def test_rbm(learning_rate=0.1, training_epochs=15,
dataset='mnist.pkl.gz', batch_size=20,
n_chains=20, n_samples=10, output_folder='rbm_plots',
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
"""
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
test_set_x, test_set_y = datasets[2]
# compute number of minibatches for training, validation and testing
n_train_batches = 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
x = T.matrix('x') # the data is presented as rasterized images
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
# 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)
# construct the RBM class
rbm = RBM(input=x, n_visible=28 * 28,
n_hidden=n_hidden, numpy_rng=rng, theano_rng=theano_rng)
# get the cost and the gradient corresponding to one step of CD-15
cost, updates = rbm.get_cost_updates(lr=learning_rate,
persistent=persistent_chain, k=15)
#################################
# 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()
# 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)]
print 'Training epoch %d, cost is ' % epoch, numpy.mean(mean_cost)
# 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
print ('Training took %f minutes' % (pretraining_time / 60.))
# end-snippet-5 start-snippet-6
#################################
# Sampling from the RBM #
#################################
# 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(
test_set_x.get_value(borrow=True)[test_idx:test_idx + n_chains],
dtype=theano.config.floatX
)
)
# end-snippet-6 start-snippet-7
plot_every = 1000
# 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=plot_every
)
# 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 - 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 = Image.fromarray(image_data)
image.save('samples.png')
# end-snippet-7
os.chdir('../')
def main(learning_rate=0.1, training_epochs=15,
dataset='mnist.pkl.gz',
batch_size=20, output_folder='dA_plots'):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
index = T.lscalar()
x = T.matrix(name='x')
rng = np.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
dae = DenoisingAutoencoder(
numpy_rng=rng,
theano_rng=theano_rng,
input=x,
n_visible=28*28,
n_hidden=500
)
cost, updates = dae.get_cost_updates(
corruption_level=0.3,
learning_rate=learning_rate
)
train_autoencoder = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
}
)
start_time = timeit.default_timer()
for epoch in xrange(training_epochs):
costs = []
for batch_index in xrange(n_train_batches):
costs.append(train_autoencoder(batch_index))
print 'Training epoch %d, mean cost ' % epoch, np.mean(costs)
end_time = timeit.default_timer()
training_time = end_time - start_time
print (
'The 0% corruption code for file ',
os.path.split(__file__)[1],
' ran for %.2fm' %(training_time / 60.)
)
image = Image.fromarray(
tile_raster_images(
X=dae.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
image.save('filters_corruption_0.png')
dae = DenoisingAutoencoder(
numpy_rng=rng,
theano_rng=theano_rng,
input=x,
n_visible=28*28,
n_hidden=500,
tied=False
)
cost, updates = dae.get_cost_updates(
corruption_level=0.3,
learning_rate=learning_rate
)
train_autoencoder = theano.function(
inputs=[index],
outputs=cost,
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
}
)
start_time = timeit.default_timer()
for epoch in xrange(training_epochs):
costs = []
for batch_index in xrange(n_train_batches):
costs.append(train_autoencoder(batch_index))
print 'Training epoch %d, mean cost ' % epoch, np.mean(costs)
end_time = timeit.default_timer()
training_time = end_time - start_time
print (
'The 30% corruption code for file ',
os.path.split(__file__)[1],
' ran for %.2fm' %(training_time / 60.)
)
image = Image.fromarray(
tile_raster_images(
X=dae.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
image.save('filters_corruption_30.png')
os.chdir('../')
def init_mlp(learning_rate, L1_reg, L2_reg, dataset, batch_size, n_hidden):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
validate_set_x, validate_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
n_validate_batches = validate_set_x.get_value(borrow=True).shape[0] / batch_size
n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size
batches = (n_train_batches, n_validate_batches, n_test_batches)
print '... building the model'
index = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
rng = np.random.RandomState(1234)
classifier = MultiLayerPerceptron(
rng=rng,
input=x,
n_in=28*28,
n_hidden=n_hidden,
n_out=10)
cost = (
classifier.negative_log_likelihood(y) +
L1_reg * classifier.L1_norm +
L2_reg * classifier.L2_norm
)
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: validate_set_x[index * batch_size: (index + 1) * batch_size],
y: validate_set_y[index * batch_size: (index + 1) * batch_size]
}
)
gparams = [T.grad(cost, param) for param in classifier.params]
updates = [
(param, param - learning_rate * gparam)
for param, gparam in zip(classifier.params, gparams)
]
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],
}
)
return train_model, validate_model, test_model, batches
def main(
finetune_learning_rate=0.1,
pretrain_epochs=15,
pretrain_learning_rate=0.001,
training_epochs=1000,
dataset='mnist.pkl.gz',
batch_size=1,
output_folder='SdA_plots/'
):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0]
n_train_batches /= batch_size
numpy_rng = np.random.RandomState(123)
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
print '... building the model'
sda = StackedDenoisingAutoencoder(
numpy_rng=numpy_rng,
n_ins=28*28,
hidden_layers_sizes=[529, 529],
n_outs=10
)
print '... getting the pretraining the model'
pretrain_fns = sda.create_pretraining_functions(
train_set_x=train_set_x,
batch_size=batch_size)
print '... pre-training the model'
start_time = timeit.default_timer()
corruption_levels = [.1, .2, .3]
for i in xrange(sda.n_layers):
for epoch in xrange(pretrain_epochs):
costs = []
for batch_index in xrange(n_train_batches):
costs.append(pretrain_fns[i](
index=batch_index,
corruption=corruption_levels[i],
learning_rate=pretrain_learning_rate)
)
print 'Pre-training layer %i, epoch %d, cost ' % (i, epoch), np.mean(costs)
end_time = timeit.default_timer()
print >> sys.stderr, ('The pretraining code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
for i, layer in enumerate(sda.dA_layers):
if i == 0:
image = Image.fromarray(
tile_raster_images(
X=layer.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
else:
image = Image.fromarray(
tile_raster_images(
X=layer.W.get_value(borrow=True).T,
img_shape=(23, 23),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
image.save('dAE_layer_pretrain_%d' % i + '.png')
print '... getting the finetunning the model'
train_fn, validatate_model, test_model = sda.build_finetune_functions(
datasets=datasets,
batch_size=batch_size,
learning_rate=finetune_learning_rate
)
print '... fine-tunning the model'
patience = 10 * n_train_batches
patience_increase = 2.
improvement_threshold = 0.995
validation_frequency = min(n_train_batches, patience / 2)
best_validation_loss = np.inf
test_score = 0.
start_time = timeit.default_timer()
done_looping = False
epoch = 0
while (epoch < training_epochs) and (not done_looping):
epoch = epoch + 1
for minibatch_index in xrange(n_train_batches):
minibatch_avg_cost = train_fn(minibatch_index)
iter = (epoch - 1) * n_train_batches + minibatch_index
if (iter + 1) % validation_frequency == 0:
validation_losses = validatate_model()
this_validation_loss = np.mean(validation_losses)
print ('epoch %i, minibatch %i/%i, validation error %f %%' %
(epoch, minibatch_index + 1, n_train_batches, this_validation_loss * 100.))
if this_validation_loss < best_validation_loss:
if this_validation_loss < best_validation_loss * improvement_threshold:
patience = max(patience, iter * patience_increase)
best_validation_loss = this_validation_loss
best_iter = iter
test_losses = test_model()
test_score = np.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
end_time = timeit.default_timer()
print (
(
'Optimization complete with best validation score of %f %%, '
'on iteration %i, '
'with test performance %f %%'
) %
(best_validation_loss * 100., best_iter + 1, test_score * 100.)
)
print >> sys.stderr, ('The training code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
for i, layer in enumerate(sda.dA_layers):
if i == 0:
image = Image.fromarray(
tile_raster_images(
X=layer.W.get_value(borrow=True).T,
img_shape=(28, 28),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
else:
image = Image.fromarray(
tile_raster_images(
X=layer.W.get_value(borrow=True).T,
img_shape=(23, 23),
tile_shape=(10, 10),
tile_spacing=(1, 1)
)
)
image.save('dAE_layer_finetune_%d' % i + '.png')
os.chdir('../')
def main(learning_rate=0.1,
training_epochs=15,
dataset='mnist.pkl.gz',
batch_size=20,
n_chains=20,
n_samples=10,
output_folder='rbm_plots',
n_hidden=500):
datasets = load_data(dataset)
train_set_x, train_set_y = datasets[0]
test_set_x, test_set_y = datasets[2]
n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size
index = T.lscalar()
x = T.matrix('x')
rng = np.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
persistent_chain = theano.shared(np.zeros((batch_size, n_hidden),
dtype=theano.config.floatX),
borrow=True)
rbm = RestrictedBolzmanMachine(
input=x,
n_visible=28 * 28,
n_hidden=n_hidden,
numpy_rng=rng,
theano_rng=theano_rng
)
cost, updates = rbm.get_cost_updates(learning_rate=learning_rate,
persistent=persistent_chain,
k=15)
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
train_rbm = theano.function(
inputs=[index],
outputs=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()
for epoch in xrange(training_epochs):
mean_cost = []
for batch_index in xrange(n_train_batches):
mean_cost += [train_rbm(batch_index)]
print 'Training epoch %d, cost is ' % epoch, np.mean(mean_cost)
plotting_start = timeit.default_timer()
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
print ('Training took %f minutes' % (pretraining_time / 60.))
number_of_test_samples = test_set_x.get_value(borrow=True).shape[0]
test_idx = rng.randint(number_of_test_samples - n_chains)
persistent_vis_chain = theano.shared(
np.asarray(
test_set_x.get_value(borrow=True)[test_idx: test_idx + n_chains],
dtype=theano.config.floatX
)
)
plot_every = 1000
(
[
presig_hids,
hid_mfs,
hid_samples,
presig_vis,
vis_mfs,
vis_samples
],
updates
) = theano.scan(
fn=rbm.gibbs_vhv,
outputs_info=[None, None, None, None, None, persistent_vis_chain],
n_steps=plot_every
)
updates.update({persistent_vis_chain: vis_samples[-1]})
sample_fn = theano.function(
inputs=[],
outputs=[
vis_mfs[-1],
vis_samples[-1]
],
updates=updates,
name='sample_fn'
)
image_data = np.zeros(
(29 * n_samples + 1, 29 * n_chains - 1),
dtype='uint8'
)
for idx in xrange(n_samples):
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)
)
image = Image.fromarray(image_data)
image.save('sample.png')
os.chdir('../')
def main():
datasets = load_data('mnist.pkl.gz')
batch_size = 20
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
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
print '... building model'
index = T.lscalar()
x = T.matrix('x')
t = T.ivector('t')
layers = [
GeneralLayer(28 * 28, 500),
GeneralLayer(500, 500),
GeneralLayer(500, 500),
GeneralLayer(500, 10, T.nnet.softmax)
]
params = []
for i, layer in enumerate(layers):
params += layer.params
if i == 0:
layer_out = layer.forward_prop(x)
else:
layer_out = layer.forward_prop(layer_out)
y = layers[-1].h
L1_norm = sum([layer.L1_norm for layer in layers])
L2_norm = sum([layer.L2_norm for layer in layers])
# Negative log likelihood
# cost = -T.mean(T.log(y)[T.arange(x.shape[0]), t]) + 0.000 * L1_norm + 0.0001 * L2_norm
cost = -T.mean((y[T.arange(x.shape[0])] - t )** 2)
error = T.mean(T.neq(T.argmax(y, axis=1), t))
# gparams = [T.grad(cost, param) for param in params]
gparams = T.grad(cost, params)
gmomentums = [
theano.shared(np.asarray(
np.zeros_like(param.get_value(borrow=True)),
dtype=theano.config.floatX)
)
for param in params
]
updates = OrderedDict()
# learning_rate = 0.1
learning_rate = 0.01
momentum = np.float(0.9)
for param, gparam, gmomentum in zip(params, gparams, gmomentums):
updates[gmomentum] = momentum * gmomentum - learning_rate * gparam
updates[param] = param + updates[gmomentum]
# updates[param] = param - learning_rate * gparam
train = theano.function(
inputs=[index],
outputs=[cost],
updates=updates,
givens={
x: train_set_x[index * batch_size: (index + 1) * batch_size],
t: train_set_y[index * batch_size: (index + 1) * batch_size]
}
)
valid = theano.function(
inputs=[index],
outputs=[error],
givens={
x: valid_set_x[index * batch_size: (index + 1) * batch_size],
t: valid_set_y[index * batch_size: (index + 1) * batch_size],
}
)
for epoch in xrange(500):
for minibatch_index in xrange(n_train_batches):
train(minibatch_index)
iter = epoch * n_train_batches + minibatch_index
if (iter + 1) % n_train_batches == 0:
validation_losses = [valid(i) for i in xrange(n_valid_batches)]
this_validation_loss = np.mean(validation_losses)
print ("EPOCH:: %i, Validation cost: %f" % (epoch+1, this_validation_loss))