# HG changeset patch # User Myriam Cote # Date 1267106748 18000 # Node ID e81241cfc2deeb4b955eb511df18696b02688535 # Parent d1bb6e06497aec02a90053142e1f0f1141b41bb4# Parent 221799d791880c1af81a904284875875a5b596a0 merge diff -r d1bb6e06497a -r e81241cfc2de baseline_algorithms/mlp/mlp_nist.py --- a/baseline_algorithms/mlp/mlp_nist.py Thu Feb 25 09:04:40 2010 -0500 +++ b/baseline_algorithms/mlp/mlp_nist.py Thu Feb 25 09:05:48 2010 -0500 @@ -30,6 +30,7 @@ import theano.tensor as T import time import theano.tensor.nnet +import pylearn from pylearn.io import filetensor as ft data_path = '/data/lisa/data/nist/by_class/' @@ -45,7 +46,7 @@ - def __init__(self, input, n_in, n_hidden, n_out): + def __init__(self, input, n_in, n_hidden, n_out,learning_rate): """Initialize the parameters for the multilayer perceptron :param input: symbolic variable that describes the input of the @@ -94,8 +95,14 @@ self.b2 = theano.shared( value = numpy.zeros((n_out,), dtype= theano.config.floatX)) + #include the learning rate in the classifer so + #we can modify it on the fly when we want + lr_value=learning_rate + self.lr=theano.shared(value=lr_value) # 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) @@ -103,6 +110,10 @@ # compute prediction as class whose probability is maximal in # symbolic form self.y_pred = T.argmax( self.p_y_given_x, axis =1) + self.y_pred_num = T.argmax( self.p_y_given_x[0:9], axis =1) + + + # L1 norm ; one regularization option is to enforce L1 norm to # be small @@ -150,21 +161,9 @@ 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,\ + adaptive_lr = 0,\ train_data = 'all/all_train_data.ft',\ train_labels = 'all/all_train_labels.ft',\ test_data = 'all/all_test_data.ft',\ @@ -178,6 +177,14 @@ nb_targets = 62): + configuration = [learning_rate,nb_max_exemples,nb_hidden,adaptive_lr] + + total_validation_error_list = [] + total_train_error_list = [] + learning_rate_list=[] + best_training_error=float('inf'); + + f = open(data_path+train_data) g= open(data_path+train_labels) @@ -235,11 +242,17 @@ y = T.lvector() # the labels are presented as 1D vector of # [long int] labels + if verbose==True: + print 'finished parsing the data' # 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) + n_out=nb_targets, + learning_rate=learning_rate) + + + # the cost we minimize during training is the negative log likelihood of # the model plus the regularization terms (L1 and L2); cost is expressed @@ -260,10 +273,10 @@ # 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 } + { classifier.W1: classifier.W1 - classifier.lr*g_W1 \ + , classifier.b1: classifier.b1 - classifier.lr*g_b1 \ + , classifier.W2: classifier.W2 - classifier.lr*g_W2 \ + , classifier.b2: classifier.b2 - classifier.lr*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 @@ -273,13 +286,17 @@ + + #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) + #2) validation error is going up twice in a row(probable overfitting) # This means we no longer stop on slow convergence as low learning rates stopped # too fast. + + # no longer relevant patience =nb_max_exemples/batch_size patience_increase = 2 # wait this much longer when a new best is # found @@ -296,9 +313,9 @@ 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 = n_iter/n_minibatches + 1 #round up 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 @@ -307,7 +324,9 @@ # 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 ] @@ -317,7 +336,7 @@ 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 @@ -325,26 +344,40 @@ this_validation_loss += test_model(x_float,y) # get the average by dividing with the number of minibatches this_validation_loss /= len(validation_batches) + #save the validation loss + total_validation_error_list.append(this_validation_loss) + + #get the training error rate + this_train_loss=0 + for x,y in train_batches: + # sum up the errors for each minibatch + x_float = x/255.0 + this_train_loss += test_model(x_float,y) + # get the average by dividing with the number of minibatches + this_train_loss /= len(train_batches) + #save the validation loss + total_train_error_list.append(this_train_loss) + if(this_train_loss best_validation_loss: + # if the validation error is going up, we are overfitting (or oscillating) + # stop converging but run at least to next validation + # to check overfitting or ocsillation + # the saved weights of the model will be a bit off in that case + 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 + # however, if adaptive_lr is true, try reducing the lr to + # get us out of an oscilliation + if adaptive_lr==1: + classifier.lr.value=classifier.lr.value/2.0 + test_score = 0. + #cap the patience so we are allowed one more validation error + #calculation before aborting + patience = iter+validation_frequency+1 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 ' validation error is going up, possibly stopping soon' 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 - + if iter>patience: + print 'we have diverged' + break + end_time = time.clock() if verbose == True: @@ -391,17 +431,25 @@ '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) + print iter + + #save the model and the weights + numpy.savez('model.npy', config=configuration, W1=classifier.W1.value,W2=classifier.W2.value, b1=classifier.b1.value,b2=classifier.b2.value) + numpy.savez('results.npy',config=configuration,total_train_error_list=total_train_error_list,total_validation_error_list=total_validation_error_list,\ + learning_rate_list=learning_rate_list) + + return (best_training_error*100.0,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,\ + (train_error,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) + nb_hidden=state.nb_hidden,\ + adaptive_lr=state.adaptive_lr) + state.train_error=train_error state.validation_error=validation_error state.test_error=test_error state.nb_exemples=nb_exemples diff -r d1bb6e06497a -r e81241cfc2de conv_mlp/convolutional_mlp.conf --- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/conv_mlp/convolutional_mlp.conf Thu Feb 25 09:05:48 2010 -0500 @@ -0,0 +1,7 @@ +learning_rate=0.01 +n_iter=1 +batch_size=20 +n_kern0=20 +n_kern1=50 +filter_shape=5 +n_layer=3 \ No newline at end of file diff -r d1bb6e06497a -r e81241cfc2de conv_mlp/convolutional_mlp.py --- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/conv_mlp/convolutional_mlp.py Thu Feb 25 09:05:48 2010 -0500 @@ -0,0 +1,472 @@ +""" +This tutorial introduces the LeNet5 neural network architecture using Theano. LeNet5 is a +convolutional neural network, good for classifying images. This tutorial shows how to build the +architecture, and comes with all the hyper-parameters you need to reproduce the paper's MNIST +results. + +The best results are obtained after X iterations of the main program loop, which takes *** +minutes on my workstation (an Intel Core i7, circa July 2009), and *** minutes on my GPU (an +NVIDIA GTX 285 graphics processor). + +This implementation simplifies the model in the following ways: + + - LeNetConvPool doesn't implement location-specific gain and bias parameters + - LeNetConvPool doesn't implement pooling by average, it implements pooling by max. + - Digit classification is implemented with a logistic regression rather than an RBF network + - LeNet5 was not fully-connected convolutions at second layer + +References: + - Y. LeCun, L. Bottou, Y. Bengio and P. Haffner: Gradient-Based Learning Applied to Document + Recognition, Proceedings of the IEEE, 86(11):2278-2324, November 1998. + http://yann.lecun.com/exdb/publis/pdf/lecun-98.pdf +""" + +import numpy, theano, cPickle, gzip, time +import theano.tensor as T +import theano.sandbox.softsign +import pylearn.datasets.MNIST +from pylearn.io import filetensor as ft +from theano.sandbox import conv, downsample + +class LeNetConvPoolLayer(object): + + def __init__(self, rng, input, filter_shape, image_shape, poolsize=(2,2)): + """ + Allocate a LeNetConvPoolLayer with shared variable internal parameters. + :type rng: numpy.random.RandomState + :param rng: a random number generator used to initialize weights + :type input: theano.tensor.dtensor4 + :param input: symbolic image tensor, of shape image_shape + :type filter_shape: tuple or list of length 4 + :param filter_shape: (number of filters, num input feature maps, + filter height,filter width) + :type image_shape: tuple or list of length 4 + :param image_shape: (batch size, num input feature maps, + image height, image width) + :type poolsize: tuple or list of length 2 + :param poolsize: the downsampling (pooling) factor (#rows,#cols) + """ + assert image_shape[1]==filter_shape[1] + self.input = input + + # initialize weight values: the fan-in of each hidden neuron is + # restricted by the size of the receptive fields. + fan_in = numpy.prod(filter_shape[1:]) + W_values = numpy.asarray( rng.uniform( \ + low = -numpy.sqrt(3./fan_in), \ + high = numpy.sqrt(3./fan_in), \ + size = filter_shape), dtype = theano.config.floatX) + self.W = theano.shared(value = W_values) + + # the bias is a 1D tensor -- one bias per output feature map + b_values = numpy.zeros((filter_shape[0],), dtype= theano.config.floatX) + self.b = theano.shared(value= b_values) + + # convolve input feature maps with filters + conv_out = conv.conv2d(input, self.W, + filter_shape=filter_shape, image_shape=image_shape) + + # downsample each feature map individually, using maxpooling + pooled_out = downsample.max_pool2D(conv_out, poolsize, ignore_border=True) + + # add the bias term. Since the bias is a vector (1D array), we first + # reshape it to a tensor of shape (1,n_filters,1,1). Each bias will thus + # be broadcasted across mini-batches and feature map width & height + self.output = T.tanh(pooled_out + self.b.dimshuffle('x', 0, 'x', 'x')) + + # store parameters of this layer + self.params = [self.W, self.b] + + +class SigmoidalLayer(object): + def __init__(self, rng, input, n_in, n_out): + """ + Typical hidden layer of a MLP: units are fully-connected and have + sigmoidal activation function. Weight matrix W is of shape (n_in,n_out) + and the bias vector b is of shape (n_out,). + + Hidden unit activation is given by: sigmoid(dot(input,W) + b) + + :type rng: numpy.random.RandomState + :param rng: a random number generator used to initialize weights + :type input: theano.tensor.dmatrix + :param input: a symbolic tensor of shape (n_examples, n_in) + :type n_in: int + :param n_in: dimensionality of input + :type n_out: int + :param n_out: number of hidden units + """ + self.input = input + + W_values = numpy.asarray( rng.uniform( \ + low = -numpy.sqrt(6./(n_in+n_out)), \ + high = numpy.sqrt(6./(n_in+n_out)), \ + size = (n_in, n_out)), dtype = theano.config.floatX) + self.W = theano.shared(value = W_values) + + b_values = numpy.zeros((n_out,), dtype= theano.config.floatX) + self.b = theano.shared(value= b_values) + + self.output = T.tanh(T.dot(input, self.W) + self.b) + self.params = [self.W, self.b] + + +class LogisticRegression(object): + """Multi-class Logistic Regression Class + + The logistic regression is fully described by a weight matrix :math:`W` + and bias vector :math:`b`. Classification is done by projecting data + points onto a set of hyperplanes, the distance to which is used to + determine a class membership probability. + """ + + def __init__(self, input, n_in, n_out): + """ Initialize the parameters of the logistic regression + :param input: symbolic variable that describes the input of the + architecture (one minibatch) + :type n_in: int + :param n_in: number of input units, the dimension of the space in + which the datapoints lie + :type n_out: int + :param n_out: number of output units, the dimension of the space in + which the labels lie + """ + + # initialize with 0 the weights W as a matrix of shape (n_in, n_out) + self.W = theano.shared( value=numpy.zeros((n_in,n_out), + dtype = theano.config.floatX) ) + # initialize the baises b as a vector of n_out 0s + self.b = theano.shared( value=numpy.zeros((n_out,), + dtype = theano.config.floatX) ) + # compute vector of class-membership probabilities in symbolic form + self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W)+self.b) + + # compute prediction as class whose probability is maximal in + # symbolic form + self.y_pred=T.argmax(self.p_y_given_x, axis=1) + + # list of parameters for this layer + self.params = [self.W, self.b] + + 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. + :param y: corresponds to a vector that gives for each example the + correct label + Note: we use the mean instead of the sum so that + the learning rate is less dependent on the batch size + """ + 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 ; zero one + loss over the size 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 load_dataset(fname,batch=20): + + # repertoire qui contient les donnees NIST + # le repertoire suivant va fonctionner si vous etes connecte sur un ordinateur + # du reseau DIRO + datapath = '/data/lisa/data/nist/by_class/' + # le fichier .ft contient chiffres NIST dans un format efficace. Les chiffres + # sont stockes dans une matrice de NxD, ou N est le nombre d'images, est D est + # le nombre de pixels par image (32x32 = 1024). Chaque pixel de l'image est une + # valeur entre 0 et 255, correspondant a un niveau de gris. Les valeurs sont + # stockees comme des uint8, donc des bytes. + f = open(datapath+'digits/digits_train_data.ft') + # Verifier que vous avez assez de memoire pour loader les donnees au complet + # dans le memoire. Sinon, utilisez ft.arraylike, une classe construite + # specialement pour des fichiers qu'on ne souhaite pas loader dans RAM. + d = ft.read(f) + + # NB: N'oubliez pas de diviser les valeurs des pixels par 255. si jamais vous + # utilisez les donnees commes entrees dans un reseaux de neurones et que vous + # voulez des entres entre 0 et 1. + # digits_train_data.ft contient les images, digits_train_labels.ft contient les + # etiquettes + f = open(datapath+'digits/digits_train_labels.ft') + labels = ft.read(f) + + + # Load the dataset + #f = gzip.open(fname,'rb') + #train_set, valid_set, test_set = cPickle.load(f) + #f.close() + + # make minibatches of size 20 + batch_size = batch # sized of the minibatch + + # Dealing with the training set + # get the list of training images (x) and their labels (y) + (train_set_x, train_set_y) = (d[:4000,:],labels[:4000]) + # initialize the list of training minibatches with empty list + train_batches = [] + for i in xrange(0, len(train_set_x), batch_size): + # add to the list of minibatches the minibatch starting at + # position i, ending at position i+batch_size + # a minibatch is a pair ; the first element of the pair is a list + # of datapoints, the second element is the list of corresponding + # labels + train_batches = train_batches + \ + [(train_set_x[i:i+batch_size], train_set_y[i:i+batch_size])] + + #print train_batches[500] + + # Dealing with the validation set + (valid_set_x, valid_set_y) = (d[4000:5000,:],labels[4000:5000]) + # initialize the list of validation minibatches + valid_batches = [] + for i in xrange(0, len(valid_set_x), batch_size): + valid_batches = valid_batches + \ + [(valid_set_x[i:i+batch_size], valid_set_y[i:i+batch_size])] + + # Dealing with the testing set + (test_set_x, test_set_y) = (d[5000:6000,:],labels[5000:6000]) + # initialize the list of testing minibatches + test_batches = [] + for i in xrange(0, len(test_set_x), batch_size): + test_batches = test_batches + \ + [(test_set_x[i:i+batch_size], test_set_y[i:i+batch_size])] + + return train_batches, valid_batches, test_batches + + +def evaluate_lenet5(learning_rate=0.1, n_iter=1, batch_size=20, n_kern0=20,n_kern1=50,filter_shape=5,n_layer=3, dataset='mnist.pkl.gz'): + rng = numpy.random.RandomState(23455) + + print 'Before load dataset' + train_batches, valid_batches, test_batches = load_dataset(dataset,batch_size) + print 'After load dataset' + + ishape = (32,32) # this is the size of NIST images + n_kern2=80 + + # allocate symbolic variables for the data + x = T.matrix('x') # rasterized images + y = T.lvector() # the labels are presented as 1D vector of [long int] labels + + + ###################### + # BUILD ACTUAL MODEL # + ###################### + + # Reshape matrix of rasterized images of shape (batch_size,28*28) + # to a 4D tensor, compatible with our LeNetConvPoolLayer + layer0_input = x.reshape((batch_size,1,32,32)) + + # Construct the first convolutional pooling layer: + # filtering reduces the image size to (32-5+1,32-5+1)=(28,28) + # maxpooling reduces this further to (28/2,28/2) = (14,14) + # 4D output tensor is thus of shape (20,20,14,14) + layer0 = LeNetConvPoolLayer(rng, input=layer0_input, + image_shape=(batch_size,1,32,32), + filter_shape=(n_kern0,1,filter_shape,filter_shape), poolsize=(2,2)) + + if(n_layer>2): + + # Construct the second convolutional pooling layer + # filtering reduces the image size to (14-5+1,14-5+1)=(10,10) + # maxpooling reduces this further to (10/2,10/2) = (5,5) + # 4D output tensor is thus of shape (20,50,5,5) + fshape=(32-filter_shape+1)/2 + layer1 = LeNetConvPoolLayer(rng, input=layer0.output, + image_shape=(batch_size,n_kern0,fshape,fshape), + filter_shape=(n_kern1,n_kern0,filter_shape,filter_shape), poolsize=(2,2)) + + else: + + fshape=(32-filter_shape+1)/2 + layer1_input = layer0.output.flatten(2) + # construct a fully-connected sigmoidal layer + layer1 = SigmoidalLayer(rng, input=layer1_input,n_in=n_kern0*fshape*fshape, n_out=500) + + layer2 = LogisticRegression(input=layer1.output, n_in=500, n_out=10) + cost = layer2.negative_log_likelihood(y) + test_model = theano.function([x,y], layer2.errors(y)) + params = layer2.params+ layer1.params + layer0.params + + + if(n_layer>3): + + fshape=(32-filter_shape+1)/2 + fshape2=(fshape-filter_shape+1)/2 + fshape3=(fshape2-filter_shape+1)/2 + layer2 = LeNetConvPoolLayer(rng, input=layer1.output, + image_shape=(batch_size,n_kern1,fshape2,fshape2), + filter_shape=(n_kern2,n_kern1,filter_shape,filter_shape), poolsize=(2,2)) + + layer3_input = layer2.output.flatten(2) + + layer3 = SigmoidalLayer(rng, input=layer3_input, + n_in=n_kern2*fshape3*fshape3, n_out=500) + + + layer4 = LogisticRegression(input=layer3.output, n_in=500, n_out=10) + + cost = layer4.negative_log_likelihood(y) + + test_model = theano.function([x,y], layer4.errors(y)) + + params = layer4.params+ layer3.params+ layer2.params+ layer1.params + layer0.params + + + elif(n_layer>2): + + fshape=(32-filter_shape+1)/2 + fshape2=(fshape-filter_shape+1)/2 + + # the SigmoidalLayer being fully-connected, it operates on 2D matrices of + # shape (batch_size,num_pixels) (i.e matrix of rasterized images). + # This will generate a matrix of shape (20,32*4*4) = (20,512) + layer2_input = layer1.output.flatten(2) + + # construct a fully-connected sigmoidal layer + layer2 = SigmoidalLayer(rng, input=layer2_input, + n_in=n_kern1*fshape2*fshape2, n_out=500) + + + # classify the values of the fully-connected sigmoidal layer + layer3 = LogisticRegression(input=layer2.output, n_in=500, n_out=10) + + # the cost we minimize during training is the NLL of the model + cost = layer3.negative_log_likelihood(y) + + # create a function to compute the mistakes that are made by the model + test_model = theano.function([x,y], layer3.errors(y)) + + # create a list of all model parameters to be fit by gradient descent + params = layer3.params+ layer2.params+ layer1.params + layer0.params + + + + + + # create a list of gradients for all model parameters + grads = T.grad(cost, params) + + # train_model is a function that updates the model parameters by SGD + # Since this model has many parameters, it would be tedious to manually + # create an update rule for each model parameter. We thus create the updates + # dictionary by automatically looping over all (params[i],grads[i]) pairs. + updates = {} + for param_i, grad_i in zip(params, grads): + updates[param_i] = param_i - learning_rate * grad_i + train_model = theano.function([x, y], cost, updates=updates) + + + ############### + # TRAIN MODEL # + ############### + + n_minibatches = len(train_batches) + + # early-stopping parameters + patience = 10000 # look as this many examples regardless + 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 # go through this many + # minibatche before checking the network + # on the validation set; in this case we + # check every epoch + + best_params = None + best_validation_loss = float('inf') + best_iter = 0 + test_score = 0. + start_time = time.clock() + + # have a maximum of `n_iter` iterations through the entire dataset + 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 ] + + if iter %100 == 0: + print 'training @ iter = ', iter + cost_ij = train_model(x,y) + + if (iter+1) % validation_frequency == 0: + + # compute zero-one loss on validation set + this_validation_loss = 0. + for x,y in valid_batches: + # sum up the errors for each minibatch + this_validation_loss += test_model(x,y) + + # get the average by dividing with the number of minibatches + this_validation_loss /= len(valid_batches) + 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) + + # 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: + test_score += test_model(x,y) + test_score /= len(test_batches) + print((' epoch %i, minibatch %i/%i, test error of best ' + 'model %f %%') % + (epoch, minibatch_index+1, n_minibatches, + test_score*100.)) + + if patience <= iter : + break + + end_time = time.clock() + print('Optimization complete.') + print('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.)) + + return (best_validation_loss * 100., test_score*100., (end_time-start_time)/60., best_iter) + +if __name__ == '__main__': + evaluate_lenet5() + +def experiment(state, channel): + print 'start experiment' + (best_validation_loss, test_score, minutes_trained, iter) = evaluate_lenet5(state.learning_rate, state.n_iter, state.batch_size, state.n_kern0, state.n_kern1, state.filter_shape, state.n_layer) + print 'end experiment' + + state.best_validation_loss = best_validation_loss + state.test_score = test_score + state.minutes_trained = minutes_trained + state.iter = iter + + return channel.COMPLETE diff -r d1bb6e06497a -r e81241cfc2de scripts/creer_jeu_occlusion.py --- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/scripts/creer_jeu_occlusion.py Thu Feb 25 09:05:48 2010 -0500 @@ -0,0 +1,42 @@ +#!/usr/bin/python +# coding: utf-8 + +''' +Sert a creer un petit jeu de donnees afin de pouvoir avoir des fragments +de lettres pour ajouter bruit d'occlusion + +Sylvain Pannetier Lebeuf dans le cadre de IFT6266, hiver 2010 + +''' + +from pylearn.io import filetensor as ft +import pylab +import random as r +from numpy import * + +nombre = 20000 #La grandeur de l'echantillon + +f = open('all_train_data.ft') #Le jeu de donnees est en local. +d = ft.read(f) +f.close() +print len(d) +random.seed(3525) + +echantillon=r.sample(xrange(len(d)),nombre) +nouveau=d[0:nombre] +for i in xrange(nombre): + nouveau[i]=d[echantillon[i]] + + +f2 = open('echantillon_occlusion.ft', 'w') +ft.write(f2,nouveau) +f2.close() + + +##Tester si ca a fonctionne +f3 = open('echantillon_occlusion.ft') + +d2=ft.read(f3) +pylab.imshow(d2[0].reshape((32,32))) +pylab.show() +f3.close() \ No newline at end of file diff -r d1bb6e06497a -r e81241cfc2de scripts/fonts_test.py --- a/scripts/fonts_test.py Thu Feb 25 09:04:40 2010 -0500 +++ b/scripts/fonts_test.py Thu Feb 25 09:05:48 2010 -0500 @@ -1,15 +1,19 @@ +#!/usr/bin/python -import random, os +import os import ImageFont, ImageDraw, Image dir1 = "/data/lisa/data/ift6266h10/allfonts/" +#dir1 = "/Tmp/allfonts/" -img = Image.new("L", (100,100)) +img = Image.new("L", (132,132)) draw = ImageDraw.Draw(img) +L = [chr(ord('0')+x) for x in range(10)] + [chr(ord('A')+x) for x in range(26)] + [chr(ord('a')+x) for x in range(26)] for f in os.listdir(dir1): try: font = ImageFont.truetype(dir1+f, 25) - textSize = font.getsize("kD9") + for l in L: + draw.text((60,60), l, font=font, fill="white") except: print dir1+f diff -r d1bb6e06497a -r e81241cfc2de scripts/imgbg_test.py --- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/scripts/imgbg_test.py Thu Feb 25 09:05:48 2010 -0500 @@ -0,0 +1,15 @@ +#!/usr/bin/python + +import Image, cPickle + +f=open('/Tmp/image_net/filelist.pkl') +image_files = cPickle.load(f) +f.close() + +for i in range(len(image_files)): + filename = '/Tmp/image_net/' + image_files[i] + try: + image = Image.open(filename).convert('L') + except: + print filename + diff -r d1bb6e06497a -r e81241cfc2de scripts/launch_generate100.py --- a/scripts/launch_generate100.py Thu Feb 25 09:04:40 2010 -0500 +++ b/scripts/launch_generate100.py Thu Feb 25 09:05:48 2010 -0500 @@ -4,9 +4,9 @@ dir1 = "/data/lisa/data/ift6266h10/" for i,s in enumerate(['valid','test']): - for c in [0.3,0.5,0.7,1]: + for j,c in enumerate([0.3,0.5,0.7,1]): l = str(c).replace('.','') - os.system("dbidispatch --condor --os=fc9 --machine=brams0c.iro.umontreal.ca ./run_pipeline.sh -o %sdata/P%s_%s_data.ft -p %sdata/P%s_%s_params -x %sdata/P%s_%s_labels.ft -f %s%s_data.ft -l %s%s_labels.ft -c %socr_%s_data.ft -d %socr_%s_labels.ft -m 0.3 -z 0.1 -a 0.1 -b 0.25 -g 0.25 -s %d" % (dir1, l, s, dir1, l, s, dir1, l, s, dir1, s, dir1, s, dir1, s, dir1, s, [20000,80000][i])) + os.system("dbidispatch --condor --os=fc9 --machine=brams0c.iro.umontreal.ca ./run_pipeline.sh -o %sdata/P%s_%s_data.ft -p %sdata/P%s_%s_params -x %sdata/P%s_%s_labels.ft -f %s%s_data.ft -l %s%s_labels.ft -c %socr_%s_data.ft -d %socr_%s_labels.ft -m 0.3 -z 0.1 -a 0.1 -b 0.25 -g 0.25 -s %d -y %d" % (dir1, l, s, dir1, l, s, dir1, l, s, dir1, s, dir1, s, dir1, s, dir1, s, [20000,80000][i], 200+i*4+j)) for i in range(100): - os.system("dbidispatch --condor --os=fc9 --machine=brams0c.iro.umontreal.ca ./run_pipeline.sh -o %sdata/P07_train%d_data.ft -p %sdata/P07_train%d_params -x %sdata/P07_train%d_labels.ft -f %strain_data.ft -l %strain_labels.ft -c %socr_train_data.ft -d %socr_train_labels.ft -m 0.7 -z 0.1 -a 0.1 -b 0.25 -g 0.25 -s 819200" % (dir1, i, dir1, i, dir1, i, dir1, dir1, dir1, dir1)) + os.system("dbidispatch --condor --os=fc9 --machine=brams0c.iro.umontreal.ca ./run_pipeline.sh -o %sdata/P07_train%d_data.ft -p %sdata/P07_train%d_params -x %sdata/P07_train%d_labels.ft -f %strain_data.ft -l %strain_labels.ft -c %socr_train_data.ft -d %socr_train_labels.ft -m 0.7 -z 0.1 -a 0.1 -b 0.25 -g 0.25 -s 819200 -y %d" % (dir1, i, dir1, i, dir1, i, dir1, dir1, dir1, dir1, 100+i)) diff -r d1bb6e06497a -r e81241cfc2de transformations/BruitGauss.py --- a/transformations/BruitGauss.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/BruitGauss.py Thu Feb 25 09:05:48 2010 -0500 @@ -22,17 +22,23 @@ ''' import numpy -import random +#import random import scipy from scipy import ndimage class BruitGauss(): - def __init__(self,complexity=1): + def __init__(self,complexity=1,seed=6378): self.nb_chngmax =10 #Le nombre de pixels changes. Seulement pour fin de calcul self.grandeurmax = 20 self.sigmamax = 6.0 self.regenerate_parameters(complexity) + self.seed=seed + + #numpy.random.seed(self.seed) + + def get_seed(self): + return self.seed def get_settings_names(self): return ['nb_chng','sigma_gauss','grandeur'] diff -r d1bb6e06497a -r e81241cfc2de transformations/DistorsionGauss.py --- a/transformations/DistorsionGauss.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/DistorsionGauss.py Thu Feb 25 09:05:48 2010 -0500 @@ -16,13 +16,21 @@ class DistorsionGauss(): - def __init__(self): + def __init__(self,seed=3459): self.ecart_type=0.1 #L'ecart type de la gaussienne self.effectuer=1 #1=on effectue et 0=rien faire + self.seed=seed + + #Les deux generateurs sont de types differents, avoir la meme seed n'a pas d'influence + #numpy.random.seed(self.seed) + #random.seed(self.seed) def get_settings_names(self): return ['effectuer'] + def get_seed(self): + return self.seed + def get_settings_names_determined_by_complexity(self,complexity): return ['ecart_type'] diff -r d1bb6e06497a -r e81241cfc2de transformations/Occlusion.py --- a/transformations/Occlusion.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/Occlusion.py Thu Feb 25 09:05:48 2010 -0500 @@ -24,7 +24,7 @@ class Occlusion(): - def __init__(self): + def __init__(self,seed=9854): #Ces 4 variables representent la taille du "crop" sur l'image2 #Ce "crop" est pris a partie de image1[15,15], le milieu de l'image1 self.haut=2 @@ -47,6 +47,9 @@ #Sert a dire si on fait quelque chose. 0=faire rien, 1 on fait quelque chose self.appliquer=1 + self.seed=seed + #numpy.random.seed(self.seed) + f3 = open('/data/lisa/data/ift6266h10/echantillon_occlusion.ft') #Doit etre sur le reseau DIRO. #f3 = open('/home/sylvain/Dropbox/Msc/IFT6266/donnees/echantillon_occlusion.ft') #Il faut arranger le path sinon @@ -59,6 +62,9 @@ def get_settings_names(self): return ['haut','bas','gauche','droite','x_arrivee','y_arrivee','endroit','rajout','appliquer'] + + def get_seed(self): + return self.seed def regenerate_parameters(self, complexity): self.haut=min(15,int(numpy.abs(numpy.random.normal(int(8*complexity),2)))) @@ -144,6 +150,7 @@ img_trans=transfo.transform_image(img.reshape((32,32))) + print transfo.get_seed() pylab.imshow(img_trans.reshape((32,32))) pylab.show() diff -r d1bb6e06497a -r e81241cfc2de transformations/PermutPixel.py --- a/transformations/PermutPixel.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/PermutPixel.py Thu Feb 25 09:05:48 2010 -0500 @@ -18,10 +18,18 @@ class PermutPixel(): - def __init__(self): + def __init__(self,seed=7152): self.nombre=10 #Le nombre de pixels a permuter self.proportion=0.3 self.effectuer=1 #1=on effectue, 0=rien faire + self.seed=seed + + #Les deux generateurs sont de types differents, avoir la meme seed n'a pas d'influence + #numpy.random.seed(self.seed) + #random.seed(self.seed) + + def get_seed(self): + return self.seed def get_settings_names(self): return ['effectuer'] diff -r d1bb6e06497a -r e81241cfc2de transformations/PoivreSel.py --- a/transformations/PoivreSel.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/PoivreSel.py Thu Feb 25 09:05:48 2010 -0500 @@ -22,11 +22,19 @@ class PoivreSel(): - def __init__(self): + def __init__(self,seed=9361): self.proportion_bruit=0.08 #Le pourcentage des pixels qui seront bruites self.nb_chng=10 #Le nombre de pixels changes. Seulement pour fin de calcul self.effectuer=1 #Vaut 1 si on effectue et 0 sinon. + self.seed=seed + #Les deux generateurs sont de types differents, avoir la meme seed n'a pas d'influence + #numpy.random.seed(self.seed) + #random.seed(self.seed) + + def get_seed(self): + return self.seed + def get_settings_names(self): return ['effectuer'] diff -r d1bb6e06497a -r e81241cfc2de transformations/Rature.py --- a/transformations/Rature.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/Rature.py Thu Feb 25 09:05:48 2010 -0500 @@ -23,7 +23,7 @@ class Rature(): - def __init__(self): + def __init__(self,seed=1256): self.angle=0 #Angle en degre de la rotation (entre 0 et 180) self.numero=0 #Le numero du 1 choisi dans la banque de 1 self.gauche=-1 #Le numero de la colonne la plus a gauche contenant le 1 @@ -41,6 +41,9 @@ self.nb_ratures=-1 #Le nombre de ratures appliques self.fini=0 #1=fini de mettre toutes les couches 0=pas fini self.complexity=0 #Pour garder en memoire la complexite si plusieurs couches sont necessaires + self.seed=seed + + #numpy.random.seed(self.seed) f3 = open('/data/lisa/data/ift6266h10/un_rature.ft') #Doit etre sur le reseau DIRO. #f3 = open('/home/sylvain/Dropbox/Msc/IFT6266/donnees/un_rature.ft') @@ -53,6 +56,9 @@ def get_settings_names(self): return ['angle','numero','faire','crop_haut','crop_gauche','largeur_bande','smooth','nb_ratures'] + + def get_seed(self): + return self.seed def regenerate_parameters(self, complexity,next_rature = False): diff -r d1bb6e06497a -r e81241cfc2de transformations/affine_transform.py --- a/transformations/affine_transform.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/affine_transform.py Thu Feb 25 09:05:48 2010 -0500 @@ -16,9 +16,8 @@ class AffineTransformation(): def __init__( self, complexity = .5): self.shape = (32,32) - self.rng = numpy.random.RandomState() self.complexity = complexity - params = self.rng.uniform(size=6) -.5 + params = numpy.random.uniform(size=6) -.5 self.a = 1. + params[0]*.6*complexity self.b = 0. + params[1]*.6*complexity self.c = params[2]*8.*complexity @@ -43,7 +42,7 @@ # the ranges are set manually as to look acceptable self.complexity = complexity - params = self.rng.uniform(size=6) -.5 + params = numpy.random.uniform(size=6) -.5 self.a = 1. + params[0]*.8*complexity self.b = 0. + params[1]*.8*complexity self.c = params[2]*9.*complexity diff -r d1bb6e06497a -r e81241cfc2de transformations/pipeline.py --- a/transformations/pipeline.py Thu Feb 25 09:04:40 2010 -0500 +++ b/transformations/pipeline.py Thu Feb 25 09:05:48 2010 -0500 @@ -37,6 +37,45 @@ DEFAULT_OCRLABEL_PATH = '/data/lisa/data/ocr_breuel/filetensor/unlv-corrected-2010-02-01-labels-shuffled.ft' ARGS_FILE = os.environ['PIPELINE_ARGS_TMPFILE'] +# PARSE COMMAND LINE ARGUMENTS +def get_argv(): + with open(ARGS_FILE) as f: + args = [l.rstrip() for l in f.readlines()] + return args + +def usage(): + print ''' +Usage: run_pipeline.sh [-m ...] [-z ...] [-o ...] [-p ...] + -m, --max-complexity: max complexity to generate for an image + -z, --probability-zero: probability of using complexity=0 for an image + -o, --output-file: full path to file to use for output of images + -p, --params-output-file: path to file to output params to + -x, --labels-output-file: path to file to output labels to + -f, --data-file: path to filetensor (.ft) data file (NIST) + -l, --label-file: path to filetensor (.ft) labels file (NIST labels) + -c, --ocr-file: path to filetensor (.ft) data file (OCR) + -d, --ocrlabel-file: path to filetensor (.ft) labels file (OCR labels) + -a, --prob-font: probability of using a raw font image + -b, --prob-captcha: probability of using a captcha image + -g, --prob-ocr: probability of using an ocr image + -y, --seed: the job seed + ''' + +try: + opts, args = getopt.getopt(get_argv(), "rm:z:o:p:x:s:f:l:c:d:a:b:g:y:", ["reload","max-complexity=", "probability-zero=", "output-file=", "params-output-file=", "labels-output-file=", +"stop-after=", "data-file=", "label-file=", "ocr-file=", "ocrlabel-file=", "prob-font=", "prob-captcha=", "prob-ocr=", "seed="]) +except getopt.GetoptError, err: + # print help information and exit: + print str(err) # will print something like "option -a not recognized" + usage() + pdb.gimp_quit(0) + sys.exit(2) + +for o, a in opts: + if o in ('-y','--seed'): + random.seed(int(a)) + numpy.random.seed(int(a)) + if DEBUG_X: import pylab pylab.ion() @@ -55,7 +94,7 @@ from add_background_image import AddBackground from affine_transform import AffineTransformation from ttf2jpg import ttf2jpg -from pycaptcha.Facade import generateCaptcha +from Facade import generateCaptcha if DEBUG: from visualizer import Visualizer @@ -254,28 +293,6 @@ ############################################################################## # MAIN -def usage(): - print ''' -Usage: run_pipeline.sh [-m ...] [-z ...] [-o ...] [-p ...] - -m, --max-complexity: max complexity to generate for an image - -z, --probability-zero: probability of using complexity=0 for an image - -o, --output-file: full path to file to use for output of images - -p, --params-output-file: path to file to output params to - -x, --labels-output-file: path to file to output labels to - -f, --data-file: path to filetensor (.ft) data file (NIST) - -l, --label-file: path to filetensor (.ft) labels file (NIST labels) - -c, --ocr-file: path to filetensor (.ft) data file (OCR) - -d, --ocrlabel-file: path to filetensor (.ft) labels file (OCR labels) - -a, --prob-font: probability of using a raw font image - -b, --prob-captcha: probability of using a captcha image - -g, --prob-ocr: probability of using an ocr image - ''' - -# See run_pipeline.py -def get_argv(): - with open(ARGS_FILE) as f: - args = [l.rstrip() for l in f.readlines()] - return args # Might be called locally or through dbidispatch. In all cases it should be # passed to the GIMP executable to be able to use GIMP filters. @@ -299,16 +316,6 @@ stop_after = None reload_mode = False - try: - opts, args = getopt.getopt(get_argv(), "rm:z:o:p:x:s:f:l:c:d:a:b:g:", ["reload","max-complexity=", "probability-zero=", "output-file=", "params-output-file=", "labels-output-file=", -"stop-after=", "data-file=", "label-file=", "ocr-file=", "ocrlabel-file=", "prob-font=", "prob-captcha=", "prob-ocr="]) - except getopt.GetoptError, err: - # print help information and exit: - print str(err) # will print something like "option -a not recognized" - usage() - pdb.gimp_quit(0) - sys.exit(2) - for o, a in opts: if o in ('-m', '--max-complexity'): max_complexity = float(a) @@ -340,6 +347,8 @@ prob_captcha = float(a) elif o in ('-g', "--prob-ocr"): prob_ocr = float(a) + elif o in ('-y', "--seed"): + pass else: assert False, "unhandled option"