Package crino :: Module module :: Class Module

Class Module

Known Subclasses:

A Module is a part of a neural network architecture, that may have parameters. Provided an input vector of fixed size, a module is able to compute an output vector, which size is specified at construction. According to their characteristics, several modules can be combined to build a new module.

If a criterion is given to a module, it is able to compute the partial gradients of its parameters, in order to perform a gradient descent.

Attention: This is an abstract class, it must be derived to be used.

Instance Methods

[hide private]

__init__(self, nOutputs, nInputs=None)
Constructs a new Module object.

source code

linkModule(self, previous)
Links the outputs of the previous module to the inputs of the current module.

source code

linkInputs(self, vector, nInputs)
Sets the symbolic vector as the inputs of the current module.

source code

function

trainFunction(self, batch_size=1, lr=0.1, downcast=None, shared_x_train=None, shared_y_train=None)
Constructs and compiles a Theano function in order to train the module.

source code

function

criterionFunction(self, downcast=None, shared_x_data=None, shared_y_data=None)
Constructs and compiles a Theano function in order to compute the criterion on a given set.

source code

function

forwardFunction(self, downcast=None, shared_x_data=None)
Constructs and compiles a Theano function in order to compute the forward on a given set.

source code

function

forward(self, x_test)
Performs the forward step on the given test example.

source code

holdFunction(self)

source code

restoreFunction(self)

source code

prepare(self)
Prepares the module before learning.

source code

prepareGeometry(self)
Sets correctly the geometry (nInputs and nOutputs) of the potential submodules.

source code

prepareParams(self)
Initializes the params of the module and its potential submodules.

source code

prepareBackup(self)
Initializes the backupParams of the module and its potential submodules.

source code

prepareOutput(self)
Computes the symbolic outputs of the module in respect to its inputs.

source code

save(self, filename)
Saves this Module to a file.

source code

Instance Variables

[hide private]

TensorVariable inputs
The symbolic inputs vector of the module, denoted x

TensorVariable outputs
The symbolic outputs vector of the module, denoted ŷ

int nInputs
The inputs size

int nOutputs
The outputs size

list params
The list of parameters

list backupParams
The list of backup parameters

bool prepared
Indicates whether the module have already been prepared

Method Details

[hide private]

init(self, nOutputs, nInputs=None)
(Constructor)

source code

Constructs a new Module object.

Parameters:

nOutputs (int) - The outputs size.
nInputs (int) - The inputs size.

linkModule(self, previous)

source code

Links the outputs of the previous module to the inputs of the current module.

Parameters:

previous (Module) - The previous module to be linked with the current module.

linkInputs(self, vector, nInputs)

source code

Sets the symbolic vector as the inputs of the current module.

Parameters:

vector (TensorVariable) - The symbolic vector that will serve as inputs.
nInputs (int) - The size of this inputs vector.

Attention: You can't change the module inputs size once it has been prepare()'d

trainFunction(self, batch_size=1, lr=0.1, downcast=None, shared_x_train=None, shared_y_train=None)

source code

Constructs and compiles a Theano function in order to train the module.

Parameters:

batch_size (int) -
The size of the batches to use for gradient descent :
- 1 for stochastic gradient descent;
- n ∈ ]1..N_train[ for mini-batch gradient descent (N_train must be a multiple of n);
- N_train for batch gradient descent.
(N_train is the total number of training examples)
lr (float) - The learning rate.
downcast (bool) - If true, allows the inputs data to be downcasted (e.g. from double to single precision floats for GPU use).

Returns: function

a Theano-function that performs one step of gradient descent