def __init__(self, input_size, hidden_sizes, output_size, **_):
"""
Parameters
----------
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
"""
super().__init__(input_size, hidden_sizes)
self.output_size = output_size
self.layer_regression = LayerRegression(self.hidden_sizes[-1],
self.output_size)
self.stopping_layer = LayerDense(self.hidden_sizes[-1] + input_size,
1,
activation="sigmoid",
name="stopping")
def __init__(self, volume_manager, input_size, hidden_sizes, output_size, activation, use_previous_direction=False, predict_offset=False,
use_layer_normalization=False, dropout_prob=0., seed=1234, **_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element X has.
hidden_sizes : int, list of int
Number of hidden units each FFNN layer should have.
output_size : int
Number of units the regression layer should have.
activation : str
Name of the activation function to use in the hidden layers
use_previous_direction : bool
Use the previous direction as an additional input
predict_offset : bool
Predict the offset from the previous direction instead (need use_previous_direction).
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations
dropout_prob : float
Dropout probability for recurrent networks. See: https://arxiv.org/pdf/1512.05287.pdf
seed : int
Random seed used for dropout normalization
"""
super().__init__(input_size, hidden_sizes, activation, use_layer_normalization, dropout_prob, seed)
self.volume_manager = volume_manager
self.output_size = output_size
self.use_previous_direction = use_previous_direction
self.predict_offset = predict_offset
if self.predict_offset:
assert self.use_previous_direction # Need previous direction to predict offset.
layer_regression_activation = "tanh" if self.predict_offset else "identity"
self.layer_regression = LayerDense(self.hidden_sizes[-1], self.output_size, activation=layer_regression_activation)
if self.dropout_prob:
p = 1 - self.dropout_prob
self.dropout_vectors[self.layer_regression.name] = self.srng.binomial(size=(self.layer_regression.input_size,), n=1, p=p, dtype=floatX) / p
def __init__(self, volume_manager, input_size, hidden_sizes, use_layer_normalization=False, use_skip_connections=False, **_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element X has.
hidden_sizes : int, list of int
Number of hidden units each FFNN layer should have.
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations
use_skip_connections : bool
Use skip connections from the input to all hidden layers in the network, and from all hidden layers to the output layer
"""
super().__init__(input_size, hidden_sizes, use_layer_normalization=use_layer_normalization, use_skip_connections=use_skip_connections)
self.volume_manager = volume_manager
self.output_size = 1 # Positive class probability
output_layer_input_size = sum(self.hidden_sizes) if self.use_skip_connections else self.hidden_sizes[-1]
self.layer_classification = LayerDense(output_layer_input_size, self.output_size, activation="sigmoid")
def __init__(self, input_size, hidden_sizes, output_size, **_):
"""
Parameters
----------
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
"""
super().__init__(input_size, hidden_sizes)
self.output_size = output_size
self.layer_regression = LayerRegression(self.hidden_sizes[-1], self.output_size)
self.stopping_layer = LayerDense(self.hidden_sizes[-1] + input_size, 1, activation="sigmoid", name="stopping")
class GRU_RegressionAndBinaryClassification(GRU):
""" A standard GRU model with both a regression output layer and
binary classification output layer.
The regression layer consists in fully connected layer (DenseLayer)
whereas the binary classification layer consists in a fully connected
layer with a sigmoid non-linearity to learn when to stop.
"""
def __init__(self, input_size, hidden_sizes, output_size, **_):
"""
Parameters
----------
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
"""
super().__init__(input_size, hidden_sizes)
self.output_size = output_size
self.layer_regression = LayerRegression(self.hidden_sizes[-1], self.output_size)
self.stopping_layer = LayerDense(self.hidden_sizes[-1] + input_size, 1, activation="sigmoid", name="stopping")
def initialize(self, weights_initializer=initer.UniformInitializer(1234)):
super().initialize(weights_initializer)
self.layer_regression.initialize(weights_initializer)
self.stopping_layer.initialize(weights_initializer)
@property
def hyperparameters(self):
hyperparameters = super().hyperparameters
hyperparameters['output_size'] = self.output_size
return hyperparameters
@property
def parameters(self):
return super().parameters + self.layer_regression.parameters + self.stopping_layer.parameters
def _fprop(self, Xi, Xi_plus1, *args):
outputs = super()._fprop(Xi, *args)
last_layer_h = outputs[len(self.hidden_sizes)-1]
regression_out = self.layer_regression.fprop(last_layer_h)
stopping = self.stopping_layer.fprop(T.concatenate([last_layer_h, Xi_plus1], axis=1))
return outputs + (regression_out, stopping)
def get_output(self, X):
outputs_info_h = []
for hidden_size in self.hidden_sizes:
outputs_info_h.append(T.zeros((X.shape[0], hidden_size)))
results, updates = theano.scan(fn=self._fprop,
outputs_info=outputs_info_h + [None, None],
sequences=[{"input": T.transpose(X, axes=(1, 0, 2)), # We want to scan over sequence elements, not the examples.
"taps": [0, 1]}],
n_steps=X.shape[1]-1)
self.graph_updates = updates
# Put back the examples so they are in the first dimension.
self.regression_out = T.transpose(results[-2], axes=(1, 0, 2))
self.stopping = T.transpose(results[-1], axes=(1, 0, 2))
return self.regression_out, self.stopping
def use(self, X):
directions = self.get_output(X)
return directions
class FFNN_Regression(FFNN):
""" A standard FFNN model with a regression layer stacked on top of it.
"""
def __init__(self, volume_manager, input_size, hidden_sizes, output_size, activation, use_previous_direction=False, predict_offset=False,
use_layer_normalization=False, dropout_prob=0., seed=1234, **_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element X has.
hidden_sizes : int, list of int
Number of hidden units each FFNN layer should have.
output_size : int
Number of units the regression layer should have.
activation : str
Name of the activation function to use in the hidden layers
use_previous_direction : bool
Use the previous direction as an additional input
predict_offset : bool
Predict the offset from the previous direction instead (need use_previous_direction).
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations
dropout_prob : float
Dropout probability for recurrent networks. See: https://arxiv.org/pdf/1512.05287.pdf
seed : int
Random seed used for dropout normalization
"""
super().__init__(input_size, hidden_sizes, activation, use_layer_normalization, dropout_prob, seed)
self.volume_manager = volume_manager
self.output_size = output_size
self.use_previous_direction = use_previous_direction
self.predict_offset = predict_offset
if self.predict_offset:
assert self.use_previous_direction # Need previous direction to predict offset.
layer_regression_activation = "tanh" if self.predict_offset else "identity"
self.layer_regression = LayerDense(self.hidden_sizes[-1], self.output_size, activation=layer_regression_activation)
if self.dropout_prob:
p = 1 - self.dropout_prob
self.dropout_vectors[self.layer_regression.name] = self.srng.binomial(size=(self.layer_regression.input_size,), n=1, p=p, dtype=floatX) / p
def initialize(self, weights_initializer=initer.UniformInitializer(1234)):
super().initialize(weights_initializer)
self.layer_regression.initialize(weights_initializer)
@property
def hyperparameters(self):
hyperparameters = super().hyperparameters
hyperparameters['output_size'] = self.output_size
hyperparameters['use_previous_direction'] = self.use_previous_direction
hyperparameters['predict_offset'] = self.predict_offset
return hyperparameters
@property
def parameters(self):
return super().parameters + self.layer_regression.parameters
def _fprop(self, Xi, *args):
# Xi.shape : (batch_size, 4) *if self.use_previous_direction, Xi.shape : (batch_size,7)
# coords + dwi ID (+ previous_direction)
# coords : streamlines 3D coordinates.
# coords.shape : (batch_size, 4) where the last column is a dwi ID.
# args.shape : n_layers * (batch_size, layer_size)
coords = Xi[:, :4]
# Get diffusion data.
# data_at_coords.shape : (batch_size, input_size)
data_at_coords = self.volume_manager.eval_at_coords(coords)
if self.use_previous_direction:
# previous_direction.shape : (batch_size, 3)
print("Using previous direction")
previous_direction = Xi[:, 4:]
fprop_input = T.concatenate([data_at_coords, previous_direction], axis=1)
else:
fprop_input = data_at_coords
# Hidden state to be passed to the next GRU iteration (next _fprop call)
# next_hidden_state.shape : n_layers * (batch_size, layer_size)
layer_outputs = super()._fprop(fprop_input)
# Compute the direction to follow for step (t)
dropout_W = self.dropout_vectors[self.layer_regression.name] if self.dropout_prob else None
regression_out = self.layer_regression.fprop(layer_outputs[-1], dropout_W)
if self.predict_offset:
print("Predicting offset")
regression_out += previous_direction # Skip-connection from the previous direction.
return layer_outputs + (regression_out,)
def make_sequence_generator(self, subject_id=0, **_):
""" Makes functions that return the prediction for x_{t+1} for every
sequence in the batch given x_{t}.
Parameters
----------
subject_id : int, optional
ID of the subject from which its diffusion data will be used. Default: 0.
"""
# Build the sequence generator as a theano function.
symb_x_t = T.matrix(name="x_t")
layer_outputs = self._fprop(symb_x_t)
# predictions.shape : (batch_size, target_size)
predictions = layer_outputs[-1]
f = theano.function(inputs=[symb_x_t], outputs=[predictions])
def _gen(x_t, states, previous_direction=None):
""" Returns the prediction for x_{t+1} for every
sequence in the batch given x_{t}.
Parameters
----------
x_t : ndarray with shape (batch_size, 3)
Streamline coordinate (x, y, z).
states : list of 2D array of shape (batch_size, hidden_size)
Currrent states of the network.
previous_direction : ndarray with shape (batch_size, 3)
If using previous direction, these should be added to the input
Returns
-------
next_x_t : ndarray with shape (batch_size, 3)
Directions to follow.
new_states : list of 2D array of shape (batch_size, hidden_size)
Updated states of the network after seeing x_t.
"""
# Append the DWI ID of each sequence after the 3D coordinates.
subject_ids = np.array([subject_id] * len(x_t), dtype=floatX)[:, None]
if not self.use_previous_direction:
x_t = np.c_[x_t, subject_ids]
else:
x_t = np.c_[x_t, subject_ids, previous_direction]
results = f(x_t)
next_x_t = results[-1]
next_x_t_both_directions = np.stack([next_x_t, -next_x_t], axis=-1)
next_x_t = next_x_t_both_directions[
(np.arange(next_x_t_both_directions.shape[0])[:, None]),
(np.arange(next_x_t_both_directions.shape[1])[None, :]),
np.argmin(np.linalg.norm(next_x_t_both_directions - previous_direction[:, :, None], axis=1), axis=1)[:, None]]
# FFNN_Regression is not a recurrent network, return original states
new_states = states
return next_x_t, new_states
return _gen
def __init__(self,
volume_manager,
input_size,
hidden_sizes,
output_size,
activation='tanh',
use_previous_direction=False,
predict_offset=False,
use_layer_normalization=False,
drop_prob=0.,
use_zoneout=False,
use_skip_connections=False,
neighborhood_radius=None,
learn_to_stop=False,
seed=1234,
**_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
activation : str
Activation function to apply on the "cell candidate"
use_previous_direction : bool
Use the previous direction as an additional input
predict_offset : bool
Predict the offset from the previous direction instead (need use_previous_direction).
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations and stabilize hidden layer evolution
drop_prob : float
Dropout/Zoneout probability for recurrent networks. See: https://arxiv.org/pdf/1512.05287.pdf & https://arxiv.org/pdf/1606.01305.pdf
use_zoneout : bool
Use zoneout implementation instead of dropout
use_skip_connections : bool
Use skip connections from the input to all hidden layers in the network, and from all hidden layers to the output layer
neighborhood_radius : float
Add signal in positions around the current streamline coordinate to the input (with given length in voxel space); None = no neighborhood
learn_to_stop : bool
Predict whether the streamline being generated should stop or not
seed : int
Random seed used for dropout normalization
"""
self.neighborhood_radius = neighborhood_radius
self.model_input_size = input_size
if self.neighborhood_radius:
self.neighborhood_directions = get_neighborhood_directions(
self.neighborhood_radius)
# Model input size is increased when using neighborhood
self.model_input_size = input_size * self.neighborhood_directions.shape[
0]
super().__init__(self.model_input_size,
hidden_sizes,
activation=activation,
use_layer_normalization=use_layer_normalization,
drop_prob=drop_prob,
use_zoneout=use_zoneout,
use_skip_connections=use_skip_connections,
seed=seed)
# Restore input size
self.input_size = input_size
self.volume_manager = volume_manager
self.output_size = output_size
self.use_previous_direction = use_previous_direction
self.predict_offset = predict_offset
self.learn_to_stop = learn_to_stop
if self.predict_offset:
assert self.use_previous_direction # Need previous direction to predict offset.
# Do not use dropout/zoneout in last hidden layer
layer_regression_activation = "tanh" if self.predict_offset else "identity"
output_layer_input_size = sum(
self.hidden_sizes
) if self.use_skip_connections else self.hidden_sizes[-1]
self.layer_regression = LayerDense(
output_layer_input_size,
self.output_size,
activation=layer_regression_activation,
name="GRU_Regression")
if self.learn_to_stop:
# Predict whether a streamline should stop or keep growing
self.layer_stopping = LayerDense(output_layer_input_size,
1,
activation='sigmoid',
name="GRU_Regression_stopping")
class GRU_Regression(GRU):
""" A standard GRU model with a regression layer stacked on top of it.
"""
def __init__(self,
volume_manager,
input_size,
hidden_sizes,
output_size,
activation='tanh',
use_previous_direction=False,
predict_offset=False,
use_layer_normalization=False,
drop_prob=0.,
use_zoneout=False,
use_skip_connections=False,
neighborhood_radius=None,
learn_to_stop=False,
seed=1234,
**_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
activation : str
Activation function to apply on the "cell candidate"
use_previous_direction : bool
Use the previous direction as an additional input
predict_offset : bool
Predict the offset from the previous direction instead (need use_previous_direction).
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations and stabilize hidden layer evolution
drop_prob : float
Dropout/Zoneout probability for recurrent networks. See: https://arxiv.org/pdf/1512.05287.pdf & https://arxiv.org/pdf/1606.01305.pdf
use_zoneout : bool
Use zoneout implementation instead of dropout
use_skip_connections : bool
Use skip connections from the input to all hidden layers in the network, and from all hidden layers to the output layer
neighborhood_radius : float
Add signal in positions around the current streamline coordinate to the input (with given length in voxel space); None = no neighborhood
learn_to_stop : bool
Predict whether the streamline being generated should stop or not
seed : int
Random seed used for dropout normalization
"""
self.neighborhood_radius = neighborhood_radius
self.model_input_size = input_size
if self.neighborhood_radius:
self.neighborhood_directions = get_neighborhood_directions(
self.neighborhood_radius)
# Model input size is increased when using neighborhood
self.model_input_size = input_size * self.neighborhood_directions.shape[
0]
super().__init__(self.model_input_size,
hidden_sizes,
activation=activation,
use_layer_normalization=use_layer_normalization,
drop_prob=drop_prob,
use_zoneout=use_zoneout,
use_skip_connections=use_skip_connections,
seed=seed)
# Restore input size
self.input_size = input_size
self.volume_manager = volume_manager
self.output_size = output_size
self.use_previous_direction = use_previous_direction
self.predict_offset = predict_offset
self.learn_to_stop = learn_to_stop
if self.predict_offset:
assert self.use_previous_direction # Need previous direction to predict offset.
# Do not use dropout/zoneout in last hidden layer
layer_regression_activation = "tanh" if self.predict_offset else "identity"
output_layer_input_size = sum(
self.hidden_sizes
) if self.use_skip_connections else self.hidden_sizes[-1]
self.layer_regression = LayerDense(
output_layer_input_size,
self.output_size,
activation=layer_regression_activation,
name="GRU_Regression")
if self.learn_to_stop:
# Predict whether a streamline should stop or keep growing
self.layer_stopping = LayerDense(output_layer_input_size,
1,
activation='sigmoid',
name="GRU_Regression_stopping")
def initialize(self, weights_initializer=initer.UniformInitializer(1234)):
super().initialize(weights_initializer)
self.layer_regression.initialize(weights_initializer)
if self.learn_to_stop:
self.layer_stopping.initialize(weights_initializer)
@property
def hyperparameters(self):
hyperparameters = super().hyperparameters
hyperparameters['output_size'] = self.output_size
hyperparameters['use_previous_direction'] = self.use_previous_direction
hyperparameters['predict_offset'] = self.predict_offset
hyperparameters['neighborhood_radius'] = self.neighborhood_radius
hyperparameters['learn_to_stop'] = self.learn_to_stop
return hyperparameters
@property
def parameters(self):
all_params = super().parameters + self.layer_regression.parameters
if self.learn_to_stop:
all_params += self.layer_stopping.parameters
return all_params
def _fprop_step(self, Xi, *args):
# Xi.shape : (batch_size, 4) *if self.use_previous_direction, Xi.shape : (batch_size,7)
# coords + dwi ID (+ previous_direction)
# coords : streamlines 3D coordinates.
# coords.shape : (batch_size, 4) where the last column is a dwi ID.
# args.shape : n_layers * (batch_size, layer_size)
batch_size = Xi.shape[0]
coords = Xi[:, :4]
# Repeat coords and apply the neighborhood transformations
if self.neighborhood_radius:
# coords.shape : (batch_size*len(neighbors_positions), 4)
coords = T.repeat(coords,
self.neighborhood_directions.shape[0],
axis=0)
coords = T.set_subtensor(
coords[:, :3],
coords[:, :3] + T.tile(self.neighborhood_directions,
(batch_size, 1)))
# Get diffusion data.
# data_at_coords.shape : (batch_size, input_size)
data_at_coords = self.volume_manager.eval_at_coords(coords)
# Concatenate back the neighborhood data into a single input vector
if self.neighborhood_radius:
data_at_coords = T.reshape(data_at_coords,
(batch_size, self.model_input_size))
if self.use_previous_direction:
# previous_direction.shape : (batch_size, 3)
previous_direction = Xi[:, 4:]
fprop_input = T.concatenate([data_at_coords, previous_direction],
axis=1)
else:
fprop_input = data_at_coords
# Hidden state to be passed to the next GRU iteration (next _fprop call)
# next_hidden_state.shape : n_layers * (batch_size, layer_size)
next_hidden_state = super()._fprop(fprop_input, *args)
# Compute the direction to follow for step (t)
output_layer_input = T.concatenate(
next_hidden_state,
axis=-1) if self.use_skip_connections else next_hidden_state[-1]
regression_out = self.layer_regression.fprop(output_layer_input)
if self.predict_offset:
regression_out += previous_direction # Skip-connection from the previous direction.
outputs = (regression_out, )
if self.learn_to_stop:
stopping_out = self.layer_stopping.fprop(output_layer_input)
outputs = (stopping_out, regression_out)
return next_hidden_state + outputs
def get_output(self, X):
# X.shape : (batch_size, seq_len, n_features=[4|7])
# For tractography n_features is (x,y,z) + (dwi_id,) + [previous_direction]
outputs_info_h = []
for hidden_size in self.hidden_sizes:
outputs_info_h.append(T.zeros((X.shape[0], hidden_size)))
outputs_info = outputs_info_h + [None]
if self.learn_to_stop:
outputs_info += [None]
results, updates = theano.scan(
fn=self._fprop_step,
# We want to scan over sequence elements, not the examples.
sequences=[T.transpose(X, axes=(1, 0, 2))],
outputs_info=outputs_info,
non_sequences=self.parameters + self.volume_manager.volumes,
strict=True)
self.graph_updates = updates
# Put back the examples so they are in the first dimension.
# regression_out.shape : (batch_size, seq_len, target_size=3)
self.regression_out = T.transpose(results[-1], axes=(1, 0, 2))
model_output = self.regression_out
if self.learn_to_stop:
self.stopping_out = T.transpose(results[-2], axes=(1, 0, 2))
model_output = (self.stopping_out, self.regression_out)
return model_output
def make_sequence_generator(self, subject_id=0, **_):
""" Makes functions that return the prediction for x_{t+1} for every
sequence in the batch given x_{t} and the current state of the model h^{l}_{t}.
Parameters
----------
subject_id : int, optional
ID of the subject from which its diffusion data will be used. Default: 0.
"""
# Build the sequence generator as a theano function.
states_h = []
for i in range(len(self.hidden_sizes)):
state_h = T.matrix(name="layer{}_state_h".format(i))
states_h.append(state_h)
symb_x_t = T.matrix(name="x_t")
new_states = self._fprop_step(symb_x_t, *states_h)
new_states_h = new_states[:len(self.hidden_sizes)]
# predictions.shape : (batch_size, target_size)
predictions = [new_states[-1]]
if self.learn_to_stop:
predictions = new_states[-2:]
f = theano.function(inputs=[symb_x_t] + states_h,
outputs=list(predictions) + list(new_states_h))
def _gen(x_t, states, previous_direction=None):
""" Returns the prediction for x_{t+1} for every
sequence in the batch given x_{t} and the current states
of the model h^{l}_{t}.
Parameters
----------
x_t : ndarray with shape (batch_size, 3)
Streamline coordinate (x, y, z).
states : list of 2D array of shape (batch_size, hidden_size)
Currrent states of the network.
previous_direction : ndarray with shape (batch_size, 3)
If using previous direction, these should be added to the input
Returns
-------
next_x_t : ndarray with shape (batch_size, 3)
Directions to follow.
new_states : list of 2D array of shape (batch_size, hidden_size)
Updated states of the network after seeing x_t.
"""
# Append the DWI ID of each sequence after the 3D coordinates.
subject_ids = np.array([subject_id] * len(x_t), dtype=floatX)[:,
None]
if not self.use_previous_direction:
x_t = np.c_[x_t, subject_ids]
else:
x_t = np.c_[x_t, subject_ids, previous_direction]
results = f(x_t, *states)
if self.learn_to_stop:
stopping = results[0]
next_x_t = results[1]
new_states = results[2:]
output = (next_x_t, stopping)
else:
next_x_t = results[0]
new_states = results[1:]
output = next_x_t
return output, new_states
return _gen
class FFNN_Classification(FFNN):
""" A standard FFNN model with a classification (sigmoid) layer stacked on top of it.
"""
def __init__(self, volume_manager, input_size, hidden_sizes, use_layer_normalization=False, use_skip_connections=False, **_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element X has.
hidden_sizes : int, list of int
Number of hidden units each FFNN layer should have.
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations
use_skip_connections : bool
Use skip connections from the input to all hidden layers in the network, and from all hidden layers to the output layer
"""
super().__init__(input_size, hidden_sizes, use_layer_normalization=use_layer_normalization, use_skip_connections=use_skip_connections)
self.volume_manager = volume_manager
self.output_size = 1 # Positive class probability
output_layer_input_size = sum(self.hidden_sizes) if self.use_skip_connections else self.hidden_sizes[-1]
self.layer_classification = LayerDense(output_layer_input_size, self.output_size, activation="sigmoid")
def initialize(self, weights_initializer=initer.UniformInitializer(1234)):
super().initialize(weights_initializer)
self.layer_classification.initialize(weights_initializer)
@property
def hyperparameters(self):
hyperparameters = super().hyperparameters
return hyperparameters
@property
def parameters(self):
return super().parameters + self.layer_classification.parameters
def _fprop(self, Xi, *args):
# Xi.shape : (batch_size, 4)
# coords + dwi ID
# coords : brain 3D coordinates.
# coords.shape : (batch_size, 4) where the last column is a dwi ID.
# args.shape : n_layers * (batch_size, layer_size)
coords = Xi[:, :4]
# Get diffusion data.
# data_at_coords.shape : (batch_size, input_size)
data_at_coords = self.volume_manager.eval_at_coords(coords)
layer_outputs = super()._fprop(data_at_coords)
# Compute positive class probability
output_layer_input = T.concatenate(layer_outputs, axis=-1) if self.use_skip_connections else layer_outputs[-1]
classification_out = self.layer_classification.fprop(output_layer_input)
# Remove single-dimension from shape
classification_out = classification_out[:, 0]
return layer_outputs + (classification_out,)
class GRU_RegressionAndBinaryClassification(GRU):
""" A standard GRU model with both a regression output layer and
binary classification output layer.
The regression layer consists in fully connected layer (DenseLayer)
whereas the binary classification layer consists in a fully connected
layer with a sigmoid non-linearity to learn when to stop.
"""
def __init__(self, input_size, hidden_sizes, output_size, **_):
"""
Parameters
----------
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
"""
super().__init__(input_size, hidden_sizes)
self.output_size = output_size
self.layer_regression = LayerRegression(self.hidden_sizes[-1],
self.output_size)
self.stopping_layer = LayerDense(self.hidden_sizes[-1] + input_size,
1,
activation="sigmoid",
name="stopping")
def initialize(self, weights_initializer=initer.UniformInitializer(1234)):
super().initialize(weights_initializer)
self.layer_regression.initialize(weights_initializer)
self.stopping_layer.initialize(weights_initializer)
@property
def hyperparameters(self):
hyperparameters = super().hyperparameters
hyperparameters['output_size'] = self.output_size
return hyperparameters
@property
def parameters(self):
return super(
).parameters + self.layer_regression.parameters + self.stopping_layer.parameters
def _fprop(self, Xi, Xi_plus1, *args):
outputs = super()._fprop(Xi, *args)
last_layer_h = outputs[len(self.hidden_sizes) - 1]
regression_out = self.layer_regression.fprop(last_layer_h)
stopping = self.stopping_layer.fprop(
T.concatenate([last_layer_h, Xi_plus1], axis=1))
return outputs + (regression_out, stopping)
def get_output(self, X):
outputs_info_h = []
for hidden_size in self.hidden_sizes:
outputs_info_h.append(T.zeros((X.shape[0], hidden_size)))
results, updates = theano.scan(
fn=self._fprop,
outputs_info=outputs_info_h + [None, None],
sequences=[{
"input": T.transpose(
X, axes=(1, 0, 2)
), # We want to scan over sequence elements, not the examples.
"taps": [0, 1]
}],
n_steps=X.shape[1] - 1)
self.graph_updates = updates
# Put back the examples so they are in the first dimension.
self.regression_out = T.transpose(results[-2], axes=(1, 0, 2))
self.stopping = T.transpose(results[-1], axes=(1, 0, 2))
return self.regression_out, self.stopping
def use(self, X):
directions = self.get_output(X)
return directions
def __init__(self,
volume_manager,
input_size,
hidden_sizes,
output_size,
activation='tanh',
use_previous_direction=False,
predict_offset=False,
use_layer_normalization=False,
drop_prob=0.,
use_zoneout=False,
use_skip_connections=False,
seed=1234,
**_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element Xi in the input sequence X has.
hidden_sizes : int, list of int
Number of hidden units each GRU should have.
output_size : int
Number of units the regression layer should have.
activation : str
Activation function to apply on the "cell candidate"
use_previous_direction : bool
Use the previous direction as an additional input
predict_offset : bool
Predict the offset from the previous direction instead (need use_previous_direction).
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations and stabilize hidden layer evolution
drop_prob : float
Dropout/Zoneout probability for recurrent networks. See: https://arxiv.org/pdf/1512.05287.pdf & https://arxiv.org/pdf/1606.01305.pdf
use_zoneout : bool
Use zoneout implementation instead of dropout
use_skip_connections : bool
Use skip connections from the input to all hidden layers in the network, and from all hidden layers to the output layer
seed : int
Random seed used for dropout normalization
"""
super().__init__(input_size,
hidden_sizes,
activation=activation,
use_layer_normalization=use_layer_normalization,
drop_prob=drop_prob,
use_zoneout=use_zoneout,
use_skip_connections=use_skip_connections,
seed=seed)
self.volume_manager = volume_manager
self.output_size = output_size
self.use_previous_direction = use_previous_direction
self.predict_offset = predict_offset
if self.predict_offset:
assert self.use_previous_direction # Need previous direction to predict offset.
# Do not use dropout/zoneout in last hidden layer
layer_regression_activation = "tanh" if self.predict_offset else "identity"
output_layer_input_size = sum(
self.hidden_sizes
) if self.use_skip_connections else self.hidden_sizes[-1]
self.layer_regression = LayerDense(
output_layer_input_size,
self.output_size,
activation=layer_regression_activation,
name="GRU_Regression")
def __init__(self,
volume_manager,
input_size,
hidden_sizes,
output_size,
activation,
use_previous_direction=False,
predict_offset=False,
use_layer_normalization=False,
dropout_prob=0.,
neighborhood_radius=False,
seed=1234,
**_):
"""
Parameters
----------
volume_manager : :class:`VolumeManger` object
Use to evaluate the diffusion signal at specific coordinates.
input_size : int
Number of units each element X has.
hidden_sizes : int, list of int
Number of hidden units each FFNN layer should have.
output_size : int
Number of units the regression layer should have.
activation : str
Name of the activation function to use in the hidden layers
use_previous_direction : bool
Use the previous direction as an additional input
predict_offset : bool
Predict the offset from the previous direction instead (need use_previous_direction).
use_layer_normalization : bool
Use LayerNormalization to normalize preactivations
dropout_prob : float
Dropout probability for recurrent networks. See: https://arxiv.org/pdf/1512.05287.pdf
neighborhood_radius : float
Add signal in positions around the current streamline coordinate to the input (with given length in voxel space); None = no neighborhood
seed : int
Random seed used for dropout normalization
"""
self.neighborhood_radius = neighborhood_radius
self.model_input_size = input_size
if self.neighborhood_radius:
self.neighborhood_directions = get_neighborhood_directions(
self.neighborhood_radius)
# Model input size is increased when using neighborhood
self.model_input_size = input_size * self.neighborhood_directions.shape[
0]
super().__init__(self.model_input_size,
hidden_sizes,
activation=activation,
use_layer_normalization=use_layer_normalization,
dropout_prob=dropout_prob,
seed=seed)
# Restore input size
self.input_size = input_size
self.volume_manager = volume_manager
self.output_size = output_size
self.use_previous_direction = use_previous_direction
self.predict_offset = predict_offset
if self.predict_offset:
assert self.use_previous_direction # Need previous direction to predict offset.
layer_regression_activation = "tanh" if self.predict_offset else "identity"
self.layer_regression = LayerDense(
self.hidden_sizes[-1],
self.output_size,
activation=layer_regression_activation)
if self.dropout_prob:
p = 1 - self.dropout_prob
self.dropout_vectors[
self.layer_regression.name] = self.srng.binomial(
size=(self.layer_regression.input_size, ),
n=1,
p=p,
dtype=floatX) / p