def build_model():
#################
# Regular model #
#################
input_size = data_sizes["sliced:data:singleslice"]
l0 = nn.layers.InputLayer(input_size)
l1a = nn.layers.dnn.Conv2DDNNLayer(l0 , filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l1b = nn.layers.dnn.Conv2DDNNLayer(l1a, filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l1 = nn.layers.dnn.MaxPool2DDNNLayer(l1b, pool_size=(2,2), stride=(2,2))
l2a = nn.layers.dnn.Conv2DDNNLayer(l1 , filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l2b = nn.layers.dnn.Conv2DDNNLayer(l2a, filter_size=(3,3), num_filters=128, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l2 = nn.layers.dnn.MaxPool2DDNNLayer(l2b, pool_size=(2,2), stride=(2,2))
l3a = nn.layers.dnn.Conv2DDNNLayer(l2 , filter_size=(3,3), num_filters=256, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l3b = nn.layers.dnn.Conv2DDNNLayer(l3a, filter_size=(3,3), num_filters=256, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l3c = nn.layers.dnn.Conv2DDNNLayer(l3b, filter_size=(3,3), num_filters=256, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l3 = nn.layers.dnn.MaxPool2DDNNLayer(l3c, pool_size=(2,2), stride=(2,2))
l4a = nn.layers.dnn.Conv2DDNNLayer(l3 , filter_size=(3,3), num_filters=512, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l4b = nn.layers.dnn.Conv2DDNNLayer(l4a, filter_size=(3,3), num_filters=512, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l4c = nn.layers.dnn.Conv2DDNNLayer(l4b, filter_size=(3,3), num_filters=512, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l4 = nn.layers.dnn.MaxPool2DDNNLayer(l4c, pool_size=(2,2), stride=(2,2))
l5a = nn.layers.dnn.Conv2DDNNLayer(l4 , filter_size=(3,3), num_filters=512, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l5b = nn.layers.dnn.Conv2DDNNLayer(l5a, filter_size=(3,3), num_filters=512, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l5c = nn.layers.dnn.Conv2DDNNLayer(l5b, filter_size=(3,3), num_filters=512, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l5 = nn.layers.dnn.MaxPool2DDNNLayer(l5c, pool_size=(2,2), stride=(2,2))
key_scale = "area_per_pixel:sax"
l_scale = nn.layers.InputLayer(data_sizes[key_scale])
# Systole Dense layers
ldsys1 = nn.layers.DenseLayer(l5, num_units=512, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify)
ldsys1drop = nn.layers.dropout(ldsys1, p=0.5)
ldsys2 = nn.layers.DenseLayer(ldsys1drop, num_units=64, W=nn.init.Orthogonal("relu"),b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify)
ldsys2drop = nn.layers.dropout(ldsys2, p=0.5)
ldsys3 = nn.layers.DenseLayer(ldsys2drop, num_units=1, b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.identity)
l_systole = layers.ScaleLayer(ldsys3, scale=l_scale)
# Diastole Dense layers
lddia1 = nn.layers.DenseLayer(l5, num_units=512, W=nn.init.Orthogonal("relu"), b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify)
lddia1drop = nn.layers.dropout(lddia1, p=0.5)
lddia2 = nn.layers.DenseLayer(lddia1drop, num_units=64, W=nn.init.Orthogonal("relu"),b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.rectify)
lddia2drop = nn.layers.dropout(lddia2, p=0.5)
lddia3 = nn.layers.DenseLayer(lddia2drop, num_units=1, b=nn.init.Constant(0.1), nonlinearity=nn.nonlinearities.identity)
l_diastole = layers.ScaleLayer(lddia3, scale=l_scale)
return {
"inputs":{
"sliced:data:singleslice": l0,
key_scale: l_scale,
},
"outputs": {
"systole:value": l_systole,
"diastole:value": l_diastole,
"systole:sigma": deep_learning_layers.FixedConstantLayer(np.ones((batch_size, 1), dtype='float32')*20./np.sqrt(test_time_augmentations)),
"diastole:sigma": deep_learning_layers.FixedConstantLayer(np.ones((batch_size, 1), dtype='float32')*30./np.sqrt(test_time_augmentations)),
},
"regularizable": {
ldsys1: l2_weight,
ldsys2: l2_weight,
ldsys3: l2_weight,
lddia1: l2_weight,
lddia2: l2_weight,
lddia3: l2_weight,
},
}
def build_model():
#################
# Regular model #
#################
input_key = "sliced:data:ax:noswitch"
data_size = data_sizes[input_key]
l0 = InputLayer(data_size)
l0r = batch_norm(reshape(l0, (-1, 1, ) + data_size[1:]))
# (batch, channel, axis, time, x, y)
# convolve over time
l1 = batch_norm(ConvolutionOverAxisLayer(l0r, num_filters=8, filter_size=(3,), axis=(3,), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.0),
))
l1m = batch_norm(MaxPoolOverAxisLayer(l1, pool_size=(4,), axis=(3,)))
# convolve over x and y
l2a = batch_norm(ConvolutionOver2DAxisLayer(l1m, num_filters=8, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.0),
))
l2b = batch_norm(ConvolutionOver2DAxisLayer(l2a, num_filters=8, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.0),
))
l2m = batch_norm(MaxPoolOver2DAxisLayer(l2b, pool_size=(2, 2), axis=(4,5)))
# convolve over x, y, time
l3a = batch_norm(ConvolutionOver3DAxisLayer(l2m, num_filters=32, filter_size=(3, 3, 3),
axis=(3,4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l3b = batch_norm(ConvolutionOver2DAxisLayer(l3a, num_filters=32, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l3m = batch_norm(MaxPoolOver2DAxisLayer(l3b, pool_size=(2, 2), axis=(4,5)))
# convolve over time
l4 = batch_norm(ConvolutionOverAxisLayer(l3m, num_filters=32, filter_size=(3,), axis=(3,), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l4m = batch_norm(MaxPoolOverAxisLayer(l4, pool_size=(2,), axis=(2,)))
# maxpool over axis
l5 = batch_norm(MaxPoolOverAxisLayer(l3m, pool_size=(4,), axis=(2,)))
# convolve over x and y
l6a = batch_norm(ConvolutionOver2DAxisLayer(l5, num_filters=128, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l6b = batch_norm(ConvolutionOver2DAxisLayer(l6a, num_filters=128, filter_size=(3, 3),
axis=(4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
))
l6m = batch_norm(MaxPoolOver2DAxisLayer(l6b, pool_size=(2, 2), axis=(4,5)))
# convolve over time and x,y, is sparse reduction layer
l7 = ConvolutionOver3DAxisLayer(l6m, num_filters=32, filter_size=(3,3,3), axis=(3,4,5), channel=1,
W=lasagne.init.Orthogonal(),
b=lasagne.init.Constant(0.1),
)
key_scale = "area_per_pixel:sax"
l_scale = InputLayer(data_sizes[key_scale])
# Systole Dense layers
ldsys1 = lasagne.layers.DenseLayer(l7, num_units=512,
W=lasagne.init.Orthogonal("relu"),
b=lasagne.init.Constant(0.1),
nonlinearity=lasagne.nonlinearities.rectify)
ldsys1drop = lasagne.layers.dropout(ldsys1, p=0.5)
ldsys2 = lasagne.layers.DenseLayer(ldsys1drop, num_units=128,
W=lasagne.init.Orthogonal("relu"),
b=lasagne.init.Constant(0.1),
nonlinearity=lasagne.nonlinearities.rectify)
ldsys2drop = lasagne.layers.dropout(ldsys2, p=0.5)
ldsys3 = lasagne.layers.DenseLayer(ldsys2drop, num_units=1,
b=lasagne.init.Constant(0.1),
nonlinearity=lasagne.nonlinearities.identity)
l_systole = layers.MuConstantSigmaErfLayer(layers.ScaleLayer(ldsys3, scale=l_scale), sigma=0.0)
# Diastole Dense layers
lddia1 = lasagne.layers.DenseLayer(l7, num_units=512,
W=lasagne.init.Orthogonal("relu"),
b=lasagne.init.Constant(0.1),
nonlinearity=lasagne.nonlinearities.rectify)
lddia1drop = lasagne.layers.dropout(lddia1, p=0.5)
lddia2 = lasagne.layers.DenseLayer(lddia1drop, num_units=128,
W=lasagne.init.Orthogonal("relu"),
b=lasagne.init.Constant(0.1),
nonlinearity=lasagne.nonlinearities.rectify)
lddia2drop = lasagne.layers.dropout(lddia2, p=0.5)
lddia3 = lasagne.layers.DenseLayer(lddia2drop, num_units=1,
b=lasagne.init.Constant(0.1),
nonlinearity=lasagne.nonlinearities.identity)
l_diastole = layers.MuConstantSigmaErfLayer(layers.ScaleLayer(lddia3, scale=l_scale), sigma=0.0)
return {
"inputs":{
input_key: l0,
key_scale: l_scale,
},
"outputs": {
"systole": l_systole,
"diastole": l_diastole,
},
"regularizable": {
ldsys1: l2_weight,
ldsys2: l2_weight,
ldsys3: l2_weight,
lddia1: l2_weight,
lddia2: l2_weight,
lddia3: l2_weight,
},
}