def test_classification_train_valid_test(self):
dataset = larochelle_etal_2007.Rectangles() # smallest one with splits
assert not hasattr(dataset, 'classification_train_valid_test_task')
train, valid, test = tasks.classification_train_valid_test(dataset)
tasks.assert_classification(*train)
tasks.assert_classification(*valid)
tasks.assert_classification(*test)
assert len(train[0]) == dataset.descr['n_train']
assert len(valid[0]) == dataset.descr['n_valid']
assert len(test[0]) == dataset.descr['n_test']
tasks.assert_classification_train_valid_test(train, valid, test)
def test_classification_train_valid_test(self):
dataset = larochelle_etal_2007.Rectangles() # smallest one with splits
assert not hasattr(dataset, 'classification_train_valid_test_task')
train, valid, test = tasks.classification_train_valid_test(dataset)
tasks.assert_classification(*train)
tasks.assert_classification(*valid)
tasks.assert_classification(*test)
assert len(train[0]) == dataset.descr['n_train']
assert len(valid[0]) == dataset.descr['n_valid']
assert len(test[0]) == dataset.descr['n_test']
tasks.assert_classification_train_valid_test(train, valid, test)
def preprocess_data(config, ctrl):
dataset = json_call(config['dataset_name'])
train, valid, test = classification_train_valid_test(dataset)
X_train, y_train = numpy.asarray(train[0]), numpy.asarray(train[1])
X_valid, y_valid = numpy.asarray(valid[0]), numpy.asarray(valid[1])
X_test, y_test = numpy.asarray(test[0]), numpy.asarray(test[1])
if config['preprocessing']['kind'] == 'pca':
# compute pca of input (TODO: retrieve only pca_whitened input)
raise NotImplementedError('rewrite since cut and paste')
(eigvals,eigvecs), centered_trainset = pylearn_pca.pca_from_examples(
X=dataset['inputs'][:dataset['n_train']],
max_energy_fraction=config['pca_energy'])
eigmean = dataset['inputs'][0] - centered_trainset[0]
whitened_inputs = pylearn_pca.pca_whiten((eigvals,eigvecs),
dataset['inputs']-eigmean)
ctrl.info('PCA kept %i of %i components'%(whitened_inputs.shape[1],
dataset['n_inputs']))
elif config['preprocessing']['kind'] == 'zca':
(eigvals,eigvecs), centered_trainset = pylearn_pca.pca_from_examples(
X=X_train,
max_energy_fraction=config['preprocessing']['energy'])
eigmean = X_train[0] - centered_trainset[0]
def whiten(X):
X = pylearn_pca.pca_whiten((eigvals,eigvecs),
X - eigmean)
X = pylearn_pca.pca_whiten_inverse((eigvals, eigvecs),
X) + eigmean
X = X.astype('float32')
X_min = X.min()
X_max = X.max()
ctrl.info('ZCA min:%f max:%f' % (X_min, X_max))
if X_min < 0 or X_max > 1.0:
ctrl.info('ZCA clamping return value to (0, 1) interval')
X = numpy.clip(X, 0, 1, out=X)
return X
X_train, X_valid, X_test = [whiten(X)
for X in [X_train, X_valid, X_test]]
elif config['preprocessing']['kind'] == 'normalize':
raise NotImplementedError('rewrite since cut and paste')
n_train=dataset['n_train']
whitened_inputs = dataset['inputs']
whitened_inputs = whitened_inputs - whitened_inputs[:n_train].mean(axis=0)
whitened_inputs /= whitened_inputs[:n_train].std(axis=0)+1e-7
elif config['preprocessing']['kind'] == 'raw':
pass
else:
raise ValueError(
'unrecognized preprocessing',
config['preprocessing']['kind'])
for Xy in 'X', 'y':
for suffix in 'train', 'valid', 'test':
varname = '%s_%s'%(Xy, suffix)
var = locals()[varname]
ctrl.info('%s shape=%s max=%f min=%f' % (
varname,
var.shape,
var.max(),
var.min()))
s_X_train = theano.shared(X_train)
s_y_train = theano.shared(y_train)
s_X_valid = theano.shared(X_valid)
s_y_valid = theano.shared(y_valid)
s_X_test = theano.shared(X_test)
s_y_test = theano.shared(y_test)
return (dataset,
(s_X_train, s_y_train),
(s_X_valid, s_y_valid),
(s_X_test, s_y_test))
def preprocess_data(config, ctrl):
dataset = json_call(config['dataset_name'])
train, valid, test = classification_train_valid_test(dataset)
X_train, y_train = numpy.asarray(train[0]), numpy.asarray(train[1])
X_valid, y_valid = numpy.asarray(valid[0]), numpy.asarray(valid[1])
X_test, y_test = numpy.asarray(test[0]), numpy.asarray(test[1])
if config['preprocessing']['kind'] == 'pca':
# compute pca of input (TODO: retrieve only pca_whitened input)
raise NotImplementedError('rewrite since cut and paste')
(eigvals, eigvecs), centered_trainset = pylearn_pca.pca_from_examples(
X=dataset['inputs'][:dataset['n_train']],
max_energy_fraction=config['pca_energy'])
eigmean = dataset['inputs'][0] - centered_trainset[0]
whitened_inputs = pylearn_pca.pca_whiten((eigvals, eigvecs),
dataset['inputs'] - eigmean)
ctrl.info('PCA kept %i of %i components' %
(whitened_inputs.shape[1], dataset['n_inputs']))
elif config['preprocessing']['kind'] == 'zca':
(eigvals, eigvecs), centered_trainset = pylearn_pca.pca_from_examples(
X=X_train, max_energy_fraction=config['preprocessing']['energy'])
eigmean = X_train[0] - centered_trainset[0]
def whiten(X):
X = pylearn_pca.pca_whiten((eigvals, eigvecs), X - eigmean)
X = pylearn_pca.pca_whiten_inverse((eigvals, eigvecs), X) + eigmean
X = X.astype('float32')
X_min = X.min()
X_max = X.max()
ctrl.info('ZCA min:%f max:%f' % (X_min, X_max))
if X_min < 0 or X_max > 1.0:
ctrl.info('ZCA clamping return value to (0, 1) interval')
X = numpy.clip(X, 0, 1, out=X)
return X
X_train, X_valid, X_test = [
whiten(X) for X in [X_train, X_valid, X_test]
]
elif config['preprocessing']['kind'] == 'normalize':
raise NotImplementedError('rewrite since cut and paste')
n_train = dataset['n_train']
whitened_inputs = dataset['inputs']
whitened_inputs = whitened_inputs - whitened_inputs[:n_train].mean(
axis=0)
whitened_inputs /= whitened_inputs[:n_train].std(axis=0) + 1e-7
elif config['preprocessing']['kind'] == 'raw':
pass
else:
raise ValueError('unrecognized preprocessing',
config['preprocessing']['kind'])
for Xy in 'X', 'y':
for suffix in 'train', 'valid', 'test':
varname = '%s_%s' % (Xy, suffix)
var = locals()[varname]
ctrl.info('%s shape=%s max=%f min=%f' %
(varname, var.shape, var.max(), var.min()))
s_X_train = theano.shared(X_train)
s_y_train = theano.shared(y_train)
s_X_valid = theano.shared(X_valid)
s_y_valid = theano.shared(y_valid)
s_X_test = theano.shared(X_test)
s_y_test = theano.shared(y_test)
return (dataset, (s_X_train, s_y_train), (s_X_valid, s_y_valid),
(s_X_test, s_y_test))
