# HG changeset patch
# User humel
# Date 1271911782 14400
# Node ID 31641a84e0ae3cba3fb68bba6ec1345cb616b181
# Parent  9a7b74927f7d9c74df77e6ddf403e1e59a6a5029
Initial commit for the experimental setup of the denoising convolutional network

diff -r 9a7b74927f7d -r 31641a84e0ae deep/convolutional_dae/salah_exp/nist_csda.py
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/deep/convolutional_dae/salah_exp/nist_csda.py	Thu Apr 22 00:49:42 2010 -0400
@@ -0,0 +1,266 @@
+#!/usr/bin/python
+# coding: utf-8
+
+import ift6266
+import pylearn
+
+import numpy 
+import theano
+import time
+
+import pylearn.version
+import theano.tensor as T
+from theano.tensor.shared_randomstreams import RandomStreams
+
+import copy
+import sys
+import os
+import os.path
+
+from jobman import DD
+import jobman, jobman.sql
+from pylearn.io import filetensor
+
+from utils import produit_cartesien_jobs
+from copy import copy
+
+from sgd_optimization_new import CSdASgdOptimizer
+
+#from ift6266.utils.scalar_series import *
+from ift6266.utils.seriestables import *
+import tables
+
+from ift6266 import datasets
+from config import *
+
+'''
+Function called by jobman upon launching each job
+Its path is the one given when inserting jobs: see EXPERIMENT_PATH
+'''
+def jobman_entrypoint(state, channel):
+    # record mercurial versions of each package
+    pylearn.version.record_versions(state,[theano,ift6266,pylearn])
+    # TODO: remove this, bad for number of simultaneous requests on DB
+    channel.save()
+
+    # For test runs, we don't want to use the whole dataset so
+    # reduce it to fewer elements if asked to.
+    rtt = None
+    #REDUCE_TRAIN_TO = 40000
+    if state.has_key('reduce_train_to'):
+        rtt = state['reduce_train_to']
+    elif REDUCE_TRAIN_TO:
+        rtt = REDUCE_TRAIN_TO
+        
+    if state.has_key('decrease_lr'):
+        decrease_lr = state['decrease_lr']
+    else :
+        decrease_lr = 0
+ 
+    n_ins = 32*32
+    n_outs = 62 # 10 digits, 26*2 (lower, capitals)
+     
+    examples_per_epoch = 100000#NIST_ALL_TRAIN_SIZE
+    
+    #To be sure variables will not be only in the if statement
+    PATH = ''
+    nom_reptrain = ''
+    nom_serie = ""
+    if state['pretrain_choice'] == 0:
+        nom_serie="series_NIST.h5"
+    elif state['pretrain_choice'] == 1:
+        nom_serie="series_P07.h5"
+
+    series = create_series(state.num_hidden_layers,nom_serie)
+
+
+    print "Creating optimizer with state, ", state
+
+    optimizer = CSdASgdOptimizer(dataset=datasets.nist_all(), 
+                                    hyperparameters=state, \
+                                    n_ins=n_ins, n_outs=n_outs,\
+                                    examples_per_epoch=examples_per_epoch, \
+                                    series=series,
+                                    max_minibatches=rtt)
+
+    parameters=[]
+    #Number of files of P07 used for pretraining
+    nb_file=0
+    if state['pretrain_choice'] == 0:
+        print('\n\tpretraining with NIST\n')
+        optimizer.pretrain(datasets.nist_all()) 
+    elif state['pretrain_choice'] == 1:
+        #To know how many file will be used during pretraining
+        nb_file = int(state['pretraining_epochs_per_layer']) 
+        state['pretraining_epochs_per_layer'] = 1 #Only 1 time over the dataset
+        if nb_file >=100:
+            sys.exit("The code does not support this much pretraining epoch (99 max with P07).\n"+
+            "You have to correct the code (and be patient, P07 is huge !!)\n"+
+             "or reduce the number of pretraining epoch to run the code (better idea).\n")
+        print('\n\tpretraining with P07')
+        optimizer.pretrain(datasets.nist_P07(min_file=0,max_file=nb_file)) 
+    channel.save()
+    
+    #Set some of the parameters used for the finetuning
+    if state.has_key('finetune_set'):
+        finetune_choice=state['finetune_set']
+    else:
+        finetune_choice=FINETUNE_SET
+    
+    if state.has_key('max_finetuning_epochs'):
+        max_finetune_epoch_NIST=state['max_finetuning_epochs']
+    else:
+        max_finetune_epoch_NIST=MAX_FINETUNING_EPOCHS
+    
+    if state.has_key('max_finetuning_epochs_P07'):
+        max_finetune_epoch_P07=state['max_finetuning_epochs_P07']
+    else:
+        max_finetune_epoch_P07=max_finetune_epoch_NIST
+    
+    #Decide how the finetune is done
+    
+    if finetune_choice == 0:
+        print('\n\n\tfinetune with NIST\n\n')
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,decrease=decrease_lr)
+        channel.save()
+    if finetune_choice == 1:
+        print('\n\n\tfinetune with P07\n\n')
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=0,decrease=decrease_lr)
+        channel.save()
+    if finetune_choice == 2:
+        print('\n\n\tfinetune with P07 followed by NIST\n\n')
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=20,decrease=decrease_lr)
+        optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=21,decrease=decrease_lr)
+        channel.save()
+    if finetune_choice == 3:
+        print('\n\n\tfinetune with NIST only on the logistic regression on top (but validation on P07).\n\
+        All hidden units output are input of the logistic regression\n\n')
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,special=1,decrease=decrease_lr)
+        
+        
+    if finetune_choice==-1:
+        print('\nSERIE OF 4 DIFFERENT FINETUNINGS')
+        print('\n\n\tfinetune with NIST\n\n')
+        sys.stdout.flush()
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,decrease=decrease_lr)
+        channel.save()
+        print('\n\n\tfinetune with P07\n\n')
+        sys.stdout.flush()
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_P07(),datasets.nist_all(),max_finetune_epoch_P07,ind_test=0,decrease=decrease_lr)
+        channel.save()
+        print('\n\n\tfinetune with P07 (done earlier) followed by NIST (written here)\n\n')
+        sys.stdout.flush()
+        optimizer.reload_parameters('params_finetune_P07.txt')
+        optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=21,decrease=decrease_lr)
+        channel.save()
+        print('\n\n\tfinetune with NIST only on the logistic regression on top.\n\
+        All hidden units output are input of the logistic regression\n\n')
+        sys.stdout.flush()
+        optimizer.reload_parameters('params_pretrain.txt')
+        optimizer.finetune(datasets.nist_all(),datasets.nist_P07(),max_finetune_epoch_NIST,ind_test=1,special=1,decrease=decrease_lr)
+        channel.save()
+    
+    channel.save()
+
+    return channel.COMPLETE
+
+# These Series objects are used to save various statistics
+# during the training.
+def create_series(num_hidden_layers, nom_serie):
+
+    # Replace series we don't want to save with DummySeries, e.g.
+    # series['training_error'] = DummySeries()
+
+    series = {}
+
+    basedir = os.getcwd()
+
+    h5f = tables.openFile(os.path.join(basedir, nom_serie), "w")
+
+    #REDUCE_EVERY=10
+    # reconstruction
+    reconstruction_base = \
+                ErrorSeries(error_name="reconstruction_error",
+                    table_name="reconstruction_error",
+                    hdf5_file=h5f,
+                    index_names=('epoch','minibatch'),
+                    title="Reconstruction error (mean over "+str(REDUCE_EVERY)+" minibatches)")
+    series['reconstruction_error'] = \
+                AccumulatorSeriesWrapper(base_series=reconstruction_base,
+                    reduce_every=REDUCE_EVERY)
+
+    # train
+    training_base = \
+                ErrorSeries(error_name="training_error",
+                    table_name="training_error",
+                    hdf5_file=h5f,
+                    index_names=('epoch','minibatch'),
+                    title="Training error (mean over "+str(REDUCE_EVERY)+" minibatches)")
+    series['training_error'] = \
+                AccumulatorSeriesWrapper(base_series=training_base,
+                    reduce_every=REDUCE_EVERY)
+
+    # valid and test are not accumulated/mean, saved directly
+    series['validation_error'] = \
+                ErrorSeries(error_name="validation_error",
+                    table_name="validation_error",
+                    hdf5_file=h5f,
+                    index_names=('epoch','minibatch'))
+
+    series['test_error'] = \
+                ErrorSeries(error_name="test_error",
+                    table_name="test_error",
+                    hdf5_file=h5f,
+                    index_names=('epoch','minibatch'))
+
+    param_names = []
+    for i in range(num_hidden_layers):
+        param_names += ['layer%d_W'%i, 'layer%d_b'%i, 'layer%d_bprime'%i]
+    param_names += ['logreg_layer_W', 'logreg_layer_b']
+
+    # comment out series we don't want to save
+    series['params'] = SharedParamsStatisticsWrapper(
+                        new_group_name="params",
+                        base_group="/",
+                        arrays_names=param_names,
+                        hdf5_file=h5f,
+                        index_names=('epoch',))
+
+    return series
+
+# Perform insertion into the Postgre DB based on combination
+# of hyperparameter values above
+# (see comment for produit_cartesien_jobs() to know how it works)
+def jobman_insert_nist():
+    jobs = produit_cartesien_jobs(JOB_VALS)
+
+    db = jobman.sql.db(JOBDB)
+    for job in jobs:
+        job.update({jobman.sql.EXPERIMENT: EXPERIMENT_PATH})
+        jobman.sql.insert_dict(job, db)
+
+    print "inserted"
+
+if __name__ == '__main__':
+
+    args = sys.argv[1:]
+
+    #if len(args) > 0 and args[0] == 'load_nist':
+    #    test_load_nist()
+
+    if len(args) > 0 and args[0] == 'jobman_insert':
+        jobman_insert_nist()
+
+    elif len(args) > 0 and args[0] == 'test_jobman_entrypoint':
+        chanmock = DD({'COMPLETE':0,'save':(lambda:None)})
+        jobman_entrypoint(DD(DEFAULT_HP_NIST), chanmock)
+
+    else:
+        print "Bad arguments"
+
diff -r 9a7b74927f7d -r 31641a84e0ae deep/convolutional_dae/salah_exp/sgd_optimization.py
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/deep/convolutional_dae/salah_exp/sgd_optimization.py	Thu Apr 22 00:49:42 2010 -0400
@@ -0,0 +1,371 @@
+#!/usr/bin/python
+# coding: utf-8
+
+# Generic SdA optimization loop, adapted from the deeplearning.net tutorial
+
+import numpy 
+import theano
+import time
+import datetime
+import theano.tensor as T
+import sys
+import pickle
+
+from jobman import DD
+import jobman, jobman.sql
+from copy import copy
+
+from stacked_dae import SdA
+
+from ift6266.utils.seriestables import *
+
+#For test purpose only
+buffersize=1000
+
+default_series = { \
+        'reconstruction_error' : DummySeries(),
+        'training_error' : DummySeries(),
+        'validation_error' : DummySeries(),
+        'test_error' : DummySeries(),
+        'params' : DummySeries()
+        }
+
+def itermax(iter, max):
+    for i,it in enumerate(iter):
+        if i >= max:
+            break
+        yield it
+
+class SdaSgdOptimizer:
+    def __init__(self, dataset, hyperparameters, n_ins, n_outs,
+                    examples_per_epoch, series=default_series, max_minibatches=None):
+        self.dataset = dataset
+        self.hp = hyperparameters
+        self.n_ins = n_ins
+        self.n_outs = n_outs
+        self.parameters_pre=[]
+   
+        self.max_minibatches = max_minibatches
+        print "SdaSgdOptimizer, max_minibatches =", max_minibatches
+
+        self.ex_per_epoch = examples_per_epoch
+        self.mb_per_epoch = examples_per_epoch / self.hp.minibatch_size
+
+        self.series = series
+
+        self.rng = numpy.random.RandomState(1234)
+
+        self.init_classifier()
+
+        sys.stdout.flush()
+
+    def init_classifier(self):
+        print "Constructing classifier"
+
+        # we don't want to save arrays in DD objects, so
+        # we recreate those arrays here
+        nhl = self.hp.num_hidden_layers
+        layers_sizes = [self.hp.hidden_layers_sizes] * nhl
+        corruption_levels = [self.hp.corruption_levels] * nhl
+
+        # construct the stacked denoising autoencoder class
+        self.classifier = SdA( \
+                          batch_size = self.hp.minibatch_size, \
+                          n_ins= self.n_ins, \
+                          hidden_layers_sizes = layers_sizes, \
+                          n_outs = self.n_outs, \
+                          corruption_levels = corruption_levels,\
+                          rng = self.rng,\
+                          pretrain_lr = self.hp.pretraining_lr, \
+                          finetune_lr = self.hp.finetuning_lr)
+
+        #theano.printing.pydotprint(self.classifier.pretrain_functions[0], "function.graph")
+
+        sys.stdout.flush()
+
+    def train(self):
+        self.pretrain(self.dataset)
+        self.finetune(self.dataset)
+
+    def pretrain(self,dataset):
+        print "STARTING PRETRAINING, time = ", datetime.datetime.now()
+        sys.stdout.flush()
+        
+        un_fichier=int(819200.0/self.hp.minibatch_size) #Number of batches in a P07 file
+
+        start_time = time.clock()  
+        ## Pre-train layer-wise 
+        for i in xrange(self.classifier.n_layers):
+            # go through pretraining epochs 
+            for epoch in xrange(self.hp.pretraining_epochs_per_layer):
+                # go through the training set
+                batch_index=0
+                count=0
+                num_files=0
+                for x,y in dataset.train(self.hp.minibatch_size):
+                    c = self.classifier.pretrain_functions[i](x)
+                    count +=1
+
+                    self.series["reconstruction_error"].append((epoch, batch_index), c)
+                    batch_index+=1
+
+                    #if batch_index % 100 == 0:
+                    #    print "100 batches"
+
+                    # useful when doing tests
+                    if self.max_minibatches and batch_index >= self.max_minibatches:
+                        break
+                    
+                    #When we pass through the data only once (the case with P07)
+                    #There is approximately 800*1024=819200 examples per file (1k per example and files are 800M)
+                    if self.hp.pretraining_epochs_per_layer == 1 and count%un_fichier == 0:
+                        print 'Pre-training layer %i, epoch %d, cost '%(i,num_files),c
+                        num_files+=1
+                        sys.stdout.flush()
+                        self.series['params'].append((num_files,), self.classifier.all_params)
+                
+                #When NIST is used
+                if self.hp.pretraining_epochs_per_layer > 1:        
+                    print 'Pre-training layer %i, epoch %d, cost '%(i,epoch),c
+                    sys.stdout.flush()
+
+                    self.series['params'].append((epoch,), self.classifier.all_params)
+     
+        end_time = time.clock()
+
+        print ('Pretraining took %f minutes' %((end_time-start_time)/60.))
+        self.hp.update({'pretraining_time': end_time-start_time})
+        
+        sys.stdout.flush()
+        
+        #To be able to load them later for tests on finetune
+        self.parameters_pre=[copy(x.value) for x in self.classifier.params]
+        f = open('params_pretrain.txt', 'w')
+        pickle.dump(self.parameters_pre,f)
+        f.close()
+
+
+    def finetune(self,dataset,dataset_test,num_finetune,ind_test,special=0,decrease=0):
+        
+        if special != 0 and special != 1:
+            sys.exit('Bad value for variable special. Must be in {0,1}')
+        print "STARTING FINETUNING, time = ", datetime.datetime.now()
+
+        minibatch_size = self.hp.minibatch_size
+        if ind_test == 0 or ind_test == 20:
+            nom_test = "NIST"
+            nom_train="P07"
+        else:
+            nom_test = "P07"
+            nom_train = "NIST"
+
+
+        # create a function to compute the mistakes that are made by the model
+        # on the validation set, or testing set
+        test_model = \
+            theano.function(
+                [self.classifier.x,self.classifier.y], self.classifier.errors)
+        #         givens = {
+        #           self.classifier.x: ensemble_x,
+        #           self.classifier.y: ensemble_y]})
+
+        validate_model = \
+            theano.function(
+                [self.classifier.x,self.classifier.y], self.classifier.errors)
+        #        givens = {
+        #           self.classifier.x: ,
+        #           self.classifier.y: ]})
+
+
+        # 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  = min(self.mb_per_epoch, patience/2)
+                                      # go through this many 
+                                      # minibatche before checking the network 
+                                      # on the validation set; in this case we 
+                                      # check every epoch 
+        if self.max_minibatches and validation_frequency > self.max_minibatches:
+            validation_frequency = self.max_minibatches / 2
+
+        best_params          = None
+        best_validation_loss = float('inf')
+        test_score           = 0.
+        start_time = time.clock()
+
+        done_looping = False
+        epoch = 0
+
+        total_mb_index = 0
+        minibatch_index = 0
+        parameters_finetune=[]
+        
+        if ind_test == 21:
+            learning_rate = self.hp.finetuning_lr / 10.0
+        else:
+            learning_rate = self.hp.finetuning_lr  #The initial finetune lr
+
+
+        while (epoch < num_finetune) and (not done_looping):
+            epoch = epoch + 1
+
+            for x,y in dataset.train(minibatch_size,bufsize=buffersize):
+                minibatch_index += 1
+                
+                
+                if special == 0:
+                    cost_ij = self.classifier.finetune(x,y,learning_rate)
+                elif special == 1:
+                    cost_ij = self.classifier.finetune2(x,y)
+                total_mb_index += 1
+
+                self.series["training_error"].append((epoch, minibatch_index), cost_ij)
+
+                if (total_mb_index+1) % validation_frequency == 0: 
+                    #minibatch_index += 1
+                    #The validation set is always NIST (we want the model to be good on NIST)
+                    if ind_test == 0 | ind_test == 20:
+                        iter=dataset_test.valid(minibatch_size,bufsize=buffersize)
+                    else:
+                        iter = dataset.valid(minibatch_size,bufsize=buffersize)
+                    if self.max_minibatches:
+                        iter = itermax(iter, self.max_minibatches)
+                    validation_losses = [validate_model(x,y) for x,y in iter]
+                    this_validation_loss = numpy.mean(validation_losses)
+
+                    self.series["validation_error"].\
+                        append((epoch, minibatch_index), this_validation_loss*100.)
+
+                    print('epoch %i, minibatch %i, validation error on NIST : %f %%' % \
+                           (epoch, minibatch_index+1, \
+                            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, total_mb_index * patience_increase)
+
+                        # save best validation score, iteration number and parameters
+                        best_validation_loss = this_validation_loss
+                        best_iter = total_mb_index
+                        parameters_finetune=[copy(x.value) for x in self.classifier.params]
+
+                        # test it on the test set
+                        iter = dataset.test(minibatch_size,bufsize=buffersize)
+                        if self.max_minibatches:
+                            iter = itermax(iter, self.max_minibatches)
+                        test_losses = [test_model(x,y) for x,y in iter]
+                        test_score = numpy.mean(test_losses)
+                        
+                        #test it on the second test set
+                        iter2 = dataset_test.test(minibatch_size,bufsize=buffersize)
+                        if self.max_minibatches:
+                            iter2 = itermax(iter2, self.max_minibatches)
+                        test_losses2 = [test_model(x,y) for x,y in iter2]
+                        test_score2 = numpy.mean(test_losses2)
+
+                        self.series["test_error"].\
+                            append((epoch, minibatch_index), test_score*100.)
+
+                        print(('     epoch %i, minibatch %i, test error on dataset %s  (train data) of best '
+                              'model %f %%') % 
+                                     (epoch, minibatch_index+1,nom_train,
+                                      test_score*100.))
+                                    
+                        print(('     epoch %i, minibatch %i, test error on dataset %s of best '
+                              'model %f %%') % 
+                                     (epoch, minibatch_index+1,nom_test,
+                                      test_score2*100.))
+                    
+                    if patience <= total_mb_index:
+                        done_looping = True
+                        break   #to exit the FOR loop
+                    
+                    sys.stdout.flush()
+
+                # useful when doing tests
+                if self.max_minibatches and minibatch_index >= self.max_minibatches:
+                    break
+            
+            if decrease == 1:
+                learning_rate /= 2 #divide the learning rate by 2 for each new epoch
+            
+            self.series['params'].append((epoch,), self.classifier.all_params)
+
+            if done_looping == True:    #To exit completly the fine-tuning
+                break   #to exit the WHILE loop
+
+        end_time = time.clock()
+        self.hp.update({'finetuning_time':end_time-start_time,\
+                    'best_validation_error':best_validation_loss,\
+                    'test_score':test_score,
+                    'num_finetuning_epochs':epoch})
+
+        print(('\nOptimization complete with best validation score of %f %%,'
+               'with test performance %f %% on dataset %s ') %  
+                     (best_validation_loss * 100., test_score*100.,nom_train))
+        print(('The test score on the %s dataset is %f')%(nom_test,test_score2*100.))
+        
+        print ('The finetuning ran for %f minutes' % ((end_time-start_time)/60.))
+        
+        sys.stdout.flush()
+        
+        #Save a copy of the parameters in a file to be able to get them in the future
+        
+        if special == 1:    #To keep a track of the value of the parameters
+            f = open('params_finetune_stanford.txt', 'w')
+            pickle.dump(parameters_finetune,f)
+            f.close()
+        
+        elif ind_test == 0 | ind_test == 20:    #To keep a track of the value of the parameters
+            f = open('params_finetune_P07.txt', 'w')
+            pickle.dump(parameters_finetune,f)
+            f.close()
+               
+
+        elif ind_test== 1:    #For the run with 2 finetunes. It will be faster.
+            f = open('params_finetune_NIST.txt', 'w')
+            pickle.dump(parameters_finetune,f)
+            f.close()
+        
+        elif ind_test== 21:    #To keep a track of the value of the parameters
+            f = open('params_finetune_P07_then_NIST.txt', 'w')
+            pickle.dump(parameters_finetune,f)
+            f.close()
+        
+
+    #Set parameters like they where right after pre-train or finetune
+    def reload_parameters(self,which):
+        
+        #self.parameters_pre=pickle.load('params_pretrain.txt')
+        f = open(which)
+        self.parameters_pre=pickle.load(f)
+        f.close()
+        for idx,x in enumerate(self.parameters_pre):
+            if x.dtype=='float64':
+                self.classifier.params[idx].value=theano._asarray(copy(x),dtype=theano.config.floatX)
+            else:
+                self.classifier.params[idx].value=copy(x)
+
+    def training_error(self,dataset):
+        # create a function to compute the mistakes that are made by the model
+        # on the validation set, or testing set
+        test_model = \
+            theano.function(
+                [self.classifier.x,self.classifier.y], self.classifier.errors)
+                
+        iter2 = dataset.train(self.hp.minibatch_size,bufsize=buffersize)
+        train_losses2 = [test_model(x,y) for x,y in iter2]
+        train_score2 = numpy.mean(train_losses2)
+        print "Training error is: " + str(train_score2)
+
+
+
+
diff -r 9a7b74927f7d -r 31641a84e0ae deep/convolutional_dae/salah_exp/stacked_convolutional_dae_uit.py
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/deep/convolutional_dae/salah_exp/stacked_convolutional_dae_uit.py	Thu Apr 22 00:49:42 2010 -0400
@@ -0,0 +1,242 @@
+import numpy
+import theano
+import time
+import sys
+import theano.tensor as T
+from theano.tensor.shared_randomstreams import RandomStreams
+import theano.sandbox.softsign
+import copy
+from theano.tensor.signal import downsample
+from theano.tensor.nnet import conv 
+
+
+import ift6266.datasets
+from ift6266.baseline.log_reg.log_reg import LogisticRegression
+
+from theano.tensor.xlogx import xlogx, xlogy0
+# it's target*log(output)
+def binary_cross_entropy(target, output, sum_axis=1):
+    XE = xlogy0(target, output) + xlogy0((1 - target), (1 - output))
+    return -T.sum(XE, axis=sum_axis)
+
+
+
+class SigmoidalLayer(object):
+    def __init__(self, rng, input, n_in, n_out):
+
+        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 dA_conv(object):
+ 
+  def __init__(self, input, filter_shape, corruption_level = 0.1, 
+               shared_W = None, shared_b = None, image_shape = None, num = 0,batch_size=20):
+
+    theano_rng = RandomStreams()
+    
+    fan_in = numpy.prod(filter_shape[1:])
+    fan_out = filter_shape[0] * numpy.prod(filter_shape[2:])
+
+    center = theano.shared(value = 1, name="center")
+    scale = theano.shared(value = 2, name="scale")
+
+    if shared_W != None and shared_b != None :
+        self.W = shared_W
+        self.b = shared_b
+    else:
+        initial_W = numpy.asarray( numpy.random.uniform(
+              low = -numpy.sqrt(6./(fan_in+fan_out)),
+              high = numpy.sqrt(6./(fan_in+fan_out)),
+              size = filter_shape), dtype = theano.config.floatX)
+        initial_b = numpy.zeros((filter_shape[0],), dtype=theano.config.floatX)
+        self.W = theano.shared(value = initial_W, name = "W")
+        self.b = theano.shared(value = initial_b, name = "b")
+    
+ 
+    initial_b_prime= numpy.zeros((filter_shape[1],),dtype=theano.config.floatX)
+
+    self.b_prime = theano.shared(value = initial_b_prime, name = "b_prime")
+ 
+    self.x = input
+
+    self.tilde_x = theano_rng.binomial( self.x.shape, 1, 1 - corruption_level,dtype=theano.config.floatX) * self.x
+
+    conv1_out = conv.conv2d(self.tilde_x, self.W, filter_shape=filter_shape,
+                            image_shape=image_shape,
+                            unroll_kern=4,unroll_batch=4, 
+                            border_mode='valid')
+
+    
+    self.y = T.tanh(conv1_out + self.b.dimshuffle('x', 0, 'x', 'x'))
+
+    
+    da_filter_shape = [ filter_shape[1], filter_shape[0], filter_shape[2],\
+                       filter_shape[3] ]
+    da_image_shape = [ batch_size, filter_shape[0], image_shape[2]-filter_shape[2]+1, 
+                       image_shape[3]-filter_shape[3]+1 ]
+    #import pdb; pdb.set_trace()
+    initial_W_prime =  numpy.asarray( numpy.random.uniform( \
+              low = -numpy.sqrt(6./(fan_in+fan_out)), \
+              high = numpy.sqrt(6./(fan_in+fan_out)), \
+              size = da_filter_shape), dtype = theano.config.floatX)
+    self.W_prime = theano.shared(value = initial_W_prime, name = "W_prime")
+
+    conv2_out = conv.conv2d(self.y, self.W_prime,
+                            filter_shape = da_filter_shape,\
+                            image_shape = da_image_shape, \
+                            unroll_kern=4,unroll_batch=4, \
+                            border_mode='full')
+
+    self.z =  (T.tanh(conv2_out + self.b_prime.dimshuffle('x', 0, 'x', 'x'))+center) / scale
+    
+    if num != 0 :
+        scaled_x = (self.x + center) / scale
+    else: 
+        scaled_x = self.x
+    self.L = - T.sum( scaled_x*T.log(self.z) + (1-scaled_x)*T.log(1-self.z), axis=1 )
+
+    self.cost = T.mean(self.L)
+
+    self.params = [ self.W, self.b, self.b_prime ] 
+
+class LeNetConvPoolLayer(object):
+
+    def __init__(self, rng, input, filter_shape, image_shape=None, poolsize=(2,2)):
+        self.input = input
+  
+        W_values = numpy.zeros(filter_shape, dtype=theano.config.floatX)
+        self.W = theano.shared(value=W_values)
+ 
+        b_values = numpy.zeros((filter_shape[0],), dtype=theano.config.floatX)
+        self.b = theano.shared(value=b_values)
+ 
+        conv_out = conv.conv2d(input, self.W,
+                filter_shape=filter_shape, image_shape=image_shape,
+                               unroll_kern=4,unroll_batch=4)
+ 
+
+        fan_in = numpy.prod(filter_shape[1:])
+        fan_out = filter_shape[0] * numpy.prod(filter_shape[2:]) / numpy.prod(poolsize)
+
+        W_bound = numpy.sqrt(6./(fan_in + fan_out))
+        self.W.value = numpy.asarray(
+                rng.uniform(low=-W_bound, high=W_bound, size=filter_shape),
+                dtype = theano.config.floatX)
+  
+
+        pooled_out = downsample.max_pool2D(conv_out, poolsize, ignore_border=True)
+ 
+        self.output = T.tanh(pooled_out + self.b.dimshuffle('x', 0, 'x', 'x'))
+        self.params = [self.W, self.b]
+ 
+
+class CSdA():
+    def __init__(self, n_ins_mlp,batch_size, conv_hidden_layers_sizes,
+                 mlp_hidden_layers_sizes, corruption_levels, rng, n_out, 
+                 pretrain_lr, finetune_lr):
+
+        # Just to make sure those are not modified somewhere else afterwards
+        hidden_layers_sizes = copy.deepcopy(mlp_hidden_layers_sizes)
+        corruption_levels = copy.deepcopy(corruption_levels)
+
+        #update_locals(self, locals())
+
+
+        
+        self.layers = []
+        self.pretrain_functions = []
+        self.params = []
+        self.n_layers = len(conv_hidden_layers_sizes)
+        self.mlp_n_layers = len(mlp_hidden_layers_sizes)
+        
+        self.x = T.matrix('x') # the data is presented as rasterized images
+        self.y = T.ivector('y') # the labels are presented as 1D vector of
+        
+        for i in xrange( self.n_layers ):
+            filter_shape=conv_hidden_layers_sizes[i][0]
+            image_shape=conv_hidden_layers_sizes[i][1]
+            max_poolsize=conv_hidden_layers_sizes[i][2]
+                
+            if i == 0 :
+                layer_input=self.x.reshape((batch_size, 1, 32, 32))
+            else:
+                layer_input=self.layers[-1].output
+            
+            layer = LeNetConvPoolLayer(rng, input=layer_input,
+                                       image_shape=image_shape,
+                                       filter_shape=filter_shape,
+                                       poolsize=max_poolsize)
+            print 'Convolutional layer', str(i+1), 'created'
+            
+            self.layers += [layer]
+            self.params += layer.params
+
+            da_layer = dA_conv(corruption_level = corruption_levels[0],
+                               input = layer_input,
+                               shared_W = layer.W, shared_b = layer.b,
+                               filter_shape=filter_shape,
+                               image_shape = image_shape, num=i , batch_size=batch_size)
+            
+            gparams = T.grad(da_layer.cost, da_layer.params)
+            
+            updates = {}
+            for param, gparam in zip(da_layer.params, gparams):
+                updates[param] = param - gparam * pretrain_lr
+            
+            update_fn = theano.function([self.x], da_layer.cost, updates = updates)
+            
+            self.pretrain_functions += [update_fn]
+            
+        for i in xrange( self.mlp_n_layers ): 
+            if i == 0 :
+                input_size = n_ins_mlp
+            else:
+                input_size = mlp_hidden_layers_sizes[i-1]
+            
+            if i == 0 :
+                if len( self.layers ) == 0 :
+                    layer_input=self.x
+                else :
+                    layer_input = self.layers[-1].output.flatten(2)
+            else:
+                layer_input = self.layers[-1].output
+            
+            layer = SigmoidalLayer(rng, layer_input, input_size,
+                                        mlp_hidden_layers_sizes[i] )
+            
+            self.layers += [layer]
+            self.params += layer.params
+            
+            print 'MLP layer', str(i+1), 'created'
+            
+        self.logLayer = LogisticRegression(input=self.layers[-1].output, \
+                                                     n_in=mlp_hidden_layers_sizes[-1], n_out=n_out)
+
+
+        self.params += self.logLayer.params
+        self.all_params = self.params
+        cost = self.logLayer.negative_log_likelihood(self.y)
+        
+        gparams = T.grad(cost, self.params)
+
+        updates = {}
+        for param,gparam in zip(self.params, gparams):
+            updates[param] = param - gparam*finetune_lr
+        
+        self.finetune = theano.function([self.x, self.y], cost, updates = updates)
+        
+        self.errors = self.logLayer.errors(self.y)
+
+
+
diff -r 9a7b74927f7d -r 31641a84e0ae deep/convolutional_dae/salah_exp/utils.py
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/deep/convolutional_dae/salah_exp/utils.py	Thu Apr 22 00:49:42 2010 -0400
@@ -0,0 +1,69 @@
+#!/usr/bin/python
+# coding: utf-8
+
+from __future__ import with_statement
+
+from jobman import DD
+
+# from pylearn codebase
+# useful in __init__(param1, param2, etc.) to save
+# values in self.param1, self.param2... just call
+# update_locals(self, locals())
+def update_locals(obj, dct):
+    if 'self' in dct:
+        del dct['self']
+    obj.__dict__.update(dct)
+
+# from a dictionary of possible values for hyperparameters, e.g.
+# hp_values = {'learning_rate':[0.1, 0.01], 'num_layers': [1,2]}
+# create a list of other dictionaries representing all the possible
+# combinations, thus in this example creating:
+# [{'learning_rate': 0.1, 'num_layers': 1}, ...]
+# (similarly for combinations (0.1, 2), (0.01, 1), (0.01, 2))
+def produit_cartesien_jobs(val_dict):
+    job_list = [DD()]
+    all_keys = val_dict.keys()
+
+    for key in all_keys:
+        possible_values = val_dict[key]
+        new_job_list = []
+        for val in possible_values:
+            for job in job_list:
+                to_insert = job.copy()
+                to_insert.update({key: val})
+                new_job_list.append(to_insert)
+        job_list = new_job_list
+
+    return job_list
+
+def test_produit_cartesien_jobs():
+    vals = {'a': [1,2], 'b': [3,4,5]}
+    print produit_cartesien_jobs(vals)
+
+
+# taken from http://stackoverflow.com/questions/276052/how-to-get-current-cpu-and-ram-usage-in-python
+"""Simple module for getting amount of memory used by a specified user's
+processes on a UNIX system.
+It uses UNIX ps utility to get the memory usage for a specified username and
+pipe it to awk for summing up per application memory usage and return the total.
+Python's Popen() from subprocess module is used for spawning ps and awk.
+
+"""
+
+import subprocess
+
+class MemoryMonitor(object):
+
+    def __init__(self, username):
+        """Create new MemoryMonitor instance."""
+        self.username = username
+
+    def usage(self):
+        """Return int containing memory used by user's processes."""
+        self.process = subprocess.Popen("ps -u %s -o rss | awk '{sum+=$1} END {print sum}'" % self.username,
+                                        shell=True,
+                                        stdout=subprocess.PIPE,
+                                        )
+        self.stdout_list = self.process.communicate()[0].split('\n')
+        return int(self.stdout_list[0])
+