def cin_resnet_block(input_shape, z_dim, padding, norm_layer, use_dropout,
use_bias, block_name, **kwargs):
num_filters = input_shape[1]
block = nn.Sequential(input_shape=input_shape, layer_name=block_name)
block.append(
nn.Conv2DLayer(input_shape,
num_filters,
3,
border_mode=padding,
no_bias=not use_bias,
activation=None,
layer_name=block_name + '/conv1'))
block.append(norm_layer(input_shape, z_dim,
layer_name=block_name + '/CIN'))
block.append(
nn.ActivationLayer(block.output_shape,
layer_name=block_name + '/relu1'))
if use_dropout:
block.append(
nn.DropoutLayer(block.output_shape,
.5,
layer_name=block_name + '/dropout'))
block.append(
nn.Conv2DLayer(block.output_shape,
num_filters,
3,
border_mode=padding,
no_bias=not use_bias,
activation=None,
layer_name=block_name + '/conv2'))
block.append(nn.InstanceNormLayer(block.output_shape, block_name + '/IN'))
return block
def _build_simple_conv_block(input_shape, num_filters, stride, border_mode,
activation, block_name, **kwargs):
block = [
nn.Conv2DLayer(input_shape,
num_filters,
3,
stride=stride,
border_mode=border_mode,
layer_name=block_name + '/conv1',
no_bias=False,
activation='linear')
]
block.append(
nn.BatchNormLayer(block[-1].output_shape,
activation=activation,
layer_name=block_name + '/conv1_bn'))
block.append(
nn.Conv2DLayer(block[-1].output_shape,
num_filters,
3,
border_mode=border_mode,
no_bias=True,
layer_name=block_name + '/conv2',
activation='linear'))
block.append(
nn.BatchNormLayer(block[-1].output_shape,
layer_name=block_name + '/conv2_bn',
activation='linear'))
return block
def prep(x):
conv = nn.Conv2DLayer((1, 3, 224, 224),
3,
1,
no_bias=False,
activation='linear',
filter_flip=False,
border_mode='valid')
kern = np.array([[0, 0, 255], [0, 255, 0], [255, 0, 0]],
'float32')[:, :, None, None]
conv.W.set_value(kern)
conv.b.set_value(np.array([-103.939, -116.779, -123.68], 'float32'))
return conv(x)
def test_partial_conv_based_padding():
shape = (1, 3, 5, 5)
num_filters = 2
filter_size = 3
x = T.ones(shape)
conv = nn.Conv2DLayer(shape,
num_filters,
filter_size,
border_mode='partial')
y = conv(x)
f = nn.function([], y)
print(f())
print(conv.update_mask.eval())
print(conv.mask_ratio)
def __init__(self, input_shape, fc=True, num_classes=1000, name='vgg19'):
super(VGG19, self).__init__(input_shape=input_shape, layer_name=name)
self.fc = fc
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
no_bias=False,
layer_name=name + '_conv1_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
no_bias=False,
layer_name=name + '_conv1_2'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool0'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
no_bias=False,
layer_name=name + '_conv2_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
no_bias=False,
layer_name=name + '_conv2_2'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool1'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
no_bias=False,
layer_name=name + '_conv3_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
no_bias=False,
layer_name=name + '_conv3_2'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
no_bias=False,
layer_name=name + '_conv3_3'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
no_bias=False,
layer_name=name + '_conv3_4'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool2'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv4_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv4_2'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv4_3'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv4_4'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool3'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv5_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv5_2'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv5_3'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
no_bias=False,
layer_name=name + '_conv5_4'))
if fc:
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool4'))
self.append(
nn.FCLayer(self.output_shape, 4096, layer_name=name + '_fc1'))
self.append(
nn.FCLayer(self.output_shape, 4096, layer_name=name + '_fc2'))
self.append(
nn.SoftmaxLayer(self.output_shape, num_classes,
name + '_softmax'))
def __init__(self, input_shape, param_file, name='vgg19'):
super(VGG19, self).__init__(input_shape=input_shape, layer_name=name)
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv1_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv1_2'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '/maxpool1'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_2'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '/maxpool2'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_1'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_2'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_3'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_4'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '/maxpool3'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
border_mode='ref',
no_bias=False,
layer_name=name + '/conv4_1'))
self.load_params(param_file)
def __init__(self, input_shape, param_file, name='Decoder'):
super(Decoder, self).__init__(input_shape=input_shape, layer_name=name)
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv1_1',
activation='relu'))
self.append(
nn.UpsamplingLayer(self.output_shape,
2,
method='nearest',
layer_name=name + '/up1'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_1',
activation='relu'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_2',
activation='relu'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_3',
activation='relu'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_4',
activation='relu'))
self.append(
nn.UpsamplingLayer(self.output_shape,
2,
method='nearest',
layer_name=name + '/up2'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_1',
activation='relu'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_2',
activation='relu'))
self.append(
nn.UpsamplingLayer(self.output_shape,
2,
method='nearest',
layer_name=name + '/up3'))
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv4_1',
activation='relu'))
self.append(
nn.Conv2DLayer(self.output_shape,
3,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
activation='tanh',
layer_name=name + '/output'))
self.load_params(param_file)
def __init__(self,
input_shape,
n_latent,
num_filters,
norm_layer,
deterministic=False,
use_bias=False,
name='Latent Encoder'):
super(LatentEncoder, self).__init__(input_shape=input_shape,
layer_name=name)
self.deterministic = deterministic
self.enc = nn.Sequential(input_shape=input_shape,
layer_name=name + '/enc')
self.enc.append(
nn.Conv2DLayer(self.enc.output_shape,
num_filters,
3,
stride=2,
no_bias=False,
activation='relu',
layer_name=name + '/conv1'))
self.enc.append(
nn.Conv2DLayer(self.enc.output_shape,
2 * num_filters,
3,
stride=2,
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv2'))
self.enc.append(norm_layer(self.enc.output_shape, name + '/norm2'))
self.enc.append(
nn.ActivationLayer(self.enc.output_shape, 'relu', name + '/act2'))
self.enc.append(
nn.Conv2DLayer(self.enc.output_shape,
4 * num_filters,
3,
stride=2,
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv3'))
self.enc.append(norm_layer(self.enc.output_shape, name + '/norm3'))
self.enc.append(
nn.ActivationLayer(self.enc.output_shape, 'relu', name + '/act3'))
self.enc.append(
nn.Conv2DLayer(self.enc.output_shape,
8 * num_filters,
3,
stride=2,
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv4'))
self.enc.append(norm_layer(self.enc.output_shape, name + '/norm4'))
self.enc.append(
nn.ActivationLayer(self.enc.output_shape, 'relu', name + '/act4'))
self.enc.append(
nn.Conv2DLayer(self.enc.output_shape,
8 * num_filters,
4,
stride=1,
no_bias=not use_bias,
activation=None,
border_mode='valid',
layer_name=name + '/conv5'))
self.enc.append(norm_layer(self.enc.output_shape, name + '/norm5'))
self.enc.append(
nn.ActivationLayer(self.enc.output_shape, 'relu', name + '/act5'))
self.enc_mu = nn.Conv2DLayer(self.enc.output_shape,
n_latent,
1,
no_bias=False,
activation=None,
layer_name=name + '/mu')
self.extend((self.enc, self.enc_mu))
if not deterministic:
self.enc_logvar = nn.Conv2DLayer(self.enc.output_shape,
n_latent,
1,
no_bias=False,
activation=None,
layer_name=name + '/logvar')
self.append(self.enc_logvar)
def __init__(self,
input_shape,
num_filters=64,
norm_layer=partial(nn.BatchNormLayer, activation=None),
use_sigmoid=False,
use_bias=True,
name='Discriminator Edges'):
super(DiscriminatorEdges, self).__init__(input_shape=input_shape,
layer_name=name)
self.append(
nn.Conv2DLayer(self.output_shape,
num_filters,
3,
stride=2,
border_mode='half',
no_bias=not use_bias,
activation='lrelu',
alpha=.2,
layer_name=name + '/conv1'))
self.append(
nn.Conv2DLayer(self.output_shape,
2 * num_filters,
3,
stride=2,
border_mode='half',
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv2'))
self.append(norm_layer(self.output_shape, name + '/bn2'))
self.append(
nn.ActivationLayer(self.output_shape,
activation='lrelu',
alpha=.2,
layer_name=name + '/act2'))
self.append(
nn.Conv2DLayer(self.output_shape,
4 * num_filters,
3,
border_mode='half',
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv3'))
self.append(norm_layer(self.output_shape, layer_name=name + '/bn3'))
self.append(
nn.ActivationLayer(self.output_shape,
activation='lrelu',
alpha=.2,
layer_name=name + '/act3'))
self.append(
nn.Conv2DLayer(self.output_shape,
4 * num_filters,
3,
border_mode='half',
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv4'))
self.append(norm_layer(self.output_shape, layer_name=name + '/bn4'))
self.append(
nn.ActivationLayer(self.output_shape,
activation='lrelu',
alpha=.2,
layer_name=name + '/act4'))
self.append(
nn.Conv2DLayer(self.output_shape,
1,
4,
border_mode='valid',
layer_name=name + '/output',
activation=None))
if use_sigmoid:
self.append(
nn.ActivationLayer(self.output_shape, 'sigmoid',
name + '/sigmoid'))
def __init__(self,
input_shape,
n_latent,
num_filters=64,
norm_layer=nn.ConditionalInstanceNorm2DLayer,
use_sigmoid=False,
use_bias=True,
name='CIN Discriminator'):
super(CINDiscriminator, self).__init__(input_shape=input_shape,
layer_name=name)
self.append(
nn.Conv2DLayer(self.output_shape,
num_filters,
4,
stride=2,
border_mode=1,
no_bias=not use_bias,
activation='lrelu',
alpha=.2,
layer_name=name + '/conv1'))
self.append(
nn.Conv2DLayer(self.output_shape,
2 * num_filters,
4,
stride=2,
border_mode=1,
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv2'))
self.append(norm_layer(self.output_shape, n_latent, name + '/bn2'))
self.append(
nn.ActivationLayer(self.output_shape,
activation='lrelu',
alpha=.2,
layer_name=name + '/act2'))
self.append(
nn.Conv2DLayer(self.output_shape,
4 * num_filters,
4,
border_mode=1,
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv3'))
self.append(
norm_layer(self.output_shape, n_latent, layer_name=name + '/bn3'))
self.append(
nn.ActivationLayer(self.output_shape,
activation='lrelu',
alpha=.2,
layer_name=name + '/act3'))
self.append(
nn.Conv2DLayer(self.output_shape,
5 * num_filters,
4,
border_mode=1,
no_bias=not use_bias,
activation=None,
layer_name=name + '/conv4'))
self.append(
norm_layer(self.output_shape, n_latent, layer_name=name + '/bn4'))
self.append(
nn.ActivationLayer(self.output_shape,
activation='lrelu',
alpha=.2,
layer_name=name + '/act4'))
self.append(
nn.Conv2DLayer(self.output_shape,
1,
4,
border_mode=1,
layer_name=name + '/output',
activation=None))
if use_sigmoid:
self.append(
nn.ActivationLayer(self.output_shape, 'sigmoid',
name + '/sigmoid'))
def __init__(self,
input_shape,
num_filters,
output_dim,
norm_layer=partial(nn.InstanceNormLayer, activation=None),
use_dropout=False,
padding='ref',
name='Resnet Generator'):
super(ResnetGen, self).__init__(input_shape=input_shape,
layer_name=name)
self.append(
nn.Conv2DLayer(input_shape,
num_filters,
7,
border_mode='ref',
activation=None,
no_bias=False,
layer_name=name + '/conv1'))
self.append(norm_layer(self.output_shape, layer_name=name + '/cin1'))
self.append(
nn.ActivationLayer(self.output_shape,
'relu',
layer_name=name + '/relu1'))
self.append(
nn.Conv2DLayer(self.output_shape,
num_filters * 2,
3,
border_mode='half',
activation=None,
no_bias=False,
layer_name=name + '/conv2'))
self.append(norm_layer(self.output_shape, layer_name=name + '/cin2'))
self.append(
nn.ActivationLayer(self.output_shape,
'relu',
layer_name=name + '/relu2'))
self.append(
nn.Conv2DLayer(self.output_shape,
num_filters * 4,
3,
stride=2,
border_mode='half',
activation=None,
no_bias=False,
layer_name=name + '/conv3'))
self.append(norm_layer(self.output_shape, layer_name=name + '/cin3'))
self.append(
nn.ActivationLayer(self.output_shape,
'relu',
layer_name=name + '/relu3'))
for i in range(3):
self.append(
ResnetBlock(self.output_shape, padding, norm_layer,
use_dropout, True,
name + '/ResBlock %d' % (i + 1)))
self.append(
nn.TransposedConvolutionalLayer(self.output_shape,
2 * num_filters,
3,
stride=(2, 2),
padding='half',
activation=None,
layer_name=name + '/deconv'))
self.append(norm_layer(self.output_shape, layer_name=name + '/cin4'))
self.append(
nn.ActivationLayer(self.output_shape, 'relu', name + '/relu4'))
self.append(
nn.Conv2DLayer(self.output_shape,
num_filters,
3,
border_mode='half',
activation=None,
no_bias=False,
layer_name=name + '/conv5'))
self.append(norm_layer(self.output_shape, layer_name=name + '/cin5'))
self.append(
nn.ActivationLayer(self.output_shape,
'relu',
layer_name=name + '/relu5'))
self.append(
nn.Conv2DLayer(self.output_shape,
output_dim,
7,
activation=None,
layer_name=name + '/output'))
self.append(
nn.ActivationLayer(self.output_shape,
'tanh',
layer_name=name + '/output_act'))
def __init__(self,
input_shape,
fc=True,
bn=False,
dropout=True,
border_mode='half',
num_classes=1000,
name='vgg16'):
super(VGG16, self).__init__(input_shape=input_shape, layer_name=name)
self.fc = fc
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv1',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn1') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu1'))
self.append(
nn.Conv2DLayer(self.output_shape,
64,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv2',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn2') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu2'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool0'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv3',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn3') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu3'))
self.append(
nn.Conv2DLayer(self.output_shape,
128,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv4',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn4') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu4'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool1'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv5',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn5') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu5'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv6',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn6') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu6'))
self.append(
nn.Conv2DLayer(self.output_shape,
256,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv7',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn7') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu7'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool2'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv8',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn8') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu8'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv9',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn9') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu9'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv10',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn10') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu10'))
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool3'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv11',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn11') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu11'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv12',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn11') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu12'))
self.append(
nn.Conv2DLayer(self.output_shape,
512,
3,
nn.HeNormal('relu'),
activation=None,
no_bias=False,
layer_name=name + '/conv13',
border_mode=border_mode))
self.append(
nn.BatchNormLayer(self.output_shape, name + '/bn13') if bn else nn.
ActivationLayer(self.output_shape, layer_name=name + '/relu13'))
if fc:
self.append(
nn.MaxPoolingLayer(self.output_shape, (2, 2),
layer_name=name + '_maxpool4'))
self.append(
nn.FCLayer(self.output_shape, 4096, layer_name=name + '_fc1'))
if dropout:
self.append(
nn.DropoutLayer(self.output_shape,
drop_prob=.5,
layer_name=name + '/dropout1'))
self.append(
nn.FCLayer(self.output_shape, 4096, layer_name=name + '_fc2'))
if dropout:
self.append(
nn.DropoutLayer(self.output_shape,
drop_prob=.5,
layer_name=name + '/dropout2'))
self.append(
nn.SoftmaxLayer(self.output_shape, num_classes,
name + '_softmax'))
def __init__(self, encoder, layer, name='Decoder'):
super(Decoder,
self).__init__(input_shape=encoder[layer - 1].output_shape,
layer_name=name)
self.enc = encoder
self.layer = layer
dec = nn.Sequential(input_shape=encoder.output_shape, layer_name=name)
dec.append(
nn.Conv2DLayer(dec.output_shape,
512,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv1_1'))
dec.append(
nn.UpsamplingLayer(dec.output_shape,
2,
method='nearest',
layer_name=name + '/up1'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
512,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_1'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
512,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_2'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
512,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_3'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
512,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv2_4'))
dec.append(
nn.UpsamplingLayer(dec.output_shape,
2,
method='nearest',
layer_name=name + '/up2'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_1'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_2'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_3'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
256,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv3_4'))
dec.append(
nn.UpsamplingLayer(dec.output_shape,
2,
method='nearest',
layer_name=name + '/up3'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
128,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv4_1'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
128,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv4_2'))
dec.append(
nn.UpsamplingLayer(dec.output_shape,
2,
method='nearest',
layer_name=name + '/up4'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
64,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv5_1'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
64,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
layer_name=name + '/conv5_2'))
dec.append(
nn.Conv2DLayer(dec.output_shape,
3,
3,
init=nn.GlorotUniform(),
border_mode='ref',
no_bias=False,
activation='tanh',
layer_name=name + '/output'))
self.append(dec[len(encoder) - layer:])