def SdAWrapper(X, batch_size = 128, layers = [512, 64], corruption_levels = [0.3, 0.3], pretrain_epochs = 100, pretrain_lr = 0.001):
X = theano.shared(np.asarray(X, dtype = theano.config.floatX), borrow = True)
n_samples, n_vars = X.get_value(borrow=True).shape
n_train_batches = n_samples / batch_size
numpy_rng = numpy.random.RandomState(23432)
###############
# BUILD MODEL #
###############
logging.info('Building model')
logging.info(str(n_vars) + ' -> ' + ' -> '.join(map(str, layers)))
sda = SdA(numpy_rng=numpy_rng, n_ins=n_vars,
hidden_layers_sizes=layers)
logging.info('Compiling training functions')
pretraining_fns = sda.pretraining_functions(train_set_x=X,
batch_size=batch_size)
#####################
# TRAINING MODEL #
#####################
logging.info('Training model')
for i in xrange(sda.n_layers):
for epoch in xrange(pretrain_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_lr))
logging.info('Training layer {}, epoch {}, cost {}'.format(
i, epoch, numpy.mean(c)))
return sda
def main():
print('Loading data from file...')
f = h5py.File('feature_matrix.h5', 'r')
data = f['data'][:]
col = f['col'][:]
row = f['row'][:]
shape = f['shape'][:]
matrix = csc_matrix(
(numpy.array(data), (numpy.array(row), numpy.array(col))),
shape=(shape[0], shape[1]),
dtype=numpy.uint8)
print(matrix.shape)
# exit(1)
batch_size = 10
n_samples, n_vars = matrix.shape
n_train_batches = n_samples / batch_size
numpy_rng = numpy.random.RandomState(23432)
# build model
print('Building model...')
sda = SdA(numpy_rng=numpy_rng,
n_ins=n_vars,
hidden_layers_sizes=[
int(sys.argv[2]),
int(sys.argv[3]),
int(sys.argv[4])
])
print('configuring...')
pretraining_fns = sda.pretraining_functions(train_set_x=matrix.todense(),
batch_size=batch_size)
print('training...')
pretraining_epochs = 15
pretrain_lr = 0.001
corruption_levels = [0.1, 0.2, 0.3] + [0.4] * sda.n_layers
for i in xrange(sda.n_layers):
for epoch in xrange(pretraining_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_lr))
y = sda.get_lowest_hidden_values(matrix)
get_y = theano.function([], y)
y_val = get_y()
print_array(y_val, index=len(sys.argv) - 1)
def main(args):
logging.info('Loading data')
X, index = load_data(args.input)
# Compute number of minibatches
batch_size = 10
n_samples, n_vars = X.get_value(borrow=True).shape
n_train_batches = n_samples / batch_size
numpy_rng = numpy.random.RandomState(23432)
###############
# BUILD MODEL #
###############
logging.info('Building model')
logging.info(str(n_vars) + ' -> ' + ' -> '.join(map(str, args.l)))
sda = SdA(numpy_rng=numpy_rng, n_ins=n_vars,
hidden_layers_sizes=args.l)
#####################
# TRAINING MODEL #
#####################
logging.info('Compiling training functions')
pretraining_fns = sda.pretraining_functions(train_set_x=X,
batch_size=batch_size)
logging.info('Training model')
pretraining_epochs = 15
pretrain_lr = 0.001
corruption_levels = [0.1, 0.2, 0.3] + [0.4] * sda.n_layers
for i in xrange(sda.n_layers):
for epoch in xrange(pretraining_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_lr))
logging.info('Training layer {}, epoch {}, cost {}'.format(
i, epoch, numpy.mean(c)))
y = sda.get_lowest_hidden_values(X)
get_y = theano.function([], y)
y_val = get_y()
print_array(y_val, index=index)
def main(args):
logging.info('Loading data')
X, index = load_data(args.input)
# Compute number of minibatches
batch_size = 10
n_samples, n_vars = X.get_value(borrow=True).shape
n_train_batches = n_samples / batch_size
numpy_rng = numpy.random.RandomState(23432)
###############
# BUILD MODEL #
###############
logging.info('Building model')
logging.info(str(n_vars) + ' -> ' + ' -> '.join(map(str, args.l)))
sda = SdA(numpy_rng=numpy_rng, n_ins=n_vars, hidden_layers_sizes=args.l)
#####################
# TRAINING MODEL #
#####################
logging.info('Compiling training functions')
pretraining_fns = sda.pretraining_functions(train_set_x=X,
batch_size=batch_size)
logging.info('Training model')
pretraining_epochs = 15
pretrain_lr = 0.001
corruption_levels = [0.1, 0.2, 0.3] + [0.4] * sda.n_layers
for i in xrange(sda.n_layers):
for epoch in xrange(pretraining_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_lr))
logging.info('Training layer {}, epoch {}, cost {}'.format(
i, epoch, numpy.mean(c)))
y = sda.get_lowest_hidden_values(X)
get_y = theano.function([], y)
y_val = get_y()
print_array(y_val, index=index)
def SdAWrapper(X,
batch_size=128,
layers=[512, 64],
corruption_levels=[0.3, 0.3],
pretrain_epochs=100,
pretrain_lr=0.001):
X = theano.shared(np.asarray(X, dtype=theano.config.floatX), borrow=True)
n_samples, n_vars = X.get_value(borrow=True).shape
n_train_batches = n_samples / batch_size
numpy_rng = numpy.random.RandomState(23432)
###############
# BUILD MODEL #
###############
logging.info('Building model')
logging.info(str(n_vars) + ' -> ' + ' -> '.join(map(str, layers)))
sda = SdA(numpy_rng=numpy_rng, n_ins=n_vars, hidden_layers_sizes=layers)
logging.info('Compiling training functions')
pretraining_fns = sda.pretraining_functions(train_set_x=X,
batch_size=batch_size)
#####################
# TRAINING MODEL #
#####################
logging.info('Training model')
for i in xrange(sda.n_layers):
for epoch in xrange(pretrain_epochs):
c = []
for batch_index in xrange(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_lr))
logging.info('Training layer {}, epoch {}, cost {}'.format(
i, epoch, numpy.mean(c)))
return sda
