def build_model(model_type, batch_size):
"""Returns a Keras model specified by name."""
if model_type == 'unet':
return u_net.get_model(input_shape=(512, 512, 3),
scales=4,
bottleneck_depth=1024,
bottleneck_layers=2)
elif model_type == 'can':
return vgg.build_can(input_shape=(512, 512, 3),
conv_channels=64,
out_channels=3)
else:
raise ValueError(model_type)
def test_one_scale(self):
model = u_net.get_model(input_shape=(64, 64, 3),
scales=1,
bottleneck_depth=128)
model.summary()
# Downscaling arm.
input_layer = model.get_layer('input')
down_conv1 = model.get_layer('down64_conv1')
down_conv2 = model.get_layer('down64_conv2')
down_pool = model.get_layer('down64_pool')
bottleneck_conv1 = model.get_layer('bottleneck_conv1')
self.assertIs(input_layer.output, down_conv1.input)
self.assertIs(down_conv1.output, down_conv2.input)
self.assertIs(down_conv2.output, down_pool.input)
self.assertIs(down_pool.output, bottleneck_conv1.input)
self.assertAllEqual(model.input_shape, [None, 64, 64, 3])
self.assertAllEqual(down_conv1.output_shape, [None, 64, 64, 64])
self.assertAllEqual(down_conv2.output_shape, [None, 64, 64, 64])
self.assertAllEqual(down_pool.output_shape, [None, 32, 32, 64])
self.assertAllEqual(bottleneck_conv1.output_shape, [None, 32, 32, 128])
# Upscaling arm.
bottleneck_conv2 = model.get_layer('bottleneck_conv2')
up_2x = model.get_layer('up64_2x')
up_2xconv = model.get_layer('up64_2xconv')
up_concat = model.get_layer('up64_concat')
up_conv1 = model.get_layer('up64_conv1')
up_conv2 = model.get_layer('up64_conv2')
output_layer = model.get_layer('output')
self.assertIs(bottleneck_conv2.output, up_2x.input)
self.assertIs(up_2x.output, up_2xconv.input)
self.assertIs(up_2xconv.output, up_concat.input[0])
self.assertIs(up_concat.output, up_conv1.input)
self.assertIs(up_conv1.output, up_conv2.input)
self.assertIs(up_conv2.output, output_layer.input)
self.assertAllEqual(bottleneck_conv2.output_shape, [None, 32, 32, 128])
self.assertAllEqual(up_2x.output_shape, [None, 64, 64, 128])
self.assertAllEqual(up_2xconv.output_shape, [None, 64, 64, 64])
self.assertAllEqual(up_concat.output_shape, [None, 64, 64, 128])
self.assertAllEqual(up_conv1.output_shape, [None, 64, 64, 64])
self.assertAllEqual(up_conv2.output_shape, [None, 64, 64, 64])
self.assertAllEqual(output_layer.output_shape, [None, 64, 64, 3])
# Skip connection.
self.assertIs(down_conv2.output, up_concat.input[1])
def test_zero_scale(self):
model = u_net.get_model(input_shape=(128, 128, 1),
scales=0,
bottleneck_depth=32)
model.summary()
input_layer = model.get_layer('input')
bottleneck_conv1 = model.get_layer('bottleneck_conv1')
bottleneck_conv2 = model.get_layer('bottleneck_conv2')
output_layer = model.get_layer('output')
self.assertIs(input_layer.output, bottleneck_conv1.input)
self.assertIs(bottleneck_conv1.output, bottleneck_conv2.input)
self.assertIs(bottleneck_conv2.output, output_layer.input)
self.assertAllEqual(model.input_shape, [None, 128, 128, 1])
self.assertAllEqual(bottleneck_conv1.output_shape,
[None, 128, 128, 32])
self.assertAllEqual(bottleneck_conv2.output_shape,
[None, 128, 128, 32])
self.assertAllEqual(model.output_shape, [None, 128, 128, 1])