=16)
o2 = m2(x)
assert o1.shape == o2.shape == (32, 28, 28, 116)
assert get_num_params(m1) == 8554
assert get_num_params(m2) == 9571
def _test_preact_conv():
x = tf.random.uniform((32, 56, 56, 64))
m = preact_conv1x1(64, 128)
m2 = preact_conv3x3(128, 32)
o1 = m(x)
assert o1.shape == (32, 56, 56, 128)
assert get_num_params(m) == 64 * 2 + 128 * 64 * (1 * 1 + 0)
o2 = m2(o1)
assert o2.shape == (32, 56, 56, 32)
assert get_num_params(m2) == 128 * 2 + 128 * 32 * (3 * 3 + 0)
def _test_mnasnet():
x = tf.random.uniform((32, 224, 224, 3))
m = get_mnasnet_a1_s1()
o = m(x)
assert o.shape == (32, 1000)
assert get_num_params(m) == 3665608
m2 = get_mnasnet_b1_s1()
o2 = m2(x)
assert o2.shape == (32, 1000)
assert get_num_params(m2) == 4383312
def _test_fire():
x = tf.random.uniform((32, 224, 224, 64))
m = Fire(64, 8, 32, 32, False)
o = m(x)
assert o.shape == (32, 224, 224, 64) and get_num_params(
m) == (64 + 1) * 8 + (8 + 1) * 32 + (8 * 9 + 1) * 32
m2 = Conv2d(64, 64, 3, 1, 1)
o = m2(x)
assert o.shape == (32, 224, 224,
64) and get_num_params(m2) == (64 * 9 + 1) * 64
def _test_sepconv_block():
x = tf.random.uniform((32, 112, 112, 32))
m = SepConv(32, 16)
o = m(x)
assert o.shape == (32, 112, 112, 16)
assert get_num_params(
m) == 32 * 1 * (3 * 3 + 0) + 32 * 2 + 32 * 16 * (1 * 1 + 0) + 16 * 2
def _test_denseblock():
x = tf.random.uniform((32, 56, 56, 64))
bn_size, growth_rate = 4, 32
m = DenseBlock(in_channels=64, num_layers=6, bn_size=bn_size, growth_rate=growth_rate)
o = m(x)
assert o.shape == (32, 56, 56, 64 + 6 * 32)
assert get_num_params(m) == 335040
def _test_basic_unit():
x = tf.random.uniform((32, 28, 28, 116))
m1 = ShuffleNetV2BasicUnit(116)
o1 = m1(x)
m2 = ShuffleNetV2BasicUnit(116, se_reduction=16)
o2 = m2(x)
m3 = ShuffleNetV2BasicUnit(116, residual=True)
o3 = m3(x)
m4 = ShuffleNetV2BasicUnit(116, se_reduction=16, residual=True)
o4 = m4(x)
for i, (o, m) in enumerate(zip([o1, o2, o3, o4], [m1, m2, m3, m4])):
assert o.shape == (32, 28, 28, 116)
if i % 2 == 0:
assert get_num_params(m) == 7598
else:
assert get_num_params(m) == 8007
def _test_googlenet():
x = tf.random.uniform((32, 224, 224, 3))
m = GoogleNet()
o, a1, a2 = m(x, training=True)
assert o.shape == a1.shape == a2.shape == (32, 1000)
o = m(x)
assert o.shape == (32, 1000)
assert get_num_params(m) == 11851864
def _test_mbconvblock():
x = tf.random.uniform((32, 112, 112, 16))
m = MBConv(in_channels=16,
out_channels=24,
kernel_size=3,
strides=2,
expansion_factor=6)
x = m(x)
assert x.shape == (32, 56, 56, 24)
conv1_expand_params = 16 * 16 * 6 * (1 * 1 + 0) + 16 * 6 * 2
conv3_params = 1 * 16 * 6 * (3 * 3 + 0) + 16 * 6 * 2
conv1_projec_params = 16 * 6 * 24 * (1 * 1 + 0) + 24 * 2
assert get_num_params(
m) == conv1_expand_params + conv3_params + conv1_projec_params
m2 = MBConv(in_channels=24,
out_channels=24,
kernel_size=3,
strides=1,
expansion_factor=6)
x = m2(x)
assert x.shape == (32, 56, 56, 24)
assert m2.residual
conv1_expand_params = 24 * 24 * 6 * (1 * 1 + 0) + 24 * 6 * 2
conv3_params = 1 * 24 * 6 * (3 * 3 + 0) + 24 * 6 * 2
conv1_projec_params = 24 * 6 * 24 * (1 * 1 + 0) + 24 * 2
assert get_num_params(
m2) == conv1_expand_params + conv3_params + conv1_projec_params
m3 = MBConv(in_channels=24,
out_channels=40,
kernel_size=5,
strides=2,
expansion_factor=3,
se_ratio=.25)
x = m3(x)
assert x.shape == (32, 28, 28, 40)
assert get_num_params(m3) == 7428
m4 = MBConv(in_channels=40,
out_channels=40,
kernel_size=5,
strides=1,
expansion_factor=3,
se_ratio=.25)
x = m4(x)
assert x.shape == (32, 28, 28, 40)
assert m4.residual
def _test_vgg():
x = tf.random.uniform((32, 224, 224, 3))
model = vgg11()
out = model(x, training=True)
assert out.shape == (32, 1000)
assert get_num_params(model) == 132866088
model = vgg13()
out = model(x, training=True)
assert out.shape == (32, 1000)
assert get_num_params(model) == 133050792
model = vgg16()
out = model(x, training=True)
assert out.shape == (32, 1000)
assert get_num_params(model) == 138361768
model = vgg19()
out = model(x, training=True)
assert out.shape == (32, 1000)
assert get_num_params(model) == 143672744
def _test_densenet():
x = tf.random.uniform((32, 224, 224, 3))
models = [get_densenet121, get_densenet161, get_densenet169, get_densenet201]
num_params = [7978856, 28681000, 14149480, 20013928]
for model, num_param in zip(models, num_params):
m = model()
o = m(x)
assert o.shape == (32, 1000)
assert get_num_params(m) == num_param
def _test_squeezenet():
x = tf.random.uniform((32, 224, 224, 3))
model = squeezenet_v1_0()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 1248424
model = squeezenet_v1_1()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 1235496
model = squeezeresnet_v1_0()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 1248424
model = squeezeresnet_v1_1()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 1235496
def _test_denseunit():
x = tf.random.uniform((32, 56, 56, 64))
bn_size, growth_rate = 4, 32
m = DenseUnit(64, bn_size=bn_size, growth_rate=growth_rate, dropout_rate=.1)
o = m(x)
assert o.shape == ((32, 56, 56, growth_rate))
conv1_params = 64 * 2 + 64 * (bn_size * growth_rate) * (1 * 1 + 0)
conv2_params = (bn_size * growth_rate) * 2 + (bn_size * growth_rate) * growth_rate * (3 * 3 + 0)
num_params = conv1_params + conv2_params
assert get_num_params(m) == num_params
def _test_alexnet(show_summary=False):
alexnet = construct_alex_net()
# alexnet.build(input_shape=(None, 224, 224, 3))
# model.compile(optimizer=tf.keras.optimizers.Adam())
x = tf.random.uniform((32, 224, 224, 3))
out = alexnet(x, training=True)
assert out.shape == (32, 1000)
assert get_num_params(alexnet) == 61100840
if show_summary:
print(alexnet.summary())
def _test_shufflenetv2():
x = tf.random.uniform((32, 224, 224, 3))
m1 = ShuffleNetV2()
o1 = m1(x)
m2 = ShuffleNetV2(se_reduction=16)
o2 = m2(x)
m3 = get_shufflenetv2(residual=True)
o3 = m3(x)
m4 = get_shufflenetv2(se_reduction=16, residual=True)
o4 = m4(x)
m5 = ShuffleNetV2(out_channels=2048, c=[244, 488, 976])
o5 = m5(x)
m6 = get_shufflenetv2(se_reduction=16,
residual=True,
n_groups=4,
width_scale='2.0')
o6 = m6(x)
assert o1.shape == o2.shape == o3.shape == o4.shape == o5.shape == o6.shape == (
32, 1000)
assert get_num_params(m1) == get_num_params(m3) == 2279760
assert get_num_params(m2) == get_num_params(m4) == 2322948
assert get_num_params(m5) == 7403600
assert get_num_params(m6) == 7594888
def _test_resnet():
x = tf.random.uniform((32, 224, 224, 3))
model = resnet18()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 11689512
model = resnet34()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 21797672
model = resnet50()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 25557032
model = resnet101()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 44549160
model = resnet152()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 60192808
model = resnext50_32x4d()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 25028904
model = resnext101_32x8d()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 44177704
model = wide_resnet50_2()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 68883240
model = wide_resnet101_2()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 126886696
model = seresnext_50()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 27559896
model = seresnext_101()
out = model(x)
assert out.shape == (32, 1000) and get_num_params(model) == 48955416
def _test_trasitblock():
x = tf.random.uniform((32, 56, 56, 256))
m = TransitBlock(in_channels=256, out_channels=128)
o = m(x)
assert o.shape == (32, 28, 28, 128)
assert get_num_params(m) == 256 * 2 + 256 * 128 * (1 * 1 + 0)
def _test_mobilenet():
x = tf.random.uniform((32, 224, 224, 3))
m = MobileNetV2()
o = m(x)
assert o.shape == (32, 1000)
assert get_num_params(m) == 3504872
