def cascade_resnet(pr,
net,
input_layer,
n=5,
nf=64,
b=lasagne.init.Constant,
**kwargs):
shape = lasagne.layers.get_output_shape(input_layer)
n_channel = shape[1]
net[pr + 'conv1'] = l.Conv(input_layer, nf, 3, b=b(), name=pr + 'conv1')
for i in range(2, n):
net[pr + 'conv%d' % i] = l.Conv(net[pr + 'conv%d' % (i - 1)],
nf,
3,
b=b(),
name=pr + 'conv%d' % i)
net[pr + 'conv_aggr'] = l.ConvAggr(net[pr + 'conv%d' % (n - 1)],
n_channel,
3,
b=b(),
name=pr + 'conv_aggr')
net[pr + 'res'] = l.ResidualLayer([net[pr + 'conv_aggr'], input_layer],
name=pr + 'res')
output_layer = net[pr + 'res']
return net, output_layer
def cascade_resnet_3d_avg(pr, net, input_layer, n=5, nf=64,
b=lasagne.init.Constant, frame_dist=range(5),
**kwargs):
shape = lasagne.layers.get_output_shape(input_layer)
n_channel = shape[1]
divide_by_n = kwargs['cascade_i'] != 0
k = (3, 3, 3)
# Data sharing layer
net[pr+'kavg'] = l.AverageInKspaceLayer([input_layer, net['mask']],
shape,
frame_dist=frame_dist,
divide_by_n=divide_by_n,
clipped=False)
# Conv layers
net[pr+'conv1'] = l.Conv3D(net[pr+'kavg'], nf, k, b=b(), name=pr+'conv1')
for i in xrange(2, n):
net[pr+'conv%d'%i] = l.Conv3D(net[pr+'conv%d'%(i-1)], nf, k, b=b(),
name=pr+'conv%d'%i)
net[pr+'conv_aggr'] = l.Conv3DAggr(net[pr+'conv%d'%(n-1)], n_channel, k,
b=b(), name=pr+'conv_aggr')
net[pr+'res'] = l.ResidualLayer([net[pr+'conv_aggr'], input_layer],
name=pr+'res')
output_layer = net[pr+'res']
return net, output_layer
def cascade_Unet(pr,
net,
input_layer,
n=5,
nf=64,
b=lasagne.init.Constant,
**kwargs):
shape = lasagne.layers.get_output_shape(input_layer)
n_channel = shape[1]
net[pr + 'conv1'] = l.Conv(input_layer, nf, 3, b=b(), name=pr + 'conv1')
#changed this part from cascade_resnet
# solved: (contracting path -> sizes need to be equal), (padding problems -> need to have same size in the end -> stride and padding parameters do not change problem)
net[pr + 'conv2'] = l.Conv(net[pr + 'conv1'],
nf,
3,
b=b(),
name=pr + 'conv2')
net[pr + 'max1'] = l.MaxPool(net[pr + 'conv2'],
2,
stride=2,
pad=(0, 0),
name=pr + 'max1')
net[pr + 'conv3'] = l.Conv(net[pr + 'max1'],
nf * 2,
3,
b=b(),
name=pr + 'conv3')
net[pr + 'conv4'] = l.Conv(net[pr + 'conv3'],
nf * 2,
3,
b=b(),
name=pr + 'conv4')
net[pr + 'max2'] = l.MaxPool(net[pr + 'conv4'],
2,
stride=2,
pad=(1, 1),
name=pr + 'max2')
net[pr + 'conv5'] = l.Conv(net[pr + 'max2'],
nf * 4,
3,
b=b(),
name=pr + 'conv5')
net[pr + 'conv6'] = l.Conv(net[pr + 'conv5'],
nf * 4,
3,
b=b(),
name=pr + 'conv6')
net[pr + 'deconv1'] = l.Deconv(net[pr + 'conv6'],
nf * 2,
2,
stride=2,
crop='full',
b=b(),
name=pr + 'deconv1')
#shape1 = lasagne.layers.get_output_shape(net[pr+'conv3'])
#shape2 = lasagne.layers.get_output_shape(net[pr+'deconv1'])
#print(shape1)
#print(shape2)
#additional input conv4 (contracting path)
net[pr + 'concat1'] = lasagne.layers.concat(
[net[pr + 'deconv1'], net[pr + 'conv4']])
net[pr + 'conv7'] = l.Conv(net[pr + 'concat1'],
nf * 2,
3,
b=b(),
name=pr + 'conv7')
net[pr + 'conv8'] = l.Conv(net[pr + 'conv7'],
nf * 2,
3,
b=b(),
name=pr + 'conv8')
net[pr + 'deconv2'] = l.Deconv(net[pr + 'conv8'],
nf,
2,
stride=2,
crop='valid',
b=b(),
name=pr + 'deconv2')
#shape3 = lasagne.layers.get_output_shape(net[pr+'conv2'])
#shape4 = lasagne.layers.get_output_shape(net[pr+'deconv2'])
#print(shape3)
#print(shape4)
#additional input conv2 (contracting path)
net[pr + 'concat2'] = lasagne.layers.concat(
[net[pr + 'deconv2'], net[pr + 'conv2']])
net[pr + 'conv9'] = l.Conv(net[pr + 'concat2'],
nf,
3,
b=b(),
name=pr + 'conv9')
net[pr + 'conv10'] = l.Conv(net[pr + 'conv9'],
nf,
3,
b=b(),
name=pr + 'conv10')
net[pr + 'conv_aggr'] = l.ConvAggr(net[pr + 'conv10'],
n_channel,
3,
pad='same',
b=b(),
name=pr + 'conv_aggr')
net[pr + 'res'] = l.ResidualLayer([net[pr + 'conv_aggr'], input_layer],
name=pr + 'res')
output_layer = lasagne.layers.reshape(net[pr + 'res'], shape)
return net, output_layer