def test_rbm_mnist(learning_rate=0.01, training_epochs=10, batch_size=20,
n_chains=30, n_samples=5, output_folder=None, isPCD=0,
n_hidden=500):
"""
Demonstrate how to train and afterwards sample from it using Theano.
This is demonstrated on MNIST.
:param learning_rate: learning rate used for training the RBM
:param training_epochs: number of epochs used for training
:param dataset: path the the pickled dataset
:param batch_size: size of a batch used to train the RBM
:param n_chains: number of parallel Gibbs chains to be used for sampling
:param n_samples: number of samples to plot for each chain
e.g.
test_rbm_mnist(output_folder='/home/eric/Desktop/rbm_plots')
"""
assert output_folder is not None
from rbm_variants import RBM_Orthogonal as RBM
# from rbm import RBM
#################################
# Data Constructing #
#################################
from sklearn.datasets import fetch_mldata
mnist = fetch_mldata('MNIST original')
from xylearn.utils.data_util import get_train_test
from xylearn.utils.data_normalization import rescale
data = get_train_test(rescale(mnist.data), mnist.target, useGPU=1, shuffle=True)
train_x, train_y = data['train']
n_vis = train_x.get_value(borrow=True).shape[1]
print numpy.linalg.matrix_rank(train_x.get_value(borrow=True))
n_train_batches = train_x.get_value(borrow=True).shape[0] / batch_size
# construct the RBM class
rbm = RBM(n_visible=n_vis, n_hidden=n_hidden, isPCD=isPCD)
train_fn = rbm.get_train_fn(train_x, batch_size)
#################################
# Training the RBM #
#################################
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
plotting_time = 0.
start_time = time.clock()
import PIL.Image
from visualizer import tile_raster_images
# go through training epochs
for epoch in xrange(training_epochs):
# go through the training set
mean_cost = []
for batch_index in xrange(n_train_batches):
# for each batch, we extract the gibbs chain
new_cost = train_fn(index=batch_index, lr=learning_rate)
mean_cost += [new_cost]
print 'Training epoch %d, cost is ' % epoch, numpy.mean(mean_cost)
# monitor projected rank
projection = rbm.project(train_x)
print 'rank: ' + str(numpy.linalg.matrix_rank(projection))
# W shape is [784 500]
# Plot filters after each training epoch
plotting_start = time.clock()
# Construct image from the weight matrix
image = PIL.Image.fromarray(tile_raster_images(
X=rbm.W.get_value(borrow=True).T,
img_shape=(28, 28), tile_shape=(20, 20),
tile_spacing=(1, 1)))
image.save('filters_at_epoch_%i.png' % epoch)
plotting_stop = time.clock()
plotting_time += (plotting_stop - plotting_start)
end_time = time.clock()
pretraining_time = (end_time - start_time) - plotting_time
print ('Training took %f minutes' % (pretraining_time / 60.))
#################################
# Sampling from the RBM #
#################################
test_idx = 1
test_x, test_y = data['test']
sample_fn = rbm.get_sampling_fn(test_x, test_idx, n_chains)
print '... begin sampling'
# plot initial image first
orig_img = test_x.get_value(borrow=True)[test_idx:test_idx + 1]
image = PIL.Image.fromarray(tile_raster_images(
X=orig_img,
img_shape=(28, 28), tile_shape=(1, 1),
tile_spacing=(1, 1)))
image.save('orig_img.png')
# create a space to store the image for plotting ( we need to leave
# room for the tile_spacing as well)
image_data = numpy.zeros((29 * n_samples + 1, 29 * n_chains - 1),
dtype='uint8')
for idx in xrange(n_samples):
# generate `plot_every` intermediate samples that we discard,
# because successive samples in the chain are too correlated
vis_mf, vis_sample = sample_fn()
print ' ... plotting sample ', idx
image_data[29 * idx:29 * idx + 28, :] = tile_raster_images(
X=vis_mf,
img_shape=(28, 28),
tile_shape=(1, n_chains),
tile_spacing=(1, 1))
# construct image
image = PIL.Image.fromarray(image_data)
image.save('samples.png')
os.chdir('../')
#################################
# Projecting from the RBM #
#################################
projection = rbm.project(train_x)
print numpy.linalg.matrix_rank(projection)
def toy_test(learning_rate=0.01, training_epochs=100, batch_size=50,
output_folder=None, isPCD=0,
n_hidden=3):
assert output_folder is not None
# toy_data, word count vector, [num_terms, num_doc].
# each cell represents the number of times a term occurs
# d1 d2 d3 d4 d5
toy_data = numpy.asarray([[0, 2, 0, 1, 0],
[9, 0, 3, 1, 1],
[4, 1, 1, 2, 1],
[10, 10, 1, 1, 0],
[1, 0, 8, 0, 10],
[0, 1, 10, 1, 0],
[1, 0, 2, 6, 1],
[0, 0, 1, 0, 0],
[1, 0, 0, 0, 0],
[1, 0, 1, 0, 0],
[1, 1, 0, 0, 1],
[10, 2, 0, 1, 0],
[0, 0, 1, 0, 10],
[1, 0, 0, 3, 0],
[0, 0, 2, 0, 1],
[10, 0, 1, 0, 0],
[0, 1, 0, 0, 0],
[0, 1, 0, 1, 0],
[1, 0, 1, 0, 0],
[1, 0, 0, 0, 1],
[1, 0, 1, 0, 0],
[0, 0, 1, 0, 0]])
# from rbm import RBM
from rbm_variants import RBM_Orthogonal as RBM
# from rbm_variants import PoissonRBM as RBM
train_x = toSharedX(toy_data, name="toy_data")
n_vis = train_x.get_value(borrow=True).shape[1]
n_samples = train_x.get_value(borrow=True).shape[0]
if batch_size >= n_samples:
batch_size = n_samples
n_train_batches = n_samples / batch_size
# construct the RBM class
rbm = RBM(n_visible=n_vis, n_hidden=n_hidden, isPCD=isPCD)
train_fn = rbm.get_train_fn(train_x, batch_size)
print "... projecting"
print rbm.project(train_x, hidSample=1)
#################################
# Training the RBM #
#################################
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
plotting_time = 0.
start_time = time.clock()
import PIL.Image
from visualizer import tile_raster_images
# go through training epochs
for epoch in xrange(training_epochs):
# go through the training set
mean_cost = []
for batch_index in xrange(n_train_batches):
# for each batch, we extract the gibbs chain
new_cost = train_fn(index=batch_index, lr=learning_rate)
mean_cost += [new_cost]
print 'Training epoch %d, cost is ' % epoch, numpy.mean(mean_cost)
if numpy.mean(mean_cost) >= 0:
break
# W shape is [784 500]
# Plot filters after each training epoch
plotting_start = time.clock()
# Construct image from the weight matrix
image = PIL.Image.fromarray(tile_raster_images(
X=rbm.W.get_value(borrow=True).T,
# weight is [n_vis, n_hidden]
# so, among 'n_hidden' rows,
# each row corresponds to propdown one hidden unit
img_shape=(1, n_vis), tile_shape=(n_hidden, 1),
tile_spacing=(1, 1)))
image.save('filters_at_epoch_%i.png' % epoch)
plotting_stop = time.clock()
plotting_time += (plotting_stop - plotting_start)
end_time = time.clock()
pretraining_time = (end_time - start_time) - plotting_time
print ('Training took %f minutes' % (pretraining_time / 60.))
print "... projecting"
print rbm.project(train_x, hidSample=1)
print "... reconstructing"
print rbm.reconstruct(train_x, showSample=1) * train_x.get_value(borrow=True)