def test_resnet50_pretrained(self):
resnet50 = resnet.ResNet50(pretrained=True, features_output_stride=16)
x_in = tf.keras.layers.Input((160, 160, 1))
x, x_mid = resnet50.make_backbone(x_in)
model = tf.keras.Model(x_in, x)
param_counts = [
np.prod(train_var.shape) for train_var in model.trainable_weights
]
with self.subTest("number of layers"):
self.assertEqual(len(model.layers),
175 + 2) # adds preprocessing layers
with self.subTest("number of trainable weights"):
self.assertEqual(len(model.trainable_weights), 212)
with self.subTest("trainable parameter count"):
self.assertEqual(np.sum(param_counts), 23534592)
with self.subTest("total parameter count"):
self.assertEqual(model.count_params(), 23587712)
with self.subTest("exact pretrained weights"):
layer = model.get_layer("conv5_block1_1_conv")
self.assertAllClose(
layer.kernel[:, :, :3, :3],
np.array([[[
[0.00946995, 0.00118913, -0.01149864],
[-0.01221565, 0.01268964, -0.00900999],
[0.05596072, 0.01261912, -0.01511286],
]]]),
)
def test_resnet50_stride16(self):
resnet50 = resnet.ResNet50(pretrained=False, features_output_stride=16)
x_in = tf.keras.layers.Input((160, 160, 1))
x, x_mid = resnet50.make_backbone(x_in)
model = tf.keras.Model(x_in, x)
param_counts = [
np.prod(train_var.shape) for train_var in model.trainable_weights
]
with self.subTest("number of layers"):
self.assertEqual(len(model.layers), 175)
with self.subTest("number of trainable weights"):
self.assertEqual(len(model.trainable_weights), 212)
with self.subTest("trainable parameter count"):
self.assertEqual(np.sum(param_counts), 23528320)
with self.subTest("total parameter count"):
self.assertEqual(model.count_params(), 23581440)
with self.subTest("output shape"):
self.assertAllEqual(model.output.shape, (None, 10, 10, 2048))
with self.subTest("feature output stride"):
self.assertEqual(
model.get_layer("conv5_block1_1_conv").strides, (1, 1))
with self.subTest("feature output dilation rate"):
self.assertEqual(
model.get_layer("conv5_block1_1_conv").dilation_rate, (2, 2))
def test_resnet50_upsampling(self):
resnet50 = resnet.ResNet50(
pretrained=False,
frozen=False,
features_output_stride=32,
upsampling_stack=upsampling.UpsamplingStack(
output_stride=4, refine_convs_filters=64),
)
x_in = tf.keras.layers.Input((160, 160, 1))
x, x_mid = resnet50.make_backbone(x_in)
with self.subTest("output shape"):
self.assertAllEqual(x.shape, (None, 160 // 4, 160 // 4, 64))
def test_resnet50_frozen(self):
resnet50 = resnet.ResNet50(pretrained=True,
frozen=True,
features_output_stride=16)
x_in = tf.keras.layers.Input((160, 160, 1))
x, x_mid = resnet50.make_backbone(x_in)
model = tf.keras.Model(x_in, x)
param_counts = [
np.prod(train_var.shape) for train_var in model.trainable_weights
]
with self.subTest("number of trainable weights"):
self.assertEqual(len(model.trainable_weights), 0)
with self.subTest("trainable parameter count"):
self.assertEqual(np.sum(param_counts), 0)
with self.subTest("total parameter count"):
self.assertEqual(model.count_params(), 23587712)
def test_resnet50(self):
resnet50 = resnet.ResNet50(pretrained=False, features_output_stride=32)
x_in = tf.keras.layers.Input((160, 160, 1))
x, x_mid = resnet50.make_backbone(x_in)
model = tf.keras.Model(x_in, x)
param_counts = [
np.prod(train_var.shape) for train_var in model.trainable_weights
]
with self.subTest("number of layers"):
self.assertEqual(len(model.layers), 175)
with self.subTest("number of trainable weights"):
self.assertEqual(len(model.trainable_weights), 212)
with self.subTest("trainable parameter count"):
self.assertEqual(np.sum(param_counts), 23528320)
with self.subTest("total parameter count"):
self.assertEqual(model.count_params(), 23581440)
with self.subTest("output shape"):
self.assertAllEqual(model.output.shape, (None, 5, 5, 2048))