def buildSectorNet(self):
sectorNet = InputLayer(self.inputShape, self.inputVar)
for i, layer in enumerate(self.layerCategory):
self.logger.debug('Build {}th conv layer'.format(i))
self.logger.debug('The output shape of {}th layer equal {}'.format(
i - 1, get_output_shape(sectorNet)))
kernelXDim = int(layer[-1])
kernelDim = (kernelXDim, ) * 3
conv3D = batch_norm(
Conv3DLayer(incoming=sectorNet,
num_filters=self.numOfFMs[i],
filter_size=kernelDim,
W=HeUniform(gain='relu'),
nonlinearity=rectify,
name='Conv3D'))
self.logger.debug(
'The shape of {}th conv3D layer equals {}'.format(
i, get_output_shape(conv3D)))
sectorNet = ConcatLayer(
[conv3D, sectorNet],
1,
cropping=['center', 'None', 'center', 'center', 'center'])
self.logger.debug(
'The shape of {}th concat layer equals {}'.format(
i, get_output_shape(sectorNet)))
assert get_output_shape(sectorNet) == (None, sum(self.numOfFMs) + 1, 1,
1, 1)
sectorNet = batch_norm(
Conv3DLayer(incoming=sectorNet,
num_filters=2,
filter_size=(1, 1, 1),
W=HeUniform(gain='relu')))
self.logger.debug('The shape of last con3D layer equals {}'.format(
get_output_shape(sectorNet)))
sectorNet = ReshapeLayer(sectorNet, ([0], -1))
self.logger.debug('The shape of ReshapeLayer equals {}'.format(
get_output_shape(sectorNet)))
sectorNet = NonlinearityLayer(sectorNet, softmax)
self.logger.debug(
'The shape of output layer, i.e. NonlinearityLayer, equals {}'.
format(get_output_shape(sectorNet)))
assert get_output_shape(sectorNet) == (None, self.numOfOutputClass)
return sectorNet
def __init__(self, input_shape=(None, 1, 33, 33, 33)):
self.cubeSize = input_shape[-1]
# Theano variables
self.input_var = T.tensor5('input_var') # input image
self.target_var = T.ivector('target_var') # target
self.logger = logging.getLogger(__name__)
input_layer = InputLayer(input_shape, self.input_var)
self.logger.info('The shape of input layer is {}'.format(
get_output_shape(input_layer)))
hidden_layer1 = Conv3DLayer(incoming=input_layer,
num_filters=16,
filter_size=(3, 3, 3),
W=HeUniform(gain='relu'),
nonlinearity=rectify)
self.logger.info('The shape of first hidden layer is {}'.format(
get_output_shape(hidden_layer1)))
hidden_layer2 = Conv3DLayer(incoming=hidden_layer1,
num_filters=32,
filter_size=(3, 3, 3),
W=HeUniform(gain='relu'),
nonlinearity=rectify)
self.logger.info('The shape of second hidden layer is {}'.format(
get_output_shape(hidden_layer2)))
hidden_layer3 = Conv3DLayer(incoming=hidden_layer2,
num_filters=2,
filter_size=(1, 1, 1),
W=HeUniform(gain='relu'),
nonlinearity=rectify)
self.logger.info('The shape of third hidden layer is {}'.format(
get_output_shape(hidden_layer3)))
shuffledLayer = DimshuffleLayer(hidden_layer3, (0, 2, 3, 4, 1))
self.logger.info('The shape of shuffled layer is {}'.format(
get_output_shape(shuffledLayer)))
reshapedLayer = ReshapeLayer(shuffledLayer, ([0], -1))
self.logger.info('The shape of reshaped layer is {}'.format(
get_output_shape(reshapedLayer)))
self.output_layer = NonlinearityLayer(reshapedLayer, softmax)
self.logger.info('The shape of output layer is {}'.format(
get_output_shape(self.output_layer)))
def lrelu_conv(l_in, feat_out, stride=1, filter_size=3):
l = NonlinearityLayer(l_in, nonlin)
return Conv3DLayer(l,
feat_out,
filter_size,
stride,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
def norm_lrelu_conv(l_in, feat_out, stride=1, filter_size=3):
if do_norm:
l_in = BatchNormLayer(l_in, axes=axes)
l = NonlinearityLayer(l_in, nonlin)
return Conv3DLayer(l,
feat_out,
filter_size,
stride,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
def conv_norm_lrelu(l_in, feat_out):
l = Conv3DLayer(l_in,
feat_out,
3,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
if do_norm:
l = BatchNormLayer(l, axes=axes)
return NonlinearityLayer(l, nonlin)
def norm_lrelu_upscale_conv_norm_lrelu(l_in, feat_out):
if do_norm:
l_in = BatchNormLayer(l_in, axes=axes)
l = NonlinearityLayer(l_in, nonlin)
l = Upscale3DLayer(l, 2)
l = Conv3DLayer(l,
feat_out,
3,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = NonlinearityLayer(l, nonlin)
return l
from lasagne.layers import *
import pickle
import theano.misc.pkl_utils
cachefile = os.path.dirname(os.path.realpath(__file__)) + "/model6tissues.pkl"
if not os.path.exists(cachefile):
l = InputLayer(shape=(None, 1, 64, 64, 64), name="input")
l_input = l
l = Conv3DLayer(l,
num_filters=16,
filter_size=(3, 3, 3),
pad='same',
name="conv",
nonlinearity=elu)
l = Conv3DLayer(l,
num_filters=16,
filter_size=(3, 3, 3),
pad='same',
name="conv",
nonlinearity=elu)
l = batch_norm(l)
li0 = l
l = MaxPool3DLayer(l, pool_size=2, name='maxpool')
l = Conv3DLayer(l,
num_filters=24,
filter_size=(3, 3, 3),
def cascade_model(options):
"""
3D cascade model using Nolearn and Lasagne
Inputs:
- model_options:
- weights_path: path to where weights should be saved
Output:
- nets = list of NeuralNets (CNN1, CNN2)
"""
# model options
channels = len(options['modalities'])
train_split_perc = options['train_split']
num_epochs = options['max_epochs']
max_epochs_patience = options['patience']
# save model to disk to re-use it. Create an experiment folder
# organize experiment
if not os.path.exists(
os.path.join(options['weight_paths'], options['experiment'])):
os.mkdir(os.path.join(options['weight_paths'], options['experiment']))
if not os.path.exists(
os.path.join(options['weight_paths'], options['experiment'],
'nets')):
os.mkdir(
os.path.join(options['weight_paths'], options['experiment'],
'nets'))
# --------------------------------------------------
# first model
# --------------------------------------------------
layer1 = InputLayer(name='in1',
shape=(None, channels) + options['patch_size'])
layer1 = batch_norm(Conv3DLayer(layer1,
name='conv1_1',
num_filters=32,
filter_size=3,
pad='same'),
name='BN1')
layer1 = Pool3DLayer(layer1,
name='avgpool_1',
mode='max',
pool_size=2,
stride=2)
layer1 = batch_norm(Conv3DLayer(layer1,
name='conv2_1',
num_filters=64,
filter_size=3,
pad='same'),
name='BN2')
layer1 = Pool3DLayer(layer1,
name='avgpoo2_1',
mode='max',
pool_size=2,
stride=2)
layer1 = DropoutLayer(layer1, name='l2drop', p=0.5)
layer1 = DenseLayer(layer1, name='d_1', num_units=256)
layer1 = DenseLayer(layer1,
name='out',
num_units=2,
nonlinearity=nonlinearities.softmax)
# save weights
net_model = 'model_1'
net_weights = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '.pkl')
net_history = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '_history.pkl')
net1 = NeuralNet(
layers=layer1,
objective_loss_function=objectives.categorical_crossentropy,
batch_iterator_train=Rotate_batch_Iterator(batch_size=128),
update=updates.adadelta,
on_epoch_finished=[
SaveWeights(net_weights, only_best=True, pickle=False),
SaveTrainingHistory(net_history),
EarlyStopping(patience=max_epochs_patience)
],
verbose=options['net_verbose'],
max_epochs=num_epochs,
train_split=TrainSplit(eval_size=train_split_perc),
)
# --------------------------------------------------
# second model
# --------------------------------------------------
layer2 = InputLayer(name='in2',
shape=(None, channels) + options['patch_size'])
layer2 = batch_norm(Conv3DLayer(layer2,
name='conv1_1',
num_filters=32,
filter_size=3,
pad='same'),
name='BN1')
layer2 = Pool3DLayer(layer2,
name='avgpool_1',
mode='max',
pool_size=2,
stride=2)
layer2 = batch_norm(Conv3DLayer(layer2,
name='conv2_1',
num_filters=64,
filter_size=3,
pad='same'),
name='BN2')
layer2 = Pool3DLayer(layer2,
name='avgpoo2_1',
mode='max',
pool_size=2,
stride=2)
layer2 = DropoutLayer(layer2, name='l2drop', p=0.5)
layer2 = DenseLayer(layer2, name='d_1', num_units=256)
layer2 = DenseLayer(layer2,
name='out',
num_units=2,
nonlinearity=nonlinearities.softmax)
# save weights
net_model = 'model_2'
net_weights2 = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '.pkl')
net_history2 = os.path.join(options['weight_paths'], options['experiment'],
'nets', net_model + '_history.pkl')
net2 = NeuralNet(
layers=layer2,
objective_loss_function=objectives.categorical_crossentropy,
batch_iterator_train=Rotate_batch_Iterator(batch_size=128),
update=updates.adadelta,
on_epoch_finished=[
SaveWeights(net_weights2, only_best=True, pickle=False),
SaveTrainingHistory(net_history2),
EarlyStopping(patience=max_epochs_patience)
],
verbose=options['net_verbose'],
max_epochs=num_epochs,
train_split=TrainSplit(eval_size=train_split_perc),
)
return [net1, net2]
def build_cnn3d(input_var=None,
w_init=None,
n_layers=(4, 2, 1),
n_filters_first=32,
imsize=[32, 32],
n_colors=3,
n_timewin=5,
isMaxpool=True,
filter_size=[(3, 1, 1), (3, 3, 3), (3, 3, 3)],
dropout=0.0,
input_dropout=0.0,
batch_norm_conv=False,
padding='same',
pool_size=[(2, 1, 1), (2, 2, 2), (2, 2, 2)],
factor=2):
"""
Builds a VGG style 3D CNN network followed by a fully-connected layer and a softmax layer.
Stacks are separated by a maxpool layer. Number of kernels in each layer is twice
the number in previous stack.
input_var: Theano variable for input to the network
outputs: pointer to the output of the last layer of network (softmax)
:param input_var: theano variable as input to the network
:param w_init: Initial weight values
:param n_layers: number of layers in each stack. An array of integers with each
value corresponding to the number of layers hesain each stack.
(e.g. [4, 2, 1] == 3 stacks with 4, 2, and 1 layers in each.
:param n_filters_first: number of filters in the first layer
:param imSize: Size of the image
:param n_colors: Number of color channels (depth)
:return: a pointer to the output of last layer
"""
weights = [] # Keeps the weights for all layers
count = 0
# If no initial weight is given, initialize with GlorotUniform
if w_init is None:
w_init = [lasagne.init.GlorotUniform()] * sum(n_layers)
# Input layer
network = InputLayer(shape=(n_timewin, None, n_colors, imsize[0],
imsize[1]),
input_var=input_var)
network = DimshuffleLayer(network, (1, 2, 0, 3, 4))
if input_dropout:
network = lasagne.layers.dropout(network, p=input_dropout)
for i, s in enumerate(n_layers):
for l in range(s):
network = Conv3DLayer(network,
num_filters=n_filters_first * (factor**i),
filter_size=filter_size[i],
W=w_init[count],
pad=padding)
count += 1
weights.append(network.W)
if dropout:
network = lasagne.layers.dropout(network, p=dropout)
if batch_norm_conv:
network = batch_norm(network)
if isMaxpool:
network = MaxPool3DLayer(network, pool_size=pool_size[i])
return network, weights
def build_net(nz=200):
input_depth = 160
input_rows = 64
input_columns = 64
#Encoder
enc = InputLayer(shape=(None, 1, input_depth, input_rows,
input_columns)) #5D tensor
enc = Conv3DLayer(incoming=enc,
num_filters=64,
filter_size=5,
stride=2,
nonlinearity=lrelu(0.2),
pad=2)
enc = Conv3DLayer(incoming=enc,
num_filters=128,
filter_size=5,
stride=2,
nonlinearity=lrelu(0.2),
pad=2)
enc = Conv3DLayer(incoming=enc,
num_filters=256,
filter_size=5,
stride=2,
nonlinearity=lrelu(0.2),
pad=2)
enc = Conv3DLayer(incoming=enc,
num_filters=256,
filter_size=5,
stride=2,
nonlinearity=lrelu(0.2),
pad=2)
enc = reshape(incoming=enc, shape=(-1, 256 * 4 * 4 * 10))
enc = DenseLayer(incoming=enc, num_units=nz, nonlinearity=sigmoid)
#Decoder
dec = InputLayer(shape=(None, nz))
dec = DenseLayer(incoming=dec, num_units=256 * 4 * 4 * 10)
dec = reshape(incoming=dec, shape=(-1, 256, 10, 4, 4))
dec = Deconv3D(incoming=dec,
num_filters=256,
filter_size=4,
stride=2,
crop=1,
nonlinearity=relu)
dec = Deconv3D(incoming=dec,
num_filters=128,
filter_size=4,
stride=2,
crop=1,
nonlinearity=relu)
dec = Deconv3D(incoming=dec,
num_filters=64,
filter_size=4,
stride=2,
crop=1,
nonlinearity=relu)
dec = Deconv3D(incoming=dec,
num_filters=1,
filter_size=4,
stride=2,
crop=1,
nonlinearity=sigmoid)
return enc, dec
def build_network(self):
self.input_var = tensor5()
self.output_var = matrix()
net = OrderedDict()
if self.instance_norm:
norm_fct = batch_norm
norm_kwargs = {'axes': (2, 3, 4)}
else:
norm_fct = batch_norm
norm_kwargs = {'axes': 'auto'}
self.input_layer = net['input'] = InputLayer(
(self.batch_size, self.n_input_channels, self.input_dim[0],
self.input_dim[1], self.input_dim[2]), self.input_var)
net['contr_1_1'] = norm_fct(
Conv3DLayer(net['input'],
self.base_n_filters,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['contr_1_2'] = norm_fct(
Conv3DLayer(net['contr_1_1'],
self.base_n_filters,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['pool1'] = Pool3DLayer(net['contr_1_2'], (1, 2, 2))
net['contr_2_1'] = norm_fct(
Conv3DLayer(net['pool1'],
self.base_n_filters * 2,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['contr_2_2'] = norm_fct(
Conv3DLayer(net['contr_2_1'],
self.base_n_filters * 2,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
l = net['pool2'] = Pool3DLayer(net['contr_2_2'], (1, 2, 2))
if self.dropout is not None:
l = DropoutLayer(l, p=self.dropout)
net['contr_3_1'] = norm_fct(
Conv3DLayer(l,
self.base_n_filters * 4,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['contr_3_2'] = norm_fct(
Conv3DLayer(net['contr_3_1'],
self.base_n_filters * 4,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
l = net['pool3'] = Pool3DLayer(net['contr_3_2'], (1, 2, 2))
if self.dropout is not None:
l = DropoutLayer(l, p=self.dropout)
net['contr_4_1'] = norm_fct(
Conv3DLayer(l,
self.base_n_filters * 8,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['contr_4_2'] = norm_fct(
Conv3DLayer(net['contr_4_1'],
self.base_n_filters * 8,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
l = net['pool4'] = Pool3DLayer(net['contr_4_2'], (1, 2, 2))
if self.dropout is not None:
l = DropoutLayer(l, p=self.dropout)
net['encode_1'] = norm_fct(
Conv3DLayer(l,
self.base_n_filters * 16,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
l = net['encode_2'] = norm_fct(
Conv3DLayer(net['encode_1'],
self.base_n_filters * 16,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['upscale1'] = Upscale3DLayer(l, (1, 2, 2))
l = net['concat1'] = ConcatLayer([net['upscale1'], net['contr_4_2']],
cropping=(None, None, "center",
"center", "center"))
if self.dropout is not None:
l = DropoutLayer(l, p=self.dropout)
net['expand_1_1'] = norm_fct(
Conv3DLayer(l,
self.base_n_filters * 8,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
l = net['expand_1_2'] = norm_fct(
Conv3DLayer(net['expand_1_1'],
self.base_n_filters * 8,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['upscale2'] = Upscale3DLayer(l, (1, 2, 2))
l = net['concat2'] = ConcatLayer([net['upscale2'], net['contr_3_2']],
cropping=(None, None, "center",
"center", "center"))
if self.dropout is not None:
l = DropoutLayer(l, p=self.dropout)
net['expand_2_1'] = norm_fct(
Conv3DLayer(l,
self.base_n_filters * 4,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
ds2 = l = net['expand_2_2'] = norm_fct(
Conv3DLayer(net['expand_2_1'],
self.base_n_filters * 4,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['upscale3'] = Upscale3DLayer(l, (1, 2, 2))
l = net['concat3'] = ConcatLayer([net['upscale3'], net['contr_2_2']],
cropping=(None, None, "center",
"center", "center"))
if self.dropout is not None:
l = DropoutLayer(l, p=self.dropout)
net['expand_3_1'] = norm_fct(
Conv3DLayer(l,
self.base_n_filters * 2,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
l = net['expand_3_2'] = norm_fct(
Conv3DLayer(net['expand_3_1'],
self.base_n_filters * 2,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['upscale4'] = Upscale3DLayer(l, (1, 2, 2))
net['concat4'] = ConcatLayer([net['upscale4'], net['contr_1_2']],
cropping=(None, None, "center", "center",
"center"))
net['expand_4_1'] = norm_fct(
Conv3DLayer(net['concat4'],
self.base_n_filters,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['expand_4_2'] = norm_fct(
Conv3DLayer(net['expand_4_1'],
self.base_n_filters,
3,
nonlinearity=self.nonlinearity,
pad=self.pad,
W=lasagne.init.HeNormal(gain="relu")), **norm_kwargs)
net['output_segmentation'] = Conv3DLayer(net['expand_4_2'],
self.num_classes,
1,
nonlinearity=None)
ds2_1x1_conv = Conv3DLayer(ds2,
self.num_classes,
1,
1,
'same',
nonlinearity=lasagne.nonlinearities.linear,
W=lasagne.init.HeNormal(gain='relu'))
ds1_ds2_sum_upscale = Upscale3DLayer(ds2_1x1_conv, (1, 2, 2))
ds3_1x1_conv = Conv3DLayer(net['expand_3_2'],
self.num_classes,
1,
1,
'same',
nonlinearity=lasagne.nonlinearities.linear,
W=lasagne.init.HeNormal(gain='relu'))
ds1_ds2_sum_upscale_ds3_sum = ElemwiseSumLayer(
(ds1_ds2_sum_upscale, ds3_1x1_conv))
ds1_ds2_sum_upscale_ds3_sum_upscale = Upscale3DLayer(
ds1_ds2_sum_upscale_ds3_sum, (1, 2, 2))
self.seg_layer = l = ElemwiseSumLayer(
(net['output_segmentation'], ds1_ds2_sum_upscale_ds3_sum_upscale))
net['dimshuffle'] = DimshuffleLayer(l, (0, 2, 3, 4, 1))
batch_size, n_z, n_rows, n_cols, _ = lasagne.layers.get_output(
net['dimshuffle']).shape
net['reshapeSeg'] = ReshapeLayer(
net['dimshuffle'],
(batch_size * n_rows * n_cols * n_z, self.num_classes))
self.output_layer = net['output_flattened'] = NonlinearityLayer(
net['reshapeSeg'], nonlinearity=lasagne.nonlinearities.softmax)
def buildBaseNet(self, inputShape=(None, 4, 25, 25, 25), forSummary=False):
if not forSummary:
message = 'Building the Architecture of BaseNet'
self.logger.info(logMessage('+', message))
baseNet = InputLayer(self.inputShape, self.inputVar)
if not forSummary:
message = 'Building the convolution layers'
self.logger.info(logMessage('-', message))
kernelShapeListLen = len(self.kernelNumList)
summary = '\n' + '.' * 130 + '\n'
summary += ' {:<15} {:<50} {:<29} {:<29}\n'.format(
'Layer', 'Input shape', 'W shape', 'Output shape')
summary += '.' * 130 + '\n'
summary += '{:<3} {:<15} {:<50} {:<29} {:<29}\n'.format(
1, 'Input', inputShape, '',
get_output_shape(baseNet, input_shapes=inputShape))
for i in xrange(kernelShapeListLen - 1):
kernelShape = self.kernelShapeList[i]
kernelNum = self.kernelNumList[i]
conv3D = Conv3DLayer(incoming=baseNet,
num_filters=kernelNum,
filter_size=kernelShape,
W=HeNormal(gain='relu'),
nonlinearity=linear,
name='Conv3D{}'.format(i))
# Just for summary the fitler shape.
WShape = conv3D.W.get_value().shape
summary += '{:<3} {:<15} {:<50} {:<29} {:<29}\n'.format(
i + 2, 'Conv3D',
get_output_shape(baseNet, input_shapes=inputShape), WShape,
get_output_shape(conv3D, input_shapes=inputShape))
batchNormLayer = BatchNormLayer(conv3D)
preluLayer = prelu(batchNormLayer)
concatLayerInputShape = '{:<25}{:<25}'.format(
get_output_shape(conv3D, input_shapes=inputShape),
get_output_shape(baseNet, input_shapes=inputShape))
baseNet = ConcatLayer(
[preluLayer, baseNet],
1,
cropping=['center', 'None', 'center', 'center', 'center'])
summary += ' {:<15} {:<50} {:<29} {:<29}\n'.format(
'Concat', concatLayerInputShape, '',
get_output_shape(baseNet, input_shapes=inputShape))
if not forSummary:
message = 'Finish Built the convolution layers'
self.logger.info(logMessage('-', message))
message = 'Building the last classfication layers'
self.logger.info(logMessage('-', message))
assert self.kernelShapeList[-1] == [1, 1, 1]
kernelShape = self.kernelShapeList[-1]
kernelNum = self.kernelNumList[-1]
conv3D = Conv3DLayer(incoming=baseNet,
num_filters=kernelNum,
filter_size=kernelShape,
W=HeNormal(gain='relu'),
nonlinearity=linear,
name='Classfication Layer')
receptiveFieldList = [
inputShape[idx] -
get_output_shape(conv3D, input_shapes=inputShape)[idx] + 1
for idx in xrange(-3, 0)
]
assert receptiveFieldList != []
receptiveFieldSet = set(receptiveFieldList)
assert len(receptiveFieldSet) == 1, (receptiveFieldSet, inputShape,
get_output_shape(
conv3D,
input_shapes=inputShape))
self.receptiveField = list(receptiveFieldSet)[0]
# Just for summary the fitler shape.
WShape = conv3D.W.get_value().shape
summary += '{:<3} {:<15} {:<50} {:<29} {:<29}\n'.format(
kernelShapeListLen + 1, 'Conv3D',
get_output_shape(baseNet, input_shapes=inputShape), WShape,
get_output_shape(conv3D, input_shapes=inputShape))
# The output shape should be (batchSize, numOfClasses, zSize, xSize, ySize).
# We will reshape it to (batchSize * zSize * xSize * ySize, numOfClasses),
# because, the softmax in theano can only receive matrix.
baseNet = DimshuffleLayer(conv3D, (0, 2, 3, 4, 1))
summary += ' {:<15} {:<50} {:<29} {:<29}\n'.format(
'Dimshuffle', get_output_shape(conv3D, input_shapes=inputShape),
'', get_output_shape(baseNet, input_shapes=inputShape))
batchSize, zSize, xSize, ySize, _ = get_output(baseNet).shape
reshapeLayerInputShape = get_output_shape(baseNet,
input_shapes=inputShape)
baseNet = ReshapeLayer(baseNet,
(batchSize * zSize * xSize * ySize, kernelNum))
summary += ' {:<15} {:<50} {:<29} {:<29}\n'.format(
'Reshape', reshapeLayerInputShape, '',
get_output_shape(baseNet, input_shapes=inputShape))
nonlinearityLayerInputShape = get_output_shape(baseNet,
input_shapes=inputShape)
baseNet = NonlinearityLayer(baseNet, softmax)
summary += ' {:<15} {:<50} {:<29} {:<29}\n'.format(
'Nonlinearity', nonlinearityLayerInputShape, '',
get_output_shape(baseNet, input_shapes=inputShape))
if not forSummary:
message = 'Finish Built the last classfication layers'
self.logger.info(logMessage('-', message))
message = 'The Receptivr Field of BaseNet equal {}'.format(
self.receptiveField)
self.logger.info(logMessage('*', message))
message = 'Finish Building the Architecture of BaseNet'
self.logger.info(logMessage('+', message))
summary += '.' * 130 + '\n'
self._summary = summary
return baseNet
output = self.merge_function(output, input)
else:
output = input
return output
# Definition of the network
conv_num_filters = 48
l = InputLayer(shape=(None, 1, 48, 72, 64), name="input")
l_input = l
# # # #
# encoding
# # # #
l = Conv3DLayer(l,
flip_filters=False,
num_filters=16,
filter_size=(1, 1, 3),
pad='valid',
name="conv")
l = Conv3DLayer(l,
flip_filters=False,
num_filters=16,
filter_size=(1, 3, 1),
pad='valid',
name="conv")
l_conv_0 = l = Conv3DLayer(l,
flip_filters=False,
num_filters=16,
filter_size=(3, 1, 1),
pad='valid',
name="conv")
def build_model(input_var=None):
# Define common parameters
input_shape = (None, 1, 145, 182, 155)
n_filters = {
'l1': 32,
'l2': 64,
'l3': 128,
'l4': 256,
'l5': 256,
'fc': 512
}
kws_conv = {
'filter_size': (3, 3, 3),
'stride': (1, 1, 1),
'pad': 'same', # (1,1,1) -- same might be faster
'W': GlorotUniform(gain='relu'),
'nonlinearity': rectify,
'flip_filters': False
}
kws_maxpool = {'pool_size': (2, 2, 2), 'stride': (2, 2, 2)}
# 'pad': (0,0,0)} -- should be defined per layer
kws_dense = {'W': GlorotUniform(gain='relu'), 'nonlinearity': rectify}
kws_dropout = {'p': 0.5}
# Define network architecture
net = InputLayer(input_shape, input_var, name='input')
# ----------- 1st layer group ---------------
net = batch_norm(
Conv3DLayer(net, n_filters['l1'], name='conv1a', **kws_conv))
net = MaxPool3DLayer(net, name='pool1', pad=(1, 0, 1), **kws_maxpool)
# ------------- 2nd layer group --------------
net = batch_norm(
Conv3DLayer(net, n_filters['l2'], name='conv2a', **kws_conv))
net = MaxPool3DLayer(net, name='pool2', pad=(1, 1, 0), **kws_maxpool)
# ----------------- 3rd layer group --------------
net = batch_norm(
Conv3DLayer(net, n_filters['l3'], name='conv3a', **kws_conv))
net = batch_norm(
Conv3DLayer(net, n_filters['l3'], name='conv3b', **kws_conv))
net = MaxPool3DLayer(net, name='pool3', pad=(1, 0, 1), **kws_maxpool)
# ----------------- 4th layer group --------------
net = batch_norm(
Conv3DLayer(net, n_filters['l4'], name='conv4a', **kws_conv))
net = batch_norm(
Conv3DLayer(net, n_filters['l4'], name='conv4b', **kws_conv))
net = MaxPool3DLayer(net, name='pool4', pad=(1, 1, 0), **kws_maxpool)
# ----------------- 5th layer group --------------
net = batch_norm(
Conv3DLayer(net, n_filters['l5'], name='conv5a', **kws_conv))
net = batch_norm(
Conv3DLayer(net, n_filters['l5'], name='conv5b', **kws_conv))
net = MaxPool3DLayer(net, name='pool5', **kws_maxpool)
# ----------------- FC layers group --------------
net = batch_norm(DenseLayer(net, n_filters['fc'], name='fc6', **kws_dense))
net = DropoutLayer(net, name='fc6_dropout', **kws_dropout)
net = batch_norm(DenseLayer(net, n_filters['fc'], name='fc7', **kws_dense))
net = DropoutLayer(net, name='fc7_dropout', **kws_dropout)
# ----------------- Output layers group --------------
net = batch_norm(DenseLayer(net, 3, nonlinearity=None, name='fc8'))
net = NonlinearityLayer(net, nonlinearity=softmax, name='prob')
return net
def build_net(input_var=None,
input_shape=(128, 128, 128),
num_output_classes=4,
num_input_channels=4,
base_n_filter=8,
do_instance_norm=True,
batch_size=None,
dropout_p=0.3,
do_norm=True):
nonlin = lasagne.nonlinearities.leaky_rectify
if do_instance_norm:
axes = (2, 3, 4)
else:
axes = 'auto'
def conv_norm_lrelu(l_in, feat_out):
l = Conv3DLayer(l_in,
feat_out,
3,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
if do_norm:
l = BatchNormLayer(l, axes=axes)
return NonlinearityLayer(l, nonlin)
def norm_lrelu_conv(l_in, feat_out, stride=1, filter_size=3):
if do_norm:
l_in = BatchNormLayer(l_in, axes=axes)
l = NonlinearityLayer(l_in, nonlin)
return Conv3DLayer(l,
feat_out,
filter_size,
stride,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
def lrelu_conv(l_in, feat_out, stride=1, filter_size=3):
l = NonlinearityLayer(l_in, nonlin)
return Conv3DLayer(l,
feat_out,
filter_size,
stride,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
def norm_lrelu_upscale_conv_norm_lrelu(l_in, feat_out):
if do_norm:
l_in = BatchNormLayer(l_in, axes=axes)
l = NonlinearityLayer(l_in, nonlin)
l = Upscale3DLayer(l, 2)
l = Conv3DLayer(l,
feat_out,
3,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = NonlinearityLayer(l, nonlin)
return l
l_in = InputLayer(shape=(batch_size, num_input_channels, input_shape[0],
input_shape[1], input_shape[2]),
input_var=input_var)
l = r = Conv3DLayer(l_in,
num_filters=base_n_filter,
filter_size=3,
stride=1,
nonlinearity=linear,
pad='same',
W=HeNormal(gain='relu'))
l = NonlinearityLayer(l, nonlin)
l = Conv3DLayer(l,
num_filters=base_n_filter,
filter_size=3,
stride=1,
nonlinearity=linear,
pad='same',
W=HeNormal(gain='relu'))
l = DropoutLayer(l, dropout_p)
l = lrelu_conv(l, base_n_filter, 1, 3)
l = ElemwiseSumLayer((l, r))
skip1 = NonlinearityLayer(l, nonlin)
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = NonlinearityLayer(l, nonlin)
l = r = Conv3DLayer(l,
base_n_filter * 2,
3,
2,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_conv(l, base_n_filter * 2)
l = DropoutLayer(l, dropout_p)
l = norm_lrelu_conv(l, base_n_filter * 2)
l = ElemwiseSumLayer((l, r))
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = skip2 = NonlinearityLayer(l, nonlin)
l = r = Conv3DLayer(l,
base_n_filter * 4,
3,
2,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_conv(l, base_n_filter * 4)
l = DropoutLayer(l, dropout_p)
l = norm_lrelu_conv(l, base_n_filter * 4)
l = ElemwiseSumLayer((l, r))
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = skip3 = NonlinearityLayer(l, nonlin)
l = r = Conv3DLayer(l,
base_n_filter * 8,
3,
2,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_conv(l, base_n_filter * 8)
l = DropoutLayer(l, dropout_p)
l = norm_lrelu_conv(l, base_n_filter * 8)
l = ElemwiseSumLayer((l, r))
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = skip4 = NonlinearityLayer(l, nonlin)
l = r = Conv3DLayer(l,
base_n_filter * 16,
3,
2,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_conv(l, base_n_filter * 16)
l = DropoutLayer(l, dropout_p)
l = norm_lrelu_conv(l, base_n_filter * 16)
l = ElemwiseSumLayer((l, r))
l = norm_lrelu_upscale_conv_norm_lrelu(l, base_n_filter * 8)
l = Conv3DLayer(l,
base_n_filter * 8,
1,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
if do_norm:
l = BatchNormLayer(l, axes=axes)
l = NonlinearityLayer(l, nonlin)
l = ConcatLayer((skip4, l), cropping=[None, None, 'center', 'center'])
l = conv_norm_lrelu(l, base_n_filter * 16)
l = Conv3DLayer(l,
base_n_filter * 8,
1,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_upscale_conv_norm_lrelu(l, base_n_filter * 4)
l = ConcatLayer((skip3, l), cropping=[None, None, 'center', 'center'])
l = ds2 = conv_norm_lrelu(l, base_n_filter * 8)
l = Conv3DLayer(l,
base_n_filter * 4,
1,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_upscale_conv_norm_lrelu(l, base_n_filter * 2)
l = ConcatLayer((skip2, l), cropping=[None, None, 'center', 'center'])
l = ds3 = conv_norm_lrelu(l, base_n_filter * 4)
l = Conv3DLayer(l,
base_n_filter * 2,
1,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
l = norm_lrelu_upscale_conv_norm_lrelu(l, base_n_filter)
l = ConcatLayer((skip1, l), cropping=[None, None, 'center', 'center'])
l = conv_norm_lrelu(l, base_n_filter * 2)
l_pred = Conv3DLayer(l,
num_output_classes,
1,
pad='same',
nonlinearity=None)
ds2_1x1_conv = Conv3DLayer(ds2,
num_output_classes,
1,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
ds1_ds2_sum_upscale = Upscale3DLayer(ds2_1x1_conv, 2)
ds3_1x1_conv = Conv3DLayer(ds3,
num_output_classes,
1,
1,
'same',
nonlinearity=linear,
W=HeNormal(gain='relu'))
ds1_ds2_sum_upscale_ds3_sum = ElemwiseSumLayer(
(ds1_ds2_sum_upscale, ds3_1x1_conv))
ds1_ds2_sum_upscale_ds3_sum_upscale = Upscale3DLayer(
ds1_ds2_sum_upscale_ds3_sum, 2)
l = seg_layer = ElemwiseSumLayer(
(l_pred, ds1_ds2_sum_upscale_ds3_sum_upscale))
l = DimshuffleLayer(l, (0, 2, 3, 4, 1))
batch_size, n_rows, n_cols, n_z, _ = lasagne.layers.get_output(l).shape
l = ReshapeLayer(l,
(batch_size * n_rows * n_cols * n_z, num_output_classes))
l = NonlinearityLayer(l, nonlinearity=lasagne.nonlinearities.softmax)
return l, seg_layer