def geomav(x):
if len(x) == 0:
return np.zeros(600)
res = np.cumsum(utils.norm_geometric_average(utils.cdf_to_pdf(x)))
return res
def build_objective(interface_layers):
# l2 regu on certain layers
l2_penalty = nn.regularization.regularize_layer_params_weighted(
interface_layers["regularizable"], nn.regularization.l2)
# build objective
return objectives.KaggleObjective(interface_layers["outputs"],
penalty=l2_penalty)
# Testing
postprocess = postprocess.postprocess
test_time_augmentations = 100 # More augmentations since a we only use single slices
tta_average_method = lambda x: np.cumsum(
utils.norm_geometric_average(utils.cdf_to_pdf(x)))
# nonlinearity putting a lower bound on it's output
def lb_softplus(lb):
return lambda x: nn.nonlinearities.softplus(x) + lb
init = nn.init.Orthogonal()
rnn_layer = functools.partial(nn.layers.RecurrentLayer,
W_in_to_hid=init,
W_hid_to_hid=init,
b=nn.init.Constant(0.1),
nonlinearity=nn.nonlinearities.rectify,
hid_init=nn.init.Constant(0.),
# Objective
l2_weight = 0.000
l2_weight_out = 0.000
def build_objective(interface_layers):
# l2 regu on certain layers
l2_penalty = nn.regularization.regularize_layer_params_weighted(
interface_layers["regularizable"], nn.regularization.l2)
# build objective
return objectives.KaggleObjective(interface_layers["outputs"], penalty=l2_penalty)
# Testing
postprocess = postprocess.postprocess
test_time_augmentations = 1000 # More augmentations since a we only use single slices
tta_average_method = lambda x: np.cumsum(utils.norm_geometric_average(utils.cdf_to_pdf(x)))
# Architecture
def build_model():
#################
# Regular model #
#################
input_size = data_sizes["sliced:data:singleslice:4ch"]
l0 = nn.layers.InputLayer(input_size)
l1a = nn.layers.dnn.Conv2DDNNLayer(l0, W=nn.init.Orthogonal("relu"), filter_size=(3,3), num_filters=64, stride=(1,1), pad="same", nonlinearity=nn.nonlinearities.rectify)
l1b = nn.layers.dnn.Conv2DDNNLayer(l1a, W=nn.init.Orthogonal("relu"), 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))
def geomav(x, **kwargs):
if len(x) == 0:
return np.zeros(600)
res = np.cumsum(utils.norm_geometric_average(utils.cdf_to_pdf(x)))
return res
}
# Objective
l2_weight = 0.000
l2_weight_out = 0.000
def build_objective(interface_layers):
# l2 regu on certain layers
l2_penalty = nn.regularization.regularize_layer_params_weighted(
interface_layers["regularizable"], nn.regularization.l2)
# build objective
return objectives.KaggleObjective(interface_layers["outputs"], penalty=l2_penalty)
# Testing
postprocess = postprocess.postprocess
test_time_augmentations = 100 # More augmentations since a we only use single slices
tta_average_method = lambda x: np.cumsum(utils.norm_geometric_average(utils.cdf_to_pdf(x)))
# nonlinearity putting a lower bound on it's output
def lb_softplus(lb):
return lambda x: nn.nonlinearities.softplus(x) + lb
init = nn.init.Orthogonal()
rnn_layer = functools.partial(nn.layers.RecurrentLayer,
W_in_to_hid=init,
W_hid_to_hid=init,
b=nn.init.Constant(0.1),
nonlinearity=nn.nonlinearities.rectify,
hid_init=nn.init.Constant(0.),
def geomav(x, *args, **kwargs):
x = x[0]
if len(x) == 0:
return np.zeros(600)
res = np.cumsum(utils.norm_geometric_average(utils.cdf_to_pdf(x)))
return res