def train_rbm(rbm, dataset, learning_rate=0.1, training_epochs=10,
batch_size=20, output_folder='rbm_plots', n_hidden=500, CD_steps=3):
train_set_x, train_set_y = dataset
# compute number of minibatches for training, validation and testing
n_train_batches = int(train_set_x.get_value(borrow=True).shape[0] / batch_size)
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
# initialize storage for the persistent chain (state = hidden
# layer of chain)
persistent_chain = theano.shared(numpy.zeros((batch_size, n_hidden),
dtype=theano.config.floatX),
borrow=True)
# get the cost and the gradient corresponding to one step of CD-1
cost, updates = rbm.get_cost_updates(lr=learning_rate,
persistent=persistent_chain, k=CD_steps)
#################################
# Training the RBM #
#################################
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
# start-snippet-5
# it is ok for a theano function to have no output
# the purpose of train_rbm is solely to update the RBM parameters
train_rbm = theano.function(
[index],
cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size]
},
name='train_rbm'
)
plotting_time = 0.
start_time = timeit.default_timer()
logFile = open('rbm_logs.txt', 'w')
logFile.write('LR: ' + str(learning_rate) + ', Epoch: ' + str(training_epochs) + ', PCD-' + str(CD_steps) + '\n\n')
# 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):
mean_cost += [train_rbm(batch_index)]
logFile.write('Training epoch %d, cost is ' + str(numpy.mean(mean_cost)) + '\n' % epoch)
# Plot filters after each training epoch
plotting_start = timeit.default_timer()
# Construct image from the weight matrix
image = Image.fromarray(
tile_raster_images(
X=rbm.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
image.save('filters_at_epoch_%i.png' % epoch)
plotting_stop = timeit.default_timer()
plotting_time += (plotting_stop - plotting_start)
end_time = timeit.default_timer()
pretraining_time = (end_time - start_time) - plotting_time
logFile.write('Training took %f minutes\n' % (pretraining_time / 60.))
logFile.close()
rbm.save_data('rbm_data')
def train_rbm(rbm,
dataset,
learning_rate=0.1,
training_epochs=10,
batch_size=20,
output_folder='rbm_plots',
n_hidden=500,
CD_steps=3):
train_set_x, train_set_y = dataset
# compute number of minibatches for training, validation and testing
n_train_batches = int(
train_set_x.get_value(borrow=True).shape[0] / batch_size)
# allocate symbolic variables for the data
index = T.lscalar() # index to a [mini]batch
# initialize storage for the persistent chain (state = hidden
# layer of chain)
persistent_chain = theano.shared(numpy.zeros((batch_size, n_hidden),
dtype=theano.config.floatX),
borrow=True)
# get the cost and the gradient corresponding to one step of CD-1
cost, updates = rbm.get_cost_updates(lr=learning_rate,
persistent=persistent_chain,
k=CD_steps)
#################################
# Training the RBM #
#################################
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
# start-snippet-5
# it is ok for a theano function to have no output
# the purpose of train_rbm is solely to update the RBM parameters
train_rbm = theano.function(
[index],
cost,
updates=updates,
givens={x: train_set_x[index * batch_size:(index + 1) * batch_size]},
name='train_rbm')
plotting_time = 0.
start_time = timeit.default_timer()
logFile = open('rbm_logs.txt', 'w')
logFile.write('LR: ' + str(learning_rate) + ', Epoch: ' +
str(training_epochs) + ', PCD-' + str(CD_steps) + '\n\n')
# 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):
mean_cost += [train_rbm(batch_index)]
logFile.write('Training epoch %d, cost is ' +
str(numpy.mean(mean_cost)) + '\n' % epoch)
# Plot filters after each training epoch
plotting_start = timeit.default_timer()
# Construct image from the weight matrix
image = Image.fromarray(
tile_raster_images(X=rbm.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)))
image.save('filters_at_epoch_%i.png' % epoch)
plotting_stop = timeit.default_timer()
plotting_time += (plotting_stop - plotting_start)
end_time = timeit.default_timer()
pretraining_time = (end_time - start_time) - plotting_time
logFile.write('Training took %f minutes\n' % (pretraining_time / 60.))
logFile.close()
rbm.save_data('rbm_data')
def sample_rbm(rbm, test_set_x, n_chains=1, n_samples=100, n_step=1000, percentage_noise = 5,n_repeat = 1):
rng = numpy.random.RandomState()
if os.path.isdir('rbm_plots'):
os.chdir('rbm_plots')
if rbm == None:
x = T.matrix('x')
theano_rng = RandomStreams(rng.randint(2 ** 30))
n_visable, n_hidden, W, hbias, vbias = get_data('tmp')
rbm = RBM(input=x, n_visible=n_visable, n_hidden=n_hidden, W=W,hbias=hbias, vbias=vbias ,numpy_rng=rng, theano_rng=theano_rng)
# find out the number of test samples
number_of_test_samples = test_set_x.get_value(borrow=True).shape[0]
# pick random test examples, with which to initialize the persistent chain
test_idx = rng.randint(number_of_test_samples - n_chains)
persistent_vis_chain = theano.shared(
numpy.asarray(
minipulate_and_save_image(test_set_x.get_value(borrow=True),test_idx, n_chains, percentage_noise, n_repeat),
dtype=theano.config.floatX
)
)
# define one step of Gibbs sampling (mf = mean-field) define a
# function that does `plot_every` steps before returning the
# sample for plotting
(
[
presig_hids,
hid_mfs,
hid_samples,
presig_vis,
vis_mfs,
vis_samples
],
updates
) = theano.scan(
rbm.gibbs_vhv,
outputs_info=[None, None, None, None, None, persistent_vis_chain],
n_steps=n_step
)
# add to updates the shared variable that takes care of our persistent
# chain :.
updates.update({persistent_vis_chain: vis_samples[-1]})
# construct the function that implements our persistent chain.
# we generate the "mean field" activations for plotting and the actual
# samples for reinitializing the state of our persistent chain
sample_fn = theano.function(
[],
[
vis_mfs[-1],
vis_samples[-1]
],
updates=updates,
name='sample_fn'
)
# 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 * n_repeat - 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()
if idx * n_step < 25:
print("normalizing")
data = persistent_vis_chain.get_value()
normal = test_normalize(data)
data[len(data) -1] = test_set_x.get_value(borrow=True)[rng.randint(number_of_test_samples - n_chains)]
persistent_vis_chain = theano.shared(
numpy.asarray(
data,
dtype=theano.config.floatX
)
)
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*n_repeat),
tile_spacing=(1, 1)
)
# construct image
image = Image.fromarray(image_data)
image.save('samples.png')
def sample_rbm(rbm,
test_set_x,
n_chains=1,
n_samples=100,
n_step=1000,
percentage_noise=5,
n_repeat=1):
rng = numpy.random.RandomState()
if os.path.isdir('rbm_plots'):
os.chdir('rbm_plots')
if rbm == None:
x = T.matrix('x')
theano_rng = RandomStreams(rng.randint(2**30))
n_visable, n_hidden, W, hbias, vbias = get_data('tmp')
rbm = RBM(input=x,
n_visible=n_visable,
n_hidden=n_hidden,
W=W,
hbias=hbias,
vbias=vbias,
numpy_rng=rng,
theano_rng=theano_rng)
# find out the number of test samples
number_of_test_samples = test_set_x.get_value(borrow=True).shape[0]
# pick random test examples, with which to initialize the persistent chain
test_idx = rng.randint(number_of_test_samples - n_chains)
persistent_vis_chain = theano.shared(
numpy.asarray(minipulate_and_save_image(
test_set_x.get_value(borrow=True), test_idx, n_chains,
percentage_noise, n_repeat),
dtype=theano.config.floatX))
# define one step of Gibbs sampling (mf = mean-field) define a
# function that does `plot_every` steps before returning the
# sample for plotting
([presig_hids, hid_mfs, hid_samples, presig_vis, vis_mfs,
vis_samples], updates) = theano.scan(
rbm.gibbs_vhv,
outputs_info=[None, None, None, None, None, persistent_vis_chain],
n_steps=n_step)
# add to updates the shared variable that takes care of our persistent
# chain :.
updates.update({persistent_vis_chain: vis_samples[-1]})
# construct the function that implements our persistent chain.
# we generate the "mean field" activations for plotting and the actual
# samples for reinitializing the state of our persistent chain
sample_fn = theano.function([], [vis_mfs[-1], vis_samples[-1]],
updates=updates,
name='sample_fn')
# 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 * n_repeat - 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()
if idx * n_step < 25:
print("normalizing")
data = persistent_vis_chain.get_value()
normal = test_normalize(data)
data[len(data) - 1] = test_set_x.get_value(
borrow=True)[rng.randint(number_of_test_samples - n_chains)]
persistent_vis_chain = theano.shared(
numpy.asarray(data, dtype=theano.config.floatX))
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 * n_repeat),
tile_spacing=(1, 1))
# construct image
image = Image.fromarray(image_data)
image.save('samples.png')