def test_upsampling_stack_upsampling_add(self):
upsampling_stack = upsampling.UpsamplingStack(
output_stride=2,
upsampling_stride=2,
skip_add=True,
transposed_conv=True,
transposed_conv_filters=16,
refine_convs=0,
)
skip_sources = [
upsampling.IntermediateFeature(tensor=tf.keras.Input((16, 16, 1)),
stride=8),
upsampling.IntermediateFeature(tensor=tf.keras.Input((32, 32, 2)),
stride=4),
]
x, intermediate_feats = upsampling_stack.make_stack(
tf.keras.Input((8, 8, 32)),
current_stride=16,
skip_sources=skip_sources)
model = tf.keras.Model(tf.keras.utils.get_source_inputs(x), x)
self.assertAllEqual(x.shape, (None, 64, 64, 16))
self.assertEqual(len(intermediate_feats), 4)
self.assertAllEqual(
model.get_layer("upsample_s16_to_s8_skip_conv1x1").output.shape,
(None, 16, 16, 16),
)
self.assertAllEqual(
model.get_layer("upsample_s8_to_s4_skip_conv1x1").output.shape,
(None, 32, 32, 16),
)
self.assertIsInstance(model.get_layer("upsample_s16_to_s8_skip_add"),
tf.keras.layers.Add)
def test_upsampling_stack_refine_convs_filter_rate(self):
upsampling_stack = upsampling.UpsamplingStack(
output_stride=2,
upsampling_stride=2,
transposed_conv=False,
refine_convs=2,
refine_convs_filters=16,
refine_convs_filters_rate=2,
)
x, intermediate_feats = upsampling_stack.make_stack(tf.keras.Input(
(4, 4, 2)),
current_stride=16)
model = tf.keras.Model(tf.keras.utils.get_source_inputs(x), x)
self.assertEqual(
model.get_layer("upsample_s16_to_s8_refine0_conv").filters, 16)
self.assertEqual(
model.get_layer("upsample_s16_to_s8_refine1_conv").filters, 16)
self.assertEqual(
model.get_layer("upsample_s8_to_s4_refine0_conv").filters, 32)
self.assertEqual(
model.get_layer("upsample_s8_to_s4_refine1_conv").filters, 32)
self.assertEqual(
model.get_layer("upsample_s4_to_s2_refine0_conv").filters, 64)
self.assertEqual(
model.get_layer("upsample_s4_to_s2_refine1_conv").filters, 64)
self.assertAllEqual(x.shape, (None, 32, 32, 64))
def test_upsampling_stack_upsampling_skip(self):
upsampling_stack = upsampling.UpsamplingStack(
output_stride=2,
upsampling_stride=2,
skip_add=False,
transposed_conv=True,
transposed_conv_filters=16,
refine_convs=0,
)
skip_sources = [
upsampling.IntermediateFeature(tensor=tf.keras.Input((16, 16, 1)),
stride=8),
upsampling.IntermediateFeature(tensor=tf.keras.Input((32, 32, 2)),
stride=4),
]
x, intermediate_feats = upsampling_stack.make_stack(
tf.keras.Input((8, 8, 32)),
current_stride=16,
skip_sources=skip_sources)
model = tf.keras.Model(tf.keras.utils.get_source_inputs(x), x)
self.assertAllEqual(x.shape, (None, 64, 64, 16))
self.assertEqual(len(intermediate_feats), 4)
self.assertIsInstance(model.layers[1], tf.keras.layers.Conv2DTranspose)
self.assertIsInstance(model.layers[2],
tf.keras.layers.BatchNormalization)
self.assertIsInstance(model.layers[4], tf.keras.layers.Activation)
self.assertIsInstance(model.layers[5], tf.keras.layers.Concatenate)
self.assertAllEqual(model.layers[5].output.shape, (None, 16, 16, 17))
self.assertIsInstance(model.layers[10], tf.keras.layers.Concatenate)
self.assertAllEqual(model.layers[10].output.shape, (None, 32, 32, 18))
def test_upsampling_stack_upsampling_stride4(self):
upsampling_stack = upsampling.UpsamplingStack(output_stride=4,
upsampling_stride=4)
x, intermediate_feats = upsampling_stack.make_stack(tf.keras.Input(
(8, 8, 32)),
current_stride=16)
self.assertAllEqual(x.shape, (None, 32, 32, 64))
self.assertEqual(len(intermediate_feats), 2)
def test_upsampling_stack_upsampling_interp(self):
upsampling_stack = upsampling.UpsamplingStack(output_stride=8,
upsampling_stride=2,
transposed_conv=False)
x, intermediate_feats = upsampling_stack.make_stack(tf.keras.Input(
(8, 8, 32)),
current_stride=16)
self.assertAllEqual(x.shape, (None, 16, 16, 64))
model = tf.keras.Model(tf.keras.utils.get_source_inputs(x), x)
self.assertIsInstance(model.layers[1], tf.keras.layers.UpSampling2D)
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_upsampling_stack(self):
upsampling_stack = upsampling.UpsamplingStack(
output_stride=4,
upsampling_stride=2,
transposed_conv=True,
transposed_conv_batchnorm=True,
refine_convs=1,
refine_convs_batchnorm=True,
)
x, intermediate_feats = upsampling_stack.make_stack(tf.keras.Input(
(8, 8, 32)),
current_stride=16)
model = tf.keras.Model(tf.keras.utils.get_source_inputs(x), x)
self.assertAllEqual(x.shape, (None, 32, 32, 64))
self.assertEqual(len(intermediate_feats), 2)
self.assertEqual(intermediate_feats[0].stride, 8)
self.assertEqual(intermediate_feats[1].stride, 4)
self.assertEqual(len(model.layers), 13)
self.assertIsInstance(model.layers[1], tf.keras.layers.Conv2DTranspose)