view linear_regression.py @ 100:574f4db76022

restructuring and added test
author Frederic Bastien <bastienf@iro.umontreal.ca>
date Tue, 06 May 2008 14:01:22 -0400
parents c4726e19b8ec
children c4916445e025
line wrap: on
line source


from learner import *
from theano import tensor as t
from compile import Function
from theano.scalar import as_scalar

# this is one of the simplest example of learner, and illustrates
# the use of theano 
class LinearRegression(Learner):
    """
    Implement linear regression, with or without L2 regularization
    (the former is called Ridge Regression and the latter Ordinary Least Squares).

    The predictor parameters are obtained analytically from the training set.
    Training can proceed sequentially (with multiple calls to update with
    different disjoint subsets of the training sets). After each call to
    update the predictor is ready to be used (and optimized for the union
    of all the training sets passed to update since construction or since
    the last call to forget).

    The L2 regularization coefficient is obtained analytically.
    For each (input[t],output[t]) pair in a minibatch,::
    
       output_t = b + W * input_t

    where b and W are obtained by minimizing::

       lambda sum_{ij} W_{ij}^2  + sum_t ||output_t - target_t||^2

    Let X be the whole training set inputs matrix (one input example per row),
    with the first column full of 1's, and Let Y the whole training set
    targets matrix (one example's target vector per row).
    Let theta = the matrix with b in its first column and W in the others,
    then each theta[:,i] is the solution of the linear system::

       XtX * theta[:,i] = XtY[:,i]

    where XtX is a (n_inputs+1)x(n_inputs+1) matrix containing X'*X
    plus lambda on the diagonal except at (0,0),
    and XtY is a (n_inputs+1)*n_outputs matrix containing X'*Y.

    The fields and attributes expected and produced by use and update are the following:

     - Input and output fields (example-wise quantities):

       - 'input' (always expected by use and update as an input_dataset field)
       - 'target' (optionally expected by use and update as an input_dataset field)
       - 'output' (optionally produced by use as an output dataset field)
       - 'squared_error' (optionally produced by use as an output dataset field, needs 'target') = example-wise squared error

     - optional input attributes (optionally expected as input_dataset attributes)

     - optional attributes (optionally expected as input_dataset attributes)
       (warning, this may be dangerous, the 'use' method will use those provided in the 
       input_dataset rather than those learned during 'update'; currently no support
       for providing these to update):
       
       - 'lambda' 
       - 'b' 
       - 'W' 
       - 'regularization_term' 
       - 'XtX'
       - 'XtY'
    """

    def attributeNames(self):
        return ["lambda","b","W","regularization_term","XtX","XtY"]
    
# definitions specifiques a la regression lineaire:


    def global_inputs(self):
        self.lambda = as_scalar(0.,'lambda')
        self.theta = t.matrix('theta')
        self.W = self.theta[:,1:] 
        self.b = self.theta[:,0]
        self.XtX = t.matrix('XtX')
        self.XtY = t.matrix('XtY')

    def global_outputs(self):
        self.regularizer = self.lambda * t.dot(self.W,self.W)
        self.loss = self.regularizer + t.sum(self.squared_error) # this only makes sense if the whole training set fits in memory in a minibatch
        self.loss_function = Function([self.W,self.lambda,self.squared_error],[self.loss])

    def initialize(self):
        self.XtX.resize((1+self.n_inputs,1+self.n_inputs))
        self.XtY.resize((1+self.n_inputs,self.n_outputs))
        self.XtX.data[:,:]=0
        self.XtY.data[:,:]=0
        numpy.diag(self.XtX.data)[1:]=self.lambda.data
        
    def updated_variables(self):
        self.new_XtX = self.XtX + t.dot(self.extended_input.T,self.extended_input)
        self.new_XtY = self.XtY + t.dot(self.extended_input.T,self.target)
        self.new_theta = t.solve(self.XtX,self.XtY)
    
    def minibatch_wise_inputs(self):
        self.input = t.matrix('input') # n_examples x n_inputs
        self.target = t.matrix('target') # n_examples x n_outputs
        
    def minibatch_wise_outputs(self):
        # self.input is a (n_examples, n_inputs) minibatch matrix
        self.extended_input = t.prepend_one_to_each_row(self.input)
        self.output = t.dot(self.input,self.W.T) + self.b  # (n_examples , n_outputs) matrix
        self.squared_error = t.sum_within_rows(t.sqr(self.output-self.target)) # (n_examples ) vector

    def attributeNames(self):
        return ["lambda","b","W","regularization_term","XtX","XtY"]

    def defaultOutputFields(self, input_fields):
        output_fields = ["output"]
        if "target" in input_fields:
            output_fields.append("squared_error")
        return output_fields
        
    # poutine generale basee sur ces fonctions

    
    def __init__(self,lambda=0.,max_memory_use=500):
        """
        @type lambda: float
        @param lambda: regularization coefficient
        """
        
        W=t.matrix('W')
        # b is a broadcastable row vector (can be replicated into
        # as many rows as there are examples in the minibach)
        b=t.row('b')
        minibatch_input = t.matrix('input') # n_examples x n_inputs
        minibatch_target = t.matrix('target') # n_examples x n_outputs
        minibatch_output = t.dot(minibatch_input,W.T) + b  # n_examples x n_outputs
        lambda = as_scalar(lambda)
        regularizer = self.lambda * t.dot(W,W)
        example_squared_error = t.sum_within_rows(t.sqr(minibatch_output-minibatch_target))
        self.output_function = Function([W,b,minibatch_input],[minibatch_output])
        self.squared_error_function = Function([minibatch_output,minibatch_target],[self.example_squared_error])
        self.loss_function = Function([W,squared_error],[self.regularizer + t.sum(self.example_squared_error)])
        self.W=None
        self.b=None
        self.XtX=None
        self.XtY=None
        
    def forget(self):
        if self.W:
            self.XtX *= 0
            self.XtY *= 0

    def use(self,input_dataset,output_fieldnames=None,copy_inputs=True):
        input_fieldnames = input_dataset.fieldNames()
        assert "input" in input_fieldnames
        if not output_fields:
            output_fields = ["output"]
            if "target" in input_fieldnames:
                output_fields += ["squared_error"]
        else:
            if "squared_error" in output_fields or "total_loss" in output_fields:
                assert "target" in input_fieldnames

        use_functions = []
        for output_fieldname in output_fieldnames:
            if output_fieldname=="output":
                use_functions.append(self.output_function)
            elif output_fieldname=="squared_error":
                use_functions.append(lambda self.output_function)
    
        n_examples = len(input_dataset)
        
        for minibatch in input_dataset.minibatches(minibatch_size=minibatch_size, allow_odd_last_minibatch=True):
            use_function(