view baseline_algorithms/mlp/mlp_nist.py @ 126:66b9513b578f

Fixed bug in pycaptcha/Facade.py
author boulanni <nicolas_boulanger@hotmail.com>
date Thu, 18 Feb 2010 14:15:47 -0500
parents 93b4b84d86cf
children f341a4efb44a
line wrap: on
line source

"""
This tutorial introduces the multilayer perceptron using Theano.  

 A multilayer perceptron is a logistic regressor where
instead of feeding the input to the logistic regression you insert a
intermidiate layer, called the hidden layer, that has a nonlinear 
activation function (usually tanh or sigmoid) . One can use many such 
hidden layers making the architecture deep. The tutorial will also tackle 
the problem of MNIST digit classification.

.. math::

    f(x) = G( b^{(2)} + W^{(2)}( s( b^{(1)} + W^{(1)} x))),

References:

    - textbooks: "Pattern Recognition and Machine Learning" - 
                 Christopher M. Bishop, section 5

TODO: recommended preprocessing, lr ranges, regularization ranges (explain 
      to do lr first, then add regularization)

"""
__docformat__ = 'restructedtext en'

import pdb
import numpy
import pylab
import theano
import theano.tensor as T
import time 
import theano.tensor.nnet
from pylearn.io import filetensor as ft

data_path = '/data/lisa/data/nist/by_class/'

class MLP(object):
    """Multi-Layer Perceptron Class

    A multilayer perceptron is a feedforward artificial neural network model 
    that has one layer or more of hidden units and nonlinear activations. 
    Intermidiate layers usually have as activation function thanh or the 
    sigmoid function  while the top layer is a softamx layer. 
    """



    def __init__(self, input, n_in, n_hidden, n_out):
        """Initialize the parameters for the multilayer perceptron

        :param input: symbolic variable that describes the input of the 
        architecture (one minibatch)

        :param n_in: number of input units, the dimension of the space in 
        which the datapoints lie

        :param n_hidden: number of hidden units 

        :param n_out: number of output units, the dimension of the space in 
        which the labels lie

        """

        # initialize the parameters theta = (W1,b1,W2,b2) ; note that this 
        # example contains only one hidden layer, but one can have as many 
        # layers as he/she wishes, making the network deeper. The only 
        # problem making the network deep this way is during learning, 
        # backpropagation being unable to move the network from the starting
        # point towards; this is where pre-training helps, giving a good 
        # starting point for backpropagation, but more about this in the 
        # other tutorials
        
        # `W1` is initialized with `W1_values` which is uniformely sampled
        # from -6./sqrt(n_in+n_hidden) and 6./sqrt(n_in+n_hidden)
        # the output of uniform if converted using asarray to dtype 
        # theano.config.floatX so that the code is runable on GPU
        W1_values = numpy.asarray( numpy.random.uniform( \
              low = -numpy.sqrt(6./(n_in+n_hidden)), \
              high = numpy.sqrt(6./(n_in+n_hidden)), \
              size = (n_in, n_hidden)), dtype = theano.config.floatX)
        # `W2` is initialized with `W2_values` which is uniformely sampled 
        # from -6./sqrt(n_hidden+n_out) and 6./sqrt(n_hidden+n_out)
        # the output of uniform if converted using asarray to dtype 
        # theano.config.floatX so that the code is runable on GPU
        W2_values = numpy.asarray( numpy.random.uniform( 
              low = -numpy.sqrt(6./(n_hidden+n_out)), \
              high= numpy.sqrt(6./(n_hidden+n_out)),\
              size= (n_hidden, n_out)), dtype = theano.config.floatX)

        self.W1 = theano.shared( value = W1_values )
        self.b1 = theano.shared( value = numpy.zeros((n_hidden,), 
                                                dtype= theano.config.floatX))
        self.W2 = theano.shared( value = W2_values )
        self.b2 = theano.shared( value = numpy.zeros((n_out,), 
                                                dtype= theano.config.floatX))

        # symbolic expression computing the values of the hidden layer
        self.hidden = T.tanh(T.dot(input, self.W1)+ self.b1)

        # symbolic expression computing the values of the top layer 
        self.p_y_given_x= T.nnet.softmax(T.dot(self.hidden, self.W2)+self.b2)

        # compute prediction as class whose probability is maximal in 
        # symbolic form
        self.y_pred = T.argmax( self.p_y_given_x, axis =1)
        
        # L1 norm ; one regularization option is to enforce L1 norm to 
        # be small 
        self.L1     = abs(self.W1).sum() + abs(self.W2).sum()

        # square of L2 norm ; one regularization option is to enforce 
        # square of L2 norm to be small
        self.L2_sqr = (self.W1**2).sum() + (self.W2**2).sum()



    def negative_log_likelihood(self, y):
        """Return the mean of the negative log-likelihood of the prediction
        of this model under a given target distribution.

        .. math::

            \frac{1}{|\mathcal{D}|}\mathcal{L} (\theta=\{W,b\}, \mathcal{D}) = 
            \frac{1}{|\mathcal{D}|}\sum_{i=0}^{|\mathcal{D}|} \log(P(Y=y^{(i)}|x^{(i)}, W,b)) \\
                \ell (\theta=\{W,b\}, \mathcal{D}) 


        :param y: corresponds to a vector that gives for each example the
        :correct label
        """
        return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]),y])




    def errors(self, y):
        """Return a float representing the number of errors in the minibatch 
        over the total number of examples of the minibatch 
        """

        # check if y has same dimension of y_pred 
        if y.ndim != self.y_pred.ndim:
            raise TypeError('y should have the same shape as self.y_pred', 
                ('y', target.type, 'y_pred', self.y_pred.type))
        # check if y is of the correct datatype        
        if y.dtype.startswith('int'):
            # the T.neq operator returns a vector of 0s and 1s, where 1
            # represents a mistake in prediction
            return T.mean(T.neq(self.y_pred, y))
        else:
            raise NotImplementedError()

#def jobman_mlp(state,channel):
#    (validation_error,test_error,nb_exemples,time)=mlp_full_nist(state.learning_rate,\
 #                                                                state.n_iter,\
 #                                                                state.batch_size,\
 #                                                                state.nb_hidden_units)
 #   state.validation_error = validation_error
 #   state.test_error = test_error
 #   state.nb_exemples = nb_exemples
  #  state.time=time
   # return channel.COMPLETE


                                                                 

def mlp_full_nist(      verbose = False,\
                        train_data = 'all/all_train_data.ft',\
                        train_labels = 'all/all_train_labels.ft',\
                        test_data = 'all/all_test_data.ft',\
                        test_labels = 'all/all_test_labels.ft',\
                        learning_rate=0.01,\
                        L1_reg = 0.00,\
                        L2_reg = 0.0001,\
                        nb_max_exemples=1000000,\
                        batch_size=20,\
                        nb_hidden = 500,\
                        nb_targets = 62):
   
    
   
    f = open(data_path+train_data)
    g= open(data_path+train_labels)
    h = open(data_path+test_data)
    i= open(data_path+test_labels)
    
    raw_train_data = ft.read(f)
    raw_train_labels = ft.read(g)
    raw_test_data = ft.read(h)
    raw_test_labels = ft.read(i)
    
    f.close()
    g.close()
    i.close()
    h.close()
    #create a validation set the same size as the test size
    #use the end of the training array for this purpose
    #discard the last remaining so we get a %batch_size number
    test_size=len(raw_test_labels)
    test_size = int(test_size/batch_size)
    test_size*=batch_size
    train_size = len(raw_train_data)
    train_size = int(train_size/batch_size)
    train_size*=batch_size
    validation_size =test_size 
    offset = train_size-test_size
    if verbose == True:
        print 'train size = %d' %train_size
        print 'test size = %d' %test_size
        print 'valid size = %d' %validation_size
        print 'offset = %d' %offset
    
    
    train_set = (raw_train_data,raw_train_labels)
    train_batches = []
    for i in xrange(0, train_size-test_size, batch_size):
        train_batches = train_batches + \
            [(raw_train_data[i:i+batch_size], raw_train_labels[i:i+batch_size])]
            
    test_batches = []
    for i in xrange(0, test_size, batch_size):
        test_batches = test_batches + \
            [(raw_test_data[i:i+batch_size], raw_test_labels[i:i+batch_size])]
    
    validation_batches = []
    for i in xrange(0, test_size, batch_size):
        validation_batches = validation_batches + \
            [(raw_train_data[offset+i:offset+i+batch_size], raw_train_labels[offset+i:offset+i+batch_size])]


    ishape     = (32,32) # this is the size of NIST images

    # allocate symbolic variables for the data
    x = T.fmatrix()  # the data is presented as rasterized images
    y = T.lvector()  # the labels are presented as 1D vector of 
                          # [long int] labels

    # construct the logistic regression class
    classifier = MLP( input=x.reshape((batch_size,32*32)),\
                        n_in=32*32,\
                        n_hidden=nb_hidden,\
                        n_out=nb_targets)

    # the cost we minimize during training is the negative log likelihood of 
    # the model plus the regularization terms (L1 and L2); cost is expressed
    # here symbolically
    cost = classifier.negative_log_likelihood(y) \
         + L1_reg * classifier.L1 \
         + L2_reg * classifier.L2_sqr 

    # compiling a theano function that computes the mistakes that are made by 
    # the model on a minibatch
    test_model = theano.function([x,y], classifier.errors(y))

    # compute the gradient of cost with respect to theta = (W1, b1, W2, b2) 
    g_W1 = T.grad(cost, classifier.W1)
    g_b1 = T.grad(cost, classifier.b1)
    g_W2 = T.grad(cost, classifier.W2)
    g_b2 = T.grad(cost, classifier.b2)

    # specify how to update the parameters of the model as a dictionary
    updates = \
        { classifier.W1: classifier.W1 - learning_rate*g_W1 \
        , classifier.b1: classifier.b1 - learning_rate*g_b1 \
        , classifier.W2: classifier.W2 - learning_rate*g_W2 \
        , classifier.b2: classifier.b2 - learning_rate*g_b2 }

    # compiling a theano function `train_model` that returns the cost, but in 
    # the same time updates the parameter of the model based on the rules 
    # defined in `updates`
    train_model = theano.function([x, y], cost, updates = updates )
    n_minibatches        = len(train_batches)

   
   
   
   #conditions for stopping the adaptation:
   #1) we have reached  nb_max_exemples (this is rounded up to be a multiple of the train size)
   #2) validation error is going up (probable overfitting)
   
   # This means we no longer stop on slow convergence as low learning rates stopped
   # too fast. 
    patience              =nb_max_exemples/batch_size
    patience_increase     = 2     # wait this much longer when a new best is 
                                  # found
    improvement_threshold = 0.995 # a relative improvement of this much is 
                                  # considered significant
    validation_frequency = n_minibatches/4
   
     

   
    best_params          = None
    best_validation_loss = float('inf')
    best_iter            = 0
    test_score           = 0.
    start_time = time.clock()
    n_iter = nb_max_exemples/batch_size  # nb of max times we are allowed to run through all exemples
    n_iter = n_iter/n_minibatches + 1
    n_iter=max(1,n_iter) # run at least once on short debug call
    # have a maximum of `n_iter` iterations through the entire dataset
   
    if verbose == True:
        print 'looping at most %d times through the data set' %n_iter
    for iter in xrange(n_iter* n_minibatches):

        # get epoch and minibatch index
        epoch           = iter / n_minibatches
        minibatch_index =  iter % n_minibatches

        # get the minibatches corresponding to `iter` modulo
        # `len(train_batches)`
        x,y = train_batches[ minibatch_index ]
        # convert to float
        x_float = x/255.0
        cost_ij = train_model(x_float,y)

        if (iter+1) % validation_frequency == 0: 
            # compute zero-one loss on validation set 
           
            this_validation_loss = 0.
            for x,y in validation_batches:
                # sum up the errors for each minibatch
                x_float = x/255.0
                this_validation_loss += test_model(x_float,y)
            # get the average by dividing with the number of minibatches
            this_validation_loss /= len(validation_batches)
            if verbose == True:
                print('epoch %i, minibatch %i/%i, validation error %f %%' % \
                    (epoch, minibatch_index+1, n_minibatches, \
                        this_validation_loss*100.))


            # if we got the best validation score until now
            if this_validation_loss < best_validation_loss:

                #improve patience if loss improvement is good enough
                if this_validation_loss < best_validation_loss *  \
                       improvement_threshold :
                    patience = max(patience, iter * patience_increase)
                elif verbose == True:
                    print 'slow convergence stop'

                # save best validation score and iteration number
                best_validation_loss = this_validation_loss
                best_iter = iter

                # test it on the test set
                test_score = 0.
                for x,y in test_batches:
                    x_float=x/255.0
                    test_score += test_model(x_float,y)
                test_score /= len(test_batches)
                if verbose == True:
                    print(('     epoch %i, minibatch %i/%i, test error of best '
                        'model %f %%') % 
                                (epoch, minibatch_index+1, n_minibatches,
                                test_score*100.))
                                
            #if the validation error is going up, we are overfitting
            #stop converging
            elif this_validation_loss > best_validation_loss:
                #calculate the test error at this point and exit
                # test it on the test set
                if verbose==True:
                    print ' We are diverging'
                best_iter = iter
                test_score = 0.
                for x,y in test_batches:
                    x_float=x/255.0
                    test_score += test_model(x_float,y)
                test_score /= len(test_batches)
                if verbose == True:
                    print ' validation error is going up, stopping now'
                    print(('     epoch %i, minibatch %i/%i, test error of best '
                        'model %f %%') % 
                                (epoch, minibatch_index+1, n_minibatches,
                                test_score*100.))
                                
                break


            
            if patience <= iter :
               break 
        

    end_time = time.clock()
    if verbose == True:
        print(('Optimization complete. Best validation score of %f %% '
            'obtained at iteration %i, with test performance %f %%') %  
                    (best_validation_loss * 100., best_iter, test_score*100.))
        print ('The code ran for %f minutes' % ((end_time-start_time)/60.))
    print iter
    return (best_validation_loss * 100.,test_score*100.,best_iter*batch_size,(end_time-start_time)/60)


if __name__ == '__main__':
    mlp_full_mnist()

def jobman_mlp_full_nist(state,channel):
    (validation_error,test_error,nb_exemples,time)=mlp_full_nist(learning_rate=state.learning_rate,\
                                                                nb_max_exemples=state.nb_max_exemples,\
                                                                nb_hidden=state.nb_hidden)
    state.validation_error=validation_error
    state.test_error=test_error
    state.nb_exemples=nb_exemples
    state.time=time
    return channel.COMPLETE