def test_convolutional_transpose():
x = tensor.tensor4('x')
num_channels = 4
num_filters = 3
image_size = (8, 6)
original_image_size = (17, 13)
batch_size = 5
filter_size = (3, 3)
step = (2, 2)
conv = ConvolutionalTranspose(filter_size,
num_filters,
num_channels,
step=step,
original_image_size=original_image_size,
image_size=image_size,
weights_init=Constant(1.),
biases_init=Constant(5.))
conv.initialize()
y = conv.apply(x)
func = function([x], y)
x_val = numpy.ones((batch_size, num_channels) + image_size,
dtype=theano.config.floatX)
expected_value = num_channels * numpy.ones((batch_size, num_filters) +
original_image_size)
expected_value[:, :, 2:-2:2, :] += num_channels
expected_value[:, :, :, 2:-2:2] += num_channels
expected_value[:, :, 2:-2:2, 2:-2:2] += num_channels
assert_allclose(func(x_val), expected_value + 5)
def test_convolutional_transpose():
x = tensor.tensor4('x')
num_channels = 4
num_filters = 3
image_size = (8, 6)
original_image_size = (17, 13)
batch_size = 5
filter_size = (3, 3)
step = (2, 2)
conv = ConvolutionalTranspose(
filter_size, num_filters, num_channels, step=step,
original_image_size=original_image_size,
image_size=image_size, weights_init=Constant(1.),
biases_init=Constant(5.))
conv.initialize()
y = conv.apply(x)
func = function([x], y)
x_val = numpy.ones((batch_size, num_channels) + image_size,
dtype=theano.config.floatX)
expected_value = num_channels * numpy.ones(
(batch_size, num_filters) + original_image_size)
expected_value[:, :, 2:-2:2, :] += num_channels
expected_value[:, :, :, 2:-2:2] += num_channels
expected_value[:, :, 2:-2:2, 2:-2:2] += num_channels
assert_allclose(func(x_val), expected_value + 5)
def test_convolutional_transpose_original_size_inference_full_padding():
brick = ConvolutionalTranspose(filter_size=(4, 5), num_filters=10,
num_channels=5, step=(3, 2),
border_mode='full',
image_size=(6, 9))
brick.allocate()
assert brick.original_image_size == (13, 13)
input_ = tensor.tensor4()
dummy = numpy.empty((4, 5, 6, 9), dtype=theano.config.floatX)
result = brick.apply(input_).eval({input_: dummy})
assert result.shape == (4, 10, 13, 13)
def test_convolutional_transpose_original_size_inference_unused_edge():
brick = ConvolutionalTranspose(filter_size=(3, 3), num_filters=10,
num_channels=5, step=(2, 2),
border_mode=(1, 1), image_size=(4, 4),
unused_edge=(1, 1))
brick.allocate()
assert brick.original_image_size == (8, 8)
input_ = tensor.tensor4()
dummy = numpy.empty((4, 5, 4, 4), dtype=theano.config.floatX)
result = brick.apply(input_).eval({input_: dummy})
assert result.shape == (4, 10, 8, 8)
def test_convolutional_transpose_original_size_inference():
brick = ConvolutionalTranspose(filter_size=(4, 5), num_filters=10,
num_channels=5, step=(3, 2),
image_size=(6, 9))
brick.allocate()
# In x: filter applied 6 times with a step of 3 and filter size of 4
# means 1 dangling pixel, total original image size of 6 * 3 + 1 == 19.
# In y: step of 2, applied 9 times, filter size of 5 means 3
# dangling pixels, so original is 2 * 9 + 3 == 21.
assert brick.original_image_size == (19, 21)
input_ = tensor.tensor4()
dummy = numpy.empty((4, 5, 6, 9), dtype=theano.config.floatX)
result = brick.apply(input_).eval({input_: dummy})
assert result.shape == (4, 10, 19, 21)
def conv_transpose_brick(filter_size, step, num_filters, border_mode='valid'):
"""Instantiates a ConvolutionalTranspose brick."""
return ConvolutionalTranspose(filter_size=(filter_size, filter_size),
step=(step, step),
border_mode=border_mode,
num_filters=num_filters,
name=name_generator())
def test_conv_transpose_exception():
brick = ConvolutionalTranspose(filter_size=(4, 5),
num_filters=10,
num_channels=5,
step=(3, 2),
tied_biases=True)
assert_raises(ValueError, brick.apply, tensor.tensor4())
def test_convolutional_transpose_original_size_inferred_conv_sequence():
brick = ConvolutionalTranspose(filter_size=(4, 5), num_filters=10,
step=(3, 2))
seq = ConvolutionalSequence([brick], num_channels=5, image_size=(6, 9))
try:
seq.allocate()
except Exception as e:
raise AssertionError('exception raised: {}: {}'.format(
e.__class__.__name__, e))
def create_model_bricks(z_dim, image_size, depth):
g_image_size = image_size
g_image_size2 = g_image_size / 2
g_image_size3 = g_image_size / 4
g_image_size4 = g_image_size / 8
g_image_size5 = g_image_size / 16
encoder_layers = []
if depth > 0:
encoder_layers = encoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv1'),
SpatialBatchNormalization(name='batch_norm1'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv2'),
SpatialBatchNormalization(name='batch_norm2'),
Rectifier(),
Convolutional(
filter_size=(2, 2), step=(2, 2), num_filters=32, name='conv3'),
SpatialBatchNormalization(name='batch_norm3'),
Rectifier()
]
if depth > 1:
encoder_layers = encoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv4'),
SpatialBatchNormalization(name='batch_norm4'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv5'),
SpatialBatchNormalization(name='batch_norm5'),
Rectifier(),
Convolutional(
filter_size=(2, 2), step=(2, 2), num_filters=64, name='conv6'),
SpatialBatchNormalization(name='batch_norm6'),
Rectifier()
]
if depth > 2:
encoder_layers = encoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv7'),
SpatialBatchNormalization(name='batch_norm7'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv8'),
SpatialBatchNormalization(name='batch_norm8'),
Rectifier(),
Convolutional(
filter_size=(2, 2), step=(2, 2), num_filters=128,
name='conv9'),
SpatialBatchNormalization(name='batch_norm9'),
Rectifier()
]
if depth > 3:
encoder_layers = encoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv10'),
SpatialBatchNormalization(name='batch_norm10'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv11'),
SpatialBatchNormalization(name='batch_norm11'),
Rectifier(),
Convolutional(filter_size=(2, 2),
step=(2, 2),
num_filters=256,
name='conv12'),
SpatialBatchNormalization(name='batch_norm12'),
Rectifier(),
]
if depth > 4:
encoder_layers = encoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=512,
name='conv13'),
SpatialBatchNormalization(name='batch_norm13'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=512,
name='conv14'),
SpatialBatchNormalization(name='batch_norm14'),
Rectifier(),
Convolutional(filter_size=(2, 2),
step=(2, 2),
num_filters=512,
name='conv15'),
SpatialBatchNormalization(name='batch_norm15'),
Rectifier()
]
decoder_layers = []
if depth > 4:
decoder_layers = decoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=512,
name='conv_n3'),
SpatialBatchNormalization(name='batch_norm_n3'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=512,
name='conv_n2'),
SpatialBatchNormalization(name='batch_norm_n2'),
Rectifier(),
ConvolutionalTranspose(
filter_size=(2, 2),
step=(2, 2),
original_image_size=(g_image_size5, g_image_size5),
num_filters=512,
name='conv_n1'),
SpatialBatchNormalization(name='batch_norm_n1'),
Rectifier()
]
if depth > 3:
decoder_layers = decoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv1'),
SpatialBatchNormalization(name='batch_norm1'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv2'),
SpatialBatchNormalization(name='batch_norm2'),
Rectifier(),
ConvolutionalTranspose(
filter_size=(2, 2),
step=(2, 2),
original_image_size=(g_image_size4, g_image_size4),
num_filters=256,
name='conv3'),
SpatialBatchNormalization(name='batch_norm3'),
Rectifier()
]
if depth > 2:
decoder_layers = decoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv4'),
SpatialBatchNormalization(name='batch_norm4'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv5'),
SpatialBatchNormalization(name='batch_norm5'),
Rectifier(),
ConvolutionalTranspose(
filter_size=(2, 2),
step=(2, 2),
original_image_size=(g_image_size3, g_image_size3),
num_filters=128,
name='conv6'),
SpatialBatchNormalization(name='batch_norm6'),
Rectifier()
]
if depth > 1:
decoder_layers = decoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv7'),
SpatialBatchNormalization(name='batch_norm7'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv8'),
SpatialBatchNormalization(name='batch_norm8'),
Rectifier(),
ConvolutionalTranspose(
filter_size=(2, 2),
step=(2, 2),
original_image_size=(g_image_size2, g_image_size2),
num_filters=64,
name='conv9'),
SpatialBatchNormalization(name='batch_norm9'),
Rectifier()
]
if depth > 0:
decoder_layers = decoder_layers + [
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv10'),
SpatialBatchNormalization(name='batch_norm10'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv11'),
SpatialBatchNormalization(name='batch_norm11'),
Rectifier(),
ConvolutionalTranspose(
filter_size=(2, 2),
step=(2, 2),
original_image_size=(g_image_size, g_image_size),
num_filters=32,
name='conv12'),
SpatialBatchNormalization(name='batch_norm12'),
Rectifier()
]
decoder_layers = decoder_layers + [
Convolutional(filter_size=(1, 1), num_filters=3, name='conv_out'),
Logistic()
]
print("creating model of depth {} with {} encoder and {} decoder layers".
format(depth, len(encoder_layers), len(decoder_layers)))
encoder_convnet = ConvolutionalSequence(
layers=encoder_layers,
num_channels=3,
image_size=(g_image_size, g_image_size),
use_bias=False,
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='encoder_convnet')
encoder_convnet.initialize()
encoder_filters = numpy.prod(encoder_convnet.get_dim('output'))
encoder_mlp = MLP(
dims=[encoder_filters, 1000, z_dim],
activations=[
Sequence([BatchNormalization(1000).apply,
Rectifier().apply],
name='activation1'),
Identity().apply
],
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='encoder_mlp')
encoder_mlp.initialize()
decoder_mlp = BatchNormalizedMLP(
activations=[Rectifier(), Rectifier()],
dims=[encoder_mlp.output_dim // 2, 1000, encoder_filters],
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='decoder_mlp')
decoder_mlp.initialize()
decoder_convnet = ConvolutionalSequence(
layers=decoder_layers,
num_channels=encoder_convnet.get_dim('output')[0],
image_size=encoder_convnet.get_dim('output')[1:],
use_bias=False,
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='decoder_convnet')
decoder_convnet.initialize()
return encoder_convnet, encoder_mlp, decoder_convnet, decoder_mlp
def create_model_bricks():
encoder_convnet = ConvolutionalSequence(
layers=[
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv1'),
SpatialBatchNormalization(name='batch_norm1'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv2'),
SpatialBatchNormalization(name='batch_norm2'),
Rectifier(),
Convolutional(filter_size=(2, 2),
step=(2, 2),
num_filters=32,
name='conv3'),
SpatialBatchNormalization(name='batch_norm3'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv4'),
SpatialBatchNormalization(name='batch_norm4'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv5'),
SpatialBatchNormalization(name='batch_norm5'),
Rectifier(),
Convolutional(filter_size=(2, 2),
step=(2, 2),
num_filters=64,
name='conv6'),
SpatialBatchNormalization(name='batch_norm6'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv7'),
SpatialBatchNormalization(name='batch_norm7'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv8'),
SpatialBatchNormalization(name='batch_norm8'),
Rectifier(),
Convolutional(filter_size=(2, 2),
step=(2, 2),
num_filters=128,
name='conv9'),
SpatialBatchNormalization(name='batch_norm9'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv10'),
SpatialBatchNormalization(name='batch_norm10'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv11'),
SpatialBatchNormalization(name='batch_norm11'),
Rectifier(),
Convolutional(filter_size=(2, 2),
step=(2, 2),
num_filters=256,
name='conv12'),
SpatialBatchNormalization(name='batch_norm12'),
Rectifier(),
],
num_channels=3,
image_size=(64, 64),
use_bias=False,
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='encoder_convnet')
encoder_convnet.initialize()
encoder_filters = numpy.prod(encoder_convnet.get_dim('output'))
encoder_mlp = MLP(
dims=[encoder_filters, 1000, 1000],
activations=[
Sequence([BatchNormalization(1000).apply,
Rectifier().apply],
name='activation1'),
Identity().apply
],
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='encoder_mlp')
encoder_mlp.initialize()
decoder_mlp = BatchNormalizedMLP(
activations=[Rectifier(), Rectifier()],
dims=[encoder_mlp.output_dim // 2, 1000, encoder_filters],
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='decoder_mlp')
decoder_mlp.initialize()
decoder_convnet = ConvolutionalSequence(
layers=[
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv1'),
SpatialBatchNormalization(name='batch_norm1'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=256,
name='conv2'),
SpatialBatchNormalization(name='batch_norm2'),
Rectifier(),
ConvolutionalTranspose(filter_size=(2, 2),
step=(2, 2),
original_image_size=(8, 8),
num_filters=256,
name='conv3'),
SpatialBatchNormalization(name='batch_norm3'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv4'),
SpatialBatchNormalization(name='batch_norm4'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=128,
name='conv5'),
SpatialBatchNormalization(name='batch_norm5'),
Rectifier(),
ConvolutionalTranspose(filter_size=(2, 2),
step=(2, 2),
original_image_size=(16, 16),
num_filters=128,
name='conv6'),
SpatialBatchNormalization(name='batch_norm6'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv7'),
SpatialBatchNormalization(name='batch_norm7'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=64,
name='conv8'),
SpatialBatchNormalization(name='batch_norm8'),
Rectifier(),
ConvolutionalTranspose(filter_size=(2, 2),
step=(2, 2),
original_image_size=(32, 32),
num_filters=64,
name='conv9'),
SpatialBatchNormalization(name='batch_norm9'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv10'),
SpatialBatchNormalization(name='batch_norm10'),
Rectifier(),
Convolutional(filter_size=(3, 3),
border_mode=(1, 1),
num_filters=32,
name='conv11'),
SpatialBatchNormalization(name='batch_norm11'),
Rectifier(),
ConvolutionalTranspose(filter_size=(2, 2),
step=(2, 2),
original_image_size=(64, 64),
num_filters=32,
name='conv12'),
SpatialBatchNormalization(name='batch_norm12'),
Rectifier(),
Convolutional(filter_size=(1, 1), num_filters=3, name='conv_out'),
Logistic(),
],
num_channels=encoder_convnet.get_dim('output')[0],
image_size=encoder_convnet.get_dim('output')[1:],
use_bias=False,
weights_init=IsotropicGaussian(0.033),
biases_init=Constant(0),
name='decoder_convnet')
decoder_convnet.initialize()
return encoder_convnet, encoder_mlp, decoder_convnet, decoder_mlp
