def initialize_diary():
diary = Diary(name='digits_vs_letters',
path='results',
overwrite=False,
fig_format='svg')
diary.add_notebook('training', verbose=True)
diary.add_notebook('validation', verbose=True)
return diary
def main(dataset_names=None):
if dataset_names is None:
dataset_names = [
'autos',
'car',
'cleveland',
'dermatology',
'ecoli',
'flare',
'glass',
'led7digit',
'lymphography',
'nursery',
'page-blocks',
'pendigits',
'satimage',
'segment',
#'shuttle',
'vehicle',
'vowel',
'yeast',
'zoo',
'auslan'
]
seed_num = 42
mc_iterations = 5
n_folds = 2
estimator_type = "svm"
# Diary to save the partial and final results
diary = Diary(name='results_Krawczyk2015',
path='results',
overwrite=False,
fig_format='svg')
# Hyperparameters for this experiment (folds, iterations, seed)
diary.add_notebook('parameters', verbose=True)
# Summary for each dataset
diary.add_notebook('datasets', verbose=False)
# Partial results for validation
diary.add_notebook('validation', verbose=True)
# Final results
diary.add_notebook('summary', verbose=True)
columns = ['dataset', 'method', 'mc', 'test_fold', 'acc']
df = MyDataFrame(columns=columns)
diary.add_entry('parameters', [
'seed', seed_num, 'mc_it', mc_iterations, 'n_folds', n_folds,
'estimator_type', estimator_type
])
data = Data(dataset_names=dataset_names)
for i, (name, dataset) in enumerate(data.datasets.iteritems()):
np.random.seed(seed_num)
dataset.print_summary()
diary.add_entry('datasets', [dataset.__str__()])
accuracies = np.zeros(mc_iterations * n_folds)
for mc in np.arange(mc_iterations):
skf = StratifiedKFold(dataset.target,
n_folds=n_folds,
shuffle=True)
test_folds = skf.test_folds
for test_fold in np.arange(n_folds):
x_train, y_train, x_test, y_test = separate_sets(
dataset.data, dataset.target, test_fold, test_folds)
if estimator_type == "svm":
est = OneClassSVM(nu=0.5, gamma=0.5)
elif estimator_type == "gmm":
est = GMM(n_components=3)
bc = BackgroundCheck(estimator=est)
oc = OcDecomposition(base_estimator=bc)
oc.fit(x_train, y_train)
accuracy = oc.accuracy(x_test, y_test)
accuracies[mc * n_folds + test_fold] = accuracy
diary.add_entry('validation', [
'dataset', name, 'method', 'our', 'mc', mc, 'test_fold',
test_fold, 'acc', accuracy
])
df = df.append_rows([[name, 'our', mc, test_fold, accuracy]])
df = df.convert_objects(convert_numeric=True)
table = df.pivot_table(values=['acc'],
index=['dataset'],
columns=['method'],
aggfunc=[np.mean, np.std])
diary.add_entry('summary', [table])
def main():
dataset_names = ['diabetes', 'ecoli', 'glass', 'heart-statlog',
'ionosphere', 'iris', 'letter', 'mfeat-karhunen',
'mfeat-morphological', 'mfeat-zernike', 'optdigits',
'pendigits', 'sonar', 'vehicle', 'waveform-5000']
data = Data(dataset_names=dataset_names)
diary = Diary(name='hempstalk', path='results', overwrite=False,
fig_format='svg')
diary.add_notebook('cross_validation')
# Columns for the DataFrame
columns=['Dataset', 'MC iteration', 'N-fold id', 'Actual class', 'Model',
'AUC', 'Prior']
# Create a DataFrame to record all intermediate results
df = pd.DataFrame(columns=columns)
mc_iterations = 10
n_folds = 10
gammas = {"diabetes":0.00005, "ecoli":0.1, "glass":0.005,
"heart-statlog":0.0001, "ionosphere":0.00005, "iris":0.0005,
"letter":0.000005, "mfeat-karhunen":0.0001,
"mfeat-morphological":0.0000001, "mfeat-zernike":0.000001,
"optdigits":0.00005, "pendigits":0.000001, "sonar":0.001,
"vehicle":0.00005, "waveform-5000":0.001}
for i, (name, dataset) in enumerate(data.datasets.iteritems()):
print('Dataset number {}'.format(i))
data.sumarize_datasets(name)
for mc in np.arange(mc_iterations):
skf = StratifiedKFold(dataset.target, n_folds=n_folds, shuffle=True)
test_folds = skf.test_folds
for test_fold in np.arange(n_folds):
x_train, y_train, x_test, y_test = separate_sets(
dataset.data, dataset.target, test_fold, test_folds)
n_training = np.alen(y_train)
w_auc_fold_dens = 0
w_auc_fold_bag = 0
w_auc_fold_com = 0
prior_sum = 0
for actual_class in dataset.classes:
tr_class = x_train[y_train == actual_class, :]
tr_class_unique_values = [np.unique(tr_class[:,column]).shape[0] for column in
range(tr_class.shape[1])]
cols_keep = np.where(np.not_equal(tr_class_unique_values,1))[0]
tr_class = tr_class[:,cols_keep]
x_test_cleaned = x_test[:,cols_keep]
t_labels = (y_test == actual_class).astype(int)
prior = np.alen(tr_class) / n_training
if np.alen(tr_class) > 1 and not all(t_labels == 0):
prior_sum += prior
n_c = tr_class.shape[1]
if n_c > np.alen(tr_class):
n_c = np.alen(tr_class)
# Train a Density estimator
model_gmm = GMM(n_components=1, covariance_type='diag')
model_gmm.fit(tr_class)
sv = OneClassSVM(nu=0.1, gamma=0.5)
bc = BackgroundCheck(estimator=sv)
bc.fit(tr_class)
svm_scores = bc.predict_proba(x_test_cleaned)[:, 1]
# Generate artificial data
new_data = model_gmm.sample(np.alen(tr_class))
# Train a Bag of Trees
bag = BaggingClassifier(
base_estimator=DecisionTreeClassifier(),
n_estimators=10)
new_data = np.vstack((tr_class, new_data))
y = np.zeros(np.alen(new_data))
y[:np.alen(tr_class)] = 1
bag.fit(new_data, y)
# Combine the results
probs = bag.predict_proba(x_test_cleaned)[:, 1]
scores = model_gmm.score(x_test_cleaned)
com_scores = (probs / np.clip(1.0 - probs, np.float32(1e-32), 1.0)) * (scores-scores.min())
# Generate our new data
# FIXME solve problem with #samples < #features
pca=True
if tr_class.shape[0] < tr_class.shape[1]:
pca=False
our_new_data = reject.create_reject_data(
tr_class, proportion=1,
method='uniform_hsphere', pca=pca,
pca_variance=0.99, pca_components=0,
hshape_cov=0, hshape_prop_in=0.99,
hshape_multiplier=1.5)
our_new_data = np.vstack((tr_class, our_new_data))
y = np.zeros(np.alen(our_new_data))
y[:np.alen(tr_class)] = 1
# Train Our Bag of Trees
our_bag = BaggingClassifier(
base_estimator=DecisionTreeClassifier(),
n_estimators=10)
our_bag.fit(our_new_data, y)
# Combine the results
our_probs = our_bag.predict_proba(x_test_cleaned)[:, 1]
our_comb_scores = (our_probs / np.clip(1.0 - our_probs,
np.float32(1e-32), 1.0)) * (scores-scores.min())
# Scores for the Density estimator
auc_dens = roc_auc_score(t_labels, scores)
# Scores for the Bag of trees
auc_bag = roc_auc_score(t_labels, probs)
# Scores for the Combined model
auc_com = roc_auc_score(t_labels, com_scores)
# Scores for our Bag of trees (trained on our data)
auc_our_bag = roc_auc_score(t_labels, our_probs)
# Scores for our Bag of trees (trained on our data)
auc_our_comb = roc_auc_score(t_labels, our_comb_scores)
# Scores for the Background Check with SVm
auc_svm = roc_auc_score(t_labels, svm_scores)
# Create a new DataFrame to append to the original one
dfaux = pd.DataFrame([[name, mc, test_fold, actual_class,
'Combined', auc_com, prior],
[name, mc, test_fold, actual_class,
'P(T$|$X)', auc_bag, prior],
[name, mc, test_fold, actual_class,
'P(X$|$A)', auc_dens, prior],
[name, mc, test_fold, actual_class,
'Our Bagg', auc_our_bag, prior],
[name, mc, test_fold, actual_class,
'Our Combined', auc_our_comb, prior],
[name, mc, test_fold, actual_class,
'SVM_BC', auc_svm, prior]],
columns=columns)
df = df.append(dfaux, ignore_index=True)
# generate_and_save_plots(t_labels, scores, diary, name, mc, test_fold,
# actual_class, 'P(X$|$A)')
# generate_and_save_plots(t_labels, probs, diary, name, mc, test_fold,
# actual_class, 'P(T$|$X)')
# generate_and_save_plots(t_labels, com_scores, diary, name, mc, test_fold,
# actual_class, 'Combined')
# generate_and_save_plots(t_labels, our_probs, diary, name, mc, test_fold,
# actual_class, 'Our_Bagg')
# generate_and_save_plots(t_labels, our_comb_scores, diary, name, mc, test_fold,
# actual_class, 'Our_Combined')
# generate_and_save_plots(t_labels, svm_scores, diary,
# name, mc, test_fold,
# actual_class, 'SVM_BC')
# Convert values to numeric
df = df.convert_objects(convert_numeric=True)
# Group everything except classes
dfgroup_classes = df.groupby(by=['Dataset', 'MC iteration', 'N-fold id',
'Model'])
# Compute the Prior sum for each dataset, iteration and fold
df['Prior_sum'] = dfgroup_classes['Prior'].transform(np.sum)
# Compute the individual weighted AUC per each class and experiment
df['wAUC'] = df.Prior * df.AUC / df.Prior_sum
# Sum the weighted AUC of each class per each experiment
series_wAUC = dfgroup_classes['wAUC'].sum()
# Transform the series to a DataFrame
df_wAUC = series_wAUC.reset_index(inplace=False)
# Compute mean and standard deviation of wAUC per Dataset and model
final_results = df_wAUC.groupby(['Dataset', 'Model'])['wAUC'].agg([np.mean,
np.std])
# Transform the series to a DataFrame
final_results.reset_index(inplace=True)
# Represent the results in a table format
final_table = final_results.pivot_table(values=['mean', 'std'],
index=['Dataset'], columns=['Model'])
# Export the results in a csv and LaTeX file
export_results(final_table)
return model_rej
def train_classifier_model(x, y):
model_clas = svm.SVC(probability=True)
#model_clas = tree.DecisionTreeClassifier(max_depth=3)
model_clas = model_clas.fit(x, y)
return model_clas
if __name__ == "__main__":
diary = Diary(name='test_rgrpg',
path='results',
overwrite=False,
fig_format='svg')
diary.add_notebook('training')
diary.add_notebook('validation')
# for i in [6]: #range(1,4):
n_iterations = 1
n_thresholds = 100
accuracies = np.empty((n_iterations, n_thresholds))
recalls = np.empty((n_iterations, n_thresholds))
for example in [2, 3, 4, 5, 6, 7, 8, 9]:
np.random.seed(42)
print('Runing example = {}'.format(example))
for iteration in range(n_iterations):
#####################################################
# TRAINING #
#####################################################
def main(dataset_names=None,
estimator_type="gmm",
mc_iterations=20,
n_folds=5,
n_ensemble=100,
seed_num=42):
if dataset_names is None:
# All the datasets used in Li2014
datasets_li2014 = [
'abalone', 'balance-scale', 'credit-approval', 'dermatology',
'ecoli', 'german', 'heart-statlog', 'hepatitis', 'horse',
'ionosphere', 'lung-cancer', 'libras-movement', 'mushroom',
'diabetes', 'landsat-satellite', 'segment', 'spambase', 'wdbc',
'wpbc', 'yeast'
]
datasets_hempstalk2008 = [
'diabetes', 'ecoli', 'glass', 'heart-statlog', 'ionosphere',
'iris', 'letter', 'mfeat-karhunen', 'mfeat-morphological',
'mfeat-zernike', 'optdigits', 'pendigits', 'sonar', 'vehicle',
'waveform-5000'
]
datasets_others = [
'diabetes', 'ecoli', 'glass', 'heart-statlog', 'ionosphere',
'iris', 'letter', 'mfeat-karhunen', 'mfeat-morphological',
'mfeat-zernike', 'optdigits', 'pendigits', 'sonar', 'vehicle',
'waveform-5000', 'scene-classification', 'tic-tac', 'autos', 'car',
'cleveland', 'dermatology', 'flare', 'page-blocks', 'segment',
'shuttle', 'vowel', 'zoo', 'abalone', 'balance-scale',
'credit-approval', 'german', 'hepatitis', 'lung-cancer'
]
# Datasets that we can add but need to be reduced
datasets_to_add = ['MNIST']
dataset_names = list(
set(datasets_li2014 + datasets_hempstalk2008 + datasets_others))
# Diary to save the partial and final results
diary = Diary(name='results_Li2014',
path='results',
overwrite=False,
fig_format='svg')
# Hyperparameters for this experiment (folds, iterations, seed)
diary.add_notebook('parameters', verbose=True)
# Summary for each dataset
diary.add_notebook('datasets', verbose=False)
# Partial results for validation
diary.add_notebook('validation', verbose=True)
# Final results
diary.add_notebook('summary', verbose=True)
columns = ['dataset', 'method', 'mc', 'test_fold', 'acc', 'logloss']
df = MyDataFrame(columns=columns)
diary.add_entry('parameters', [
'seed', seed_num, 'mc_it', mc_iterations, 'n_folds', n_folds,
'n_ensemble', n_ensemble, 'estimator_type', estimator_type
])
data = Data(dataset_names=dataset_names)
for name, dataset in data.datasets.iteritems():
if name in ['letter', 'shuttle']:
dataset.reduce_number_instances(0.1)
export_datasets_description_to_latex(data, path=diary.path)
for i, (name, dataset) in enumerate(data.datasets.iteritems()):
np.random.seed(seed_num)
dataset.print_summary()
diary.add_entry('datasets', [dataset.__str__()])
for mc in np.arange(mc_iterations):
skf = StratifiedKFold(dataset.target,
n_folds=n_folds,
shuffle=True)
test_folds = skf.test_folds
for test_fold in np.arange(n_folds):
x_train, y_train, x_test, y_test = separate_sets(
dataset.data, dataset.target, test_fold, test_folds)
# Binary discriminative classifier
sv = SVC(kernel='linear', probability=True)
# Density estimator for the background check
if estimator_type == "svm":
gamma = 1.0 / x_train.shape[1]
est = OneClassSVM(nu=0.1, gamma=gamma)
elif estimator_type == "gmm":
est = GMM(n_components=1)
elif estimator_type == "gmm3":
est = GMM(n_components=3)
elif estimator_type == "mymvn":
est = MyMultivariateNormal()
# Multiclass discriminative model with one-vs-one binary class.
ovo = OvoClassifier(base_classifier=sv)
classifier = ConfidentClassifier(classifier=ovo,
estimator=est,
mu=0.5,
m=0.5)
ensemble = Ensemble(base_classifier=classifier,
n_ensemble=n_ensemble)
# classifier = ConfidentClassifier(classifier=sv,
# estimator=est, mu=0.5,
# m=0.5)
# ovo = OvoClassifier(base_classifier=classifier)
# ensemble = Ensemble(base_classifier=ovo,
# n_ensemble=n_ensemble)
xs_bootstrap, ys_bootstrap = ensemble.fit(x_train, y_train)
accuracy = ensemble.accuracy(x_test, y_test)
log_loss = ensemble.log_loss(x_test, y_test)
diary.add_entry('validation', [
'dataset', name, 'method', 'our', 'mc', mc, 'test_fold',
test_fold, 'acc', accuracy, 'logloss', log_loss
])
df = df.append_rows(
[[name, 'our', mc, test_fold, accuracy, log_loss]])
# Li2014: EP-CC model
# The classification confidence is used in learning the weights
# of the base classifier as well as in weighted voting.
ensemble_li = Ensemble(n_ensemble=n_ensemble, lambd=1e-8)
ensemble_li.fit(x_train,
y_train,
xs=xs_bootstrap,
ys=ys_bootstrap)
accuracy_li = ensemble_li.accuracy(x_test, y_test)
log_loss_li = ensemble_li.log_loss(x_test, y_test)
diary.add_entry('validation', [
'dataset', name, 'method', 'Li2014', 'mc', mc, 'test_fold',
test_fold, 'acc', accuracy_li, 'logloss', log_loss_li
])
df = df.append_rows(
[[name, 'Li2014', mc, test_fold, accuracy_li,
log_loss_li]])
export_summary(df, diary)
batch_size=5000
inner_batch_size=5000
nb_classes=2
noise_proportion=0.25
score_lin=np.linspace(0,1,100)
minibatch_method='lineal' # 'random', 'lineal'
n_hidden=[25, 25]
output_activation= 'sigmoid' # 'isotonic_regression' # sigmoid
if nb_classes == 2:
loss='binary_crossentropy'
else:
loss='categorical_crossentropy'
diary = Diary(name='experiment', path='results')
diary.add_notebook('hyperparameters')
diary.add_entry('hyperparameters', ['train_size', train_size])
diary.add_entry('hyperparameters', ['num_classes', nb_classes])
diary.add_entry('hyperparameters', ['batch_size', batch_size])
diary.add_entry('hyperparameters', ['inner_batch_size', inner_batch_size])
diary.add_entry('hyperparameters', ['minibatch_method', minibatch_method])
diary.add_entry('hyperparameters', ['output_activation', output_activation])
diary.add_entry('hyperparameters', ['loss', loss])
diary.add_entry('hyperparameters', ['optimizer', optimizer.get_config()['name']])
for key, value in optimizer.get_config().iteritems():
diary.add_entry('hyperparameters', [key, value])
diary.add_entry('hyperparameters', ['binarize', binarize])
diary.add_entry('hyperparameters', ['add_noise', add_noise])
diary.add_entry('hyperparameters', ['noise', noise_proportion])
diary.add_notebook('training')
diary.add_notebook('validation')
def main(dataset_names=None,
estimator_type="kernel",
mc_iterations=1,
n_folds=10,
seed_num=42):
if dataset_names is None:
dataset_names = ['glass', 'hepatitis', 'ionosphere', 'vowel']
bandwidths_o_norm = {
'glass': 0.09,
'hepatitis': 0.105,
'ionosphere': 0.039,
'vowel': 0.075
}
bandwidths_bc = {
'glass': 0.09,
'hepatitis': 0.105,
'ionosphere': 0.039,
'vowel': 0.0145
}
bandwidths_t_norm = {
'glass': 0.336,
'hepatitis': 0.015,
'ionosphere': 0.0385,
'vowel': 0.0145
}
tuned_mus = {
'glass': [0.094, 0.095, 0.2, 0.0, 0.0, 0.1],
'vowel': [0.0, 0.0, 0.5, 0.5, 0.5, 0.0]
}
tuned_ms = {
'glass': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
'vowel': [1.0, 1.0, 1.0, 1.0, 1.0, 1.0]
}
bandwidth_o_norm = 0.05
bandwidth_t_norm = 0.05
bandwidth_bc = 0.05
# Diary to save the partial and final results
diary = Diary(name='results_Tax2008',
path='results',
overwrite=False,
fig_format='svg')
# Hyperparameters for this experiment (folds, iterations, seed)
diary.add_notebook('parameters', verbose=True)
# Summary for each dataset
diary.add_notebook('datasets', verbose=False)
# Partial results for validation
diary.add_notebook('validation', verbose=True)
# Final results
diary.add_notebook('summary', verbose=True)
columns = ['dataset', 'method', 'mc', 'test_fold', 'acc']
df = MyDataFrame(columns=columns)
diary.add_entry('parameters', [
'seed', seed_num, 'mc_it', mc_iterations, 'n_folds', n_folds,
'estimator_type', estimator_type, 'bw_o', bandwidth_o_norm, 'bw_t',
bandwidth_t_norm, 'bw_bc', bandwidth_bc
])
data = Data(dataset_names=dataset_names)
for name, dataset in data.datasets.iteritems():
if name in ['letter', 'shuttle']:
dataset.reduce_number_instances(0.1)
export_datasets_description_to_latex(data, path=diary.path)
for i, (name, dataset) in enumerate(data.datasets.iteritems()):
np.random.seed(seed_num)
dataset.print_summary()
diary.add_entry('datasets', [dataset.__str__()])
# accuracies_tuned = np.zeros(mc_iterations * n_folds)
# if name in bandwidths_o_norm.keys():
# bandwidth_o_norm = bandwidths_o_norm[name]
# bandwidth_t_norm = bandwidths_t_norm[name]
# bandwidth_bc = bandwidths_bc[name]
# else:
# bandwidth_o_norm = np.mean(bandwidths_o_norm.values())
# bandwidth_t_norm = np.mean(bandwidths_t_norm.values())
# bandwidth_bc = np.mean(bandwidths_bc.values())
for mc in np.arange(mc_iterations):
skf = StratifiedKFold(dataset.target,
n_folds=n_folds,
shuffle=True)
test_folds = skf.test_folds
for test_fold in np.arange(n_folds):
x_train, y_train, x_test, y_test = separate_sets(
dataset.data, dataset.target, test_fold, test_folds)
# if name in ['glass', 'hepatitis', 'ionosphere', 'thyroid',
# 'iris', 'heart-statlog', 'diabetes', 'abalone',
# 'mushroom', 'spambase']:
x_test, y_test = generate_outliers(x_test, y_test)
# elif name == 'vowel':
# x_train = x_train[y_train <= 5]
# y_train = y_train[y_train <= 5]
# y_test[y_test > 5] = 6
# elif dataset.n_classes > 2:
# x_train = x_train[y_train <= dataset.n_classes/2]
# y_train = y_train[y_train <= dataset.n_classes/2]
# y_test[y_test > dataset.n_classes/2] = dataset.n_classes+1
# else:
# continue
if estimator_type == "svm":
est = OneClassSVM(nu=0.5, gamma=1.0 / x_train.shape[1])
elif estimator_type == "gmm":
est = GMM(n_components=1)
elif estimator_type == "gmm3":
est = GMM(n_components=3)
elif estimator_type == "kernel":
est = MyMultivariateKernelDensity(kernel='gaussian',
bandwidth=bandwidth_bc)
estimators = None
bcs = None
if estimator_type == "kernel":
estimators, bcs = fit_estimators(
MyMultivariateKernelDensity(kernel='gaussian',
bandwidth=bandwidth_bc),
x_train, y_train)
# Untuned background check
bc = BackgroundCheck(estimator=est, mu=0.0, m=1.0)
oc = OcDecomposition(base_estimator=bc)
if estimators is None:
oc.fit(x_train, y_train)
else:
oc.set_estimators(bcs, x_train, y_train)
accuracy = oc.accuracy(x_test, y_test)
diary.add_entry('validation', [
'dataset', name, 'method', 'BC', 'mc', mc, 'test_fold',
test_fold, 'acc', accuracy
])
df = df.append_rows([[name, 'BC', mc, test_fold, accuracy]])
e = MyMultivariateKernelDensity(kernel='gaussian',
bandwidth=bandwidth_o_norm)
oc_o_norm = OcDecomposition(base_estimator=e,
normalization="O-norm")
if estimators is None:
oc_o_norm.fit(x_train, y_train)
else:
oc_o_norm.set_estimators(estimators, x_train, y_train)
accuracy_o_norm = oc_o_norm.accuracy(x_test, y_test)
diary.add_entry('validation', [
'dataset', name, 'method', 'O-norm', 'mc', mc, 'test_fold',
test_fold, 'acc', accuracy_o_norm
])
df = df.append_rows(
[[name, 'O-norm', mc, test_fold, accuracy_o_norm]])
e = MyMultivariateKernelDensity(kernel='gaussian',
bandwidth=bandwidth_t_norm)
oc_t_norm = OcDecomposition(base_estimator=e,
normalization="T-norm")
if estimators is None:
oc_t_norm.fit(x_train, y_train)
else:
oc_t_norm.set_estimators(estimators, x_train, y_train)
accuracy_t_norm = oc_t_norm.accuracy(x_test, y_test)
diary.add_entry('validation', [
'dataset', name, 'method', 'T-norm', 'mc', mc, 'test_fold',
test_fold, 'acc', accuracy_t_norm
])
df = df.append_rows(
[[name, 'T-norm', mc, test_fold, accuracy_t_norm]])
# Tuned background check
# if name in tuned_mus.keys():
# mus = tuned_mus[name]
# ms = tuned_ms[name]
# else:
# mus = None
# ms = None
# bc = BackgroundCheck(estimator=est, mu=0.0, m=1.0)
# oc_tuned = OcDecomposition(base_estimator=bc)
# oc_tuned.fit(x_train, y_train, mus=mus, ms=ms)
# accuracy_tuned = oc_tuned.accuracy(x_test, y_test, mus=mus,
# ms=ms)
# accuracies_tuned[mc * n_folds + test_fold] = accuracy_tuned
# diary.add_entry('validation', ['dataset', name,
# 'method', 'BC-tuned',
# 'mc', mc,
# 'test_fold', test_fold,
# 'acc', accuracy_tuned])
# df = df.append_rows([[name, 'BC-tuned', mc, test_fold,
# accuracy_tuned]])
export_summary(df, diary)
def sgd_optimization_gauss(learning_rate=0.13, n_epochs=1000,
batch_size=600):
"""
Demonstrate stochastic gradient descent optimization of a log-linear
model
:type learning_rate: float
:param learning_rate: learning rate used (factor for the stochastic
gradient)
:type n_epochs: int
:param n_epochs: maximal number of epochs to run the optimizer
"""
#datasets = load_data(dataset)
diary = Diary(name='experiment', path='results')
diary.add_notebook('training')
diary.add_notebook('validation')
diary.add_notebook('data')
samples=[4000,10000]
diary.add_entry('data', ['samples', samples])
diary.add_entry('data', ['num_classes', len(samples)])
diary.add_entry('data', ['batch_size', batch_size])
#means=[[0,0],[5,5]]
#cov=[[[1,0],[0,1]],[[3,0],[0,3]]]
#diary.add_entry('data', ['means', means])
#diary.add_entry('data', ['covariance', cov])
#datasets = generate_gaussian_data(means, cov, samples)
datasets = generate_opposite_cs_data(samples)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
diary.add_entry('data', ['train_size', len(train_set_y.eval())])
diary.add_entry('data', ['valid_size', len(valid_set_y.eval())])
diary.add_entry('data', ['test_size', len(test_set_y.eval())])
pt = PresentationTier()
pt.plot_samples(train_set_x.eval(), train_set_y.eval())
# compute number of minibatches for training, validation and testing
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
delta = 20
x_min = numpy.min(train_set_x.eval(),axis=0)
x_max = numpy.max(train_set_x.eval(),axis=0)
x1_lin = numpy.linspace(x_min[0], x_max[0], delta)
x2_lin = numpy.linspace(x_min[1], x_max[1], delta)
MX1, MX2 = numpy.meshgrid(x1_lin, x2_lin)
x_grid = numpy.asarray([MX1.flatten(),MX2.flatten()]).T
grid_set_x = theano.shared(numpy.asarray(x_grid,
dtype=theano.config.floatX),
borrow=True)
n_grid_batches = grid_set_x.get_value(borrow=True).shape[0] / batch_size
######################
# BUILD ACTUAL MODEL #
######################
print '... building the model'
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # the data is presented as rasterized images
y = T.ivector('y') # the labels are presented as 1D vector of
# [int] labels
# construct the logistic regression class
# Each MNIST image has size 28*28
n_in = train_set_x.eval().shape[-1]
n_out = max(train_set_y.eval()) + 1
classifier = LogisticRegression(input=x, n_in=n_in, n_out=n_out)
# the cost we minimize during training is the negative log likelihood of
# the model in symbolic format
cost = classifier.negative_log_likelihood(y)
# compiling a Theano function that computes the mistakes that are made by
# the model on a minibatch
test_model = theano.function(inputs=[index],
outputs=classifier.errors(y),
givens={
x: test_set_x[index * batch_size: (index + 1) * batch_size],
y: test_set_y[index * batch_size: (index + 1) * batch_size]})
validate_model = theano.function(inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]})
# Scores
grid_scores_model = theano.function(inputs=[],
outputs=classifier.scores(),
givens={
x: grid_set_x})
training_scores_model = theano.function(
inputs=[index],
outputs=classifier.scores(),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
}
)
validation_scores_model = theano.function(
inputs=[index],
outputs=classifier.scores(),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
}
)
# compute the gradient of cost with respect to theta = (W,b)
g_w = T.grad(cost=cost, wrt=classifier.w)
g_b = T.grad(cost=cost, wrt=classifier.b)
# specify how to update the parameters of the model as a list of
# (variable, update expression) pairs.
updates = [(classifier.w, classifier.w - learning_rate * g_w),
(classifier.b, classifier.b - learning_rate * g_b)]
# 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(inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]})
# Accuracy
validation_accuracy_model = theano.function(
inputs=[index],
outputs=classifier.accuracy(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]
}
)
training_accuracy_model = theano.function(
inputs=[index],
outputs=classifier.accuracy(y),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]
}
)
# Loss
training_error_model = theano.function(inputs=[index],
outputs=classifier.errors(y),
givens={
x: train_set_x[index * batch_size:(index + 1) * batch_size],
y: train_set_y[index * batch_size:(index + 1) * batch_size]})
validation_error_model = theano.function(inputs=[index],
outputs=classifier.errors(y),
givens={
x: valid_set_x[index * batch_size:(index + 1) * batch_size],
y: valid_set_y[index * batch_size:(index + 1) * batch_size]})
###############
# TRAIN MODEL #
###############
print '... training the model'
# early-stopping parameters
patience = 5000 # 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(n_train_batches, patience / 2)
# go through this many
# minibatche before checking the network
# on the validation set; in this case we
# check every epoch
print('Creating error and accuracy vectors')
error_train = numpy.zeros(n_epochs+1)
error_val = numpy.zeros(n_epochs+1)
accuracy_train = numpy.zeros(n_epochs+1)
accuracy_val = numpy.zeros(n_epochs+1)
# Results for Isotonic Regression
error_train_ir = numpy.zeros(n_epochs+1)
error_val_ir = numpy.zeros(n_epochs+1)
accuracy_train_ir = numpy.zeros(n_epochs+1)
accuracy_val_ir = numpy.zeros(n_epochs+1)
best_params = None
best_validation_loss = numpy.inf
test_score = 0.
start_time = time.clock()
ir = IsotonicRegression(increasing=True, out_of_bounds='clip',
y_min=0, y_max=1)
done_looping = False
epoch = 0
CS = None
while (epoch < n_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
minibatch_avg_cost = train_model(minibatch_index)
# iteration number
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
# compute zero-one loss on validation set
validation_losses = [validate_model(i)
for i in xrange(n_valid_batches)]
this_validation_loss = numpy.mean(validation_losses)
print('epoch %i, minibatch %i/%i, validation error %f %%' % \
(epoch, minibatch_index + 1, n_train_batches,
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)
best_validation_loss = this_validation_loss
# test it on the test set
test_losses = [test_model(i)
for i in xrange(n_test_batches)]
test_score = numpy.mean(test_losses)
print((' epoch %i, minibatch %i/%i, test error of best'
' model %f %%') %
(epoch, minibatch_index + 1, n_train_batches,
test_score * 100.))
if patience <= iter:
done_looping = True
break
scores_grid = grid_scores_model()
fig = pt.update_contourline(grid_set_x.eval(), scores_grid, delta)
diary.save_figure(fig, filename='contour_lines', extension='svg')
scores_train = numpy.asarray([training_scores_model(i) for i
in range(n_train_batches)]).flatten()
scores_val = numpy.asarray([validation_scores_model(i) for i
in range(n_valid_batches)]).flatten()
print('Learning Isotonic Regression from TRAINING set')
ir.fit(scores_train, train_set_y.eval())
scores_train_ir = ir.predict(scores_train)
print('IR predict validation probabilities')
scores_val_ir = ir.predict(scores_val)
scores_set = (scores_train, scores_val, scores_train_ir,
scores_val_ir)
labels_set = (train_set_y.eval(), valid_set_y.eval(),
train_set_y.eval(), valid_set_y.eval())
legend = ['train', 'valid', 'iso. train', 'iso. valid']
fig = pt.plot_reliability_diagram(scores_set, labels_set, legend)
diary.save_figure(fig, filename='reliability_diagram', extension='svg')
# TODO add reliability map
scores_set = (scores_train)
prob_set = (train_set_y.eval())
fig = pt.plot_reliability_map(scores_set, labels_set, legend)
diary.save_figure(fig, filename='reliability_map', extension='svg')
fig = pt.plot_histogram_scores(scores_set)
diary.save_figure(fig, filename='histogram_scores', extension='svg')
# Performance
accuracy_train[epoch] = numpy.asarray([training_accuracy_model(i) for i
in range(n_train_batches)]).flatten().mean()
accuracy_val[epoch] = numpy.asarray([validation_accuracy_model(i) for i
in range(n_valid_batches)]).flatten().mean()
error_train[epoch] = numpy.asarray([training_error_model(i) for i
in range(n_train_batches)]).flatten().mean()
error_val[epoch] = numpy.asarray([validation_error_model(i) for i
in range(n_valid_batches)]).flatten().mean()
accuracy_train_ir[epoch] = compute_accuracy(scores_train_ir, train_set_y.eval())
accuracy_val_ir[epoch] = compute_accuracy(scores_val_ir, valid_set_y.eval())
error_train_ir[epoch] = compute_cross_entropy(scores_train_ir, train_set_y.eval())
error_val_ir[epoch] = compute_cross_entropy(scores_val_ir, valid_set_y.eval())
diary.add_entry('training', [error_train[epoch], accuracy_train[epoch]])
diary.add_entry('validation', [error_val[epoch], accuracy_val[epoch]])
accuracy_set = (accuracy_train[1:epoch], accuracy_val[1:epoch],
accuracy_train_ir[1:epoch], accuracy_val_ir[1:epoch])
fig = pt.plot_accuracy(accuracy_set, legend)
diary.save_figure(fig, filename='accuracy', extension='svg')
error_set = (error_train[1:epoch], error_val[1:epoch],
error_train_ir[1:epoch], error_val_ir[1:epoch])
fig = pt.plot_error(error_set, legend, 'cross-entropy')
diary.save_figure(fig, filename='error', extension='svg')
pt.update_contourline(grid_set_x.eval(), scores_grid, delta,
clabel=True)
end_time = time.clock()
print(('Optimization complete with best validation score of %f %%,'
'with test performance %f %%') %
(best_validation_loss * 100., test_score * 100.))
print 'The code run for %d epochs, with %f epochs/sec' % (
epoch, 1. * epoch / (end_time - start_time))
