def test_set_up(self):
tr = rbm_config.TrainParam(learning_rate=0.05,
momentum_type=rbm_config.CLASSICAL,
momentum=0.5,
weight_decay=0,
sparsity_constraint=False,
sparsity_target=0.1**9,
sparsity_cost=10**8,
sparsity_decay=0.9,
epochs=5)
config = rbm_config.RBMConfig()
config.v_n = 3
config.h_n = 2
config.v_unit = rbm_units.GaussianVisibleUnit
config.progress_logger = rbm_logger.ProgressLogger()
config.train_params = tr
np_rand = np.random.RandomState(123)
# Weights
W = np_rand.uniform(low=-1. / 10, high=-1. / 10,
size=(3, 2)).astype(np.float32)
vb = np.array([-0.1, 0, 0.1], dtype=np.float32)
hb = np.array([0.01, -0.01], dtype=np.float32)
Wt = theano.shared(W, name='W')
vbt = theano.shared(vb, name='vbias')
hbt = theano.shared(hb, name='hbias')
g_rbm = rbm.RBM(config, W=Wt, h_bias=hbt, v_bias=vbt)
self.assertTrue(g_rbm)
self.assertTrue(isinstance(g_rbm.v_unit,
rbm_units.GaussianVisibleUnit))
self.assertTrue(isinstance(g_rbm.h_unit, rbm_units.RBMUnit))
self.assertTrue(
np.count_nonzero(g_rbm.W.get_value(borrow=True) - W) == 0)
self.assertTrue(
np.count_nonzero(g_rbm.v_bias.get_value(borrow=True) - vb) == 0)
self.assertTrue(
np.count_nonzero(g_rbm.h_bias.get_value(borrow=True) - hb) == 0)
x = sklearn.preprocessing.scale(
np.array([[200.0, 188., 7.], [150.0, 128., 0.], [250.0, 98., 3.]],
dtype=theano.config.floatX))
v = theano.shared(x)
_, _, h = g_rbm.sample_h_given_v(v)
_, _, vs = g_rbm.sample_v_given_h(h)
_, _, hs = g_rbm.sample_h_given_v(vs)
dw = T.dot(v.T, h) - T.dot(vs.T, hs)
dv = T.sum(v - vs, axis=0)
dh = T.sum(h - hs, axis=0)
gr = g_rbm.get_partial_derivatives(v, None)['gradients']
gdw, gdv, gdh = gr[0], gr[1], gr[2]
print gdw, gdv, gdh
compute_derivative = theano.function([], [dw, dv, dh, gdw, gdv, gdh])
for i in xrange(20):
a, b, c, d, e, f = compute_derivative()
# print a, b, c
print d, e, f
def train_kanade():
print "Testing RBM"
data_manager = store.StorageManager('Kanade/SimpleRBMTest')
# Load mnist hand digits
datasets = loader.load_kanade(n=500,
set_name='25_25',
emotions=['happy', 'sadness'],
pre={'scale2unit': True})
train_x, train_y = datasets[0]
sparsity_constraint = True
# Initialise the RBM and training parameters
tr = rbm_config.TrainParam(learning_rate=0.0001,
momentum_type=rbm_config.NESTEROV,
momentum=0.9,
weight_decay=0.001,
sparsity_constraint=sparsity_constraint,
sparsity_target=0.01,
sparsity_cost=1,
sparsity_decay=0.9,
dropout=True,
epochs=100)
n_visible = train_x.get_value().shape[1]
n_hidden = 500
config = rbm_config.RBMConfig(
v_n=n_visible,
v2_n=n_visible,
h_n=n_hidden,
v_unit=rbm_units.GaussianVisibleUnit,
associative=False,
cd_type=rbm_config.CLASSICAL,
cd_steps=1,
train_params=tr,
progress_logger=rbm_logger.ProgressLogger(img_shape=(25, 25)))
rbm = RBM(config)
print "... initialised RBM"
# Train RBM
rbm.train(train_x)
# Test RBM
rbm.reconstruct(train_x, k=5, plot_n=10, plot_every=1)
# Store Parameters
data_manager.persist(rbm)
def test_rbm():
print "Testing RBM"
# Load mnist hand digits
datasets = mnist_loader.load_digits(n=[100, 0, 100], digits=[1])
train_set_x, train_set_y = datasets[0]
test_set_x, test_set_y = datasets[2]
# Initialise the RBM and training parameters
tr = rbm_config.TrainParam(learning_rate=0.01,
momentum_type=NESTEROV,
momentum=0.5,
weight_decay=0.01,
sparsity_constraint=True,
sparsity_target=0.01,
sparsity_cost=0.01,
sparsity_decay=0.1)
n_visible = train_set_x.get_value().shape[1]
n_hidden = 2
config = rbm_config.RBMConfig(v_n=n_visible,
v2_n=n_visible,
h_n=n_hidden,
associative=False,
cd_type=CLASSICAL,
cd_steps=1,
train_params=tr,
progress_logger=ProgressLogger())
rbm = RBM(config)
print "... initialised RBM"
curr_dir = store.move_to(str(rbm))
print "... moved to {}".format(curr_dir)
# Train RBM
rbm.train(train_set_x)
# Test RBM
rbm.reconstruct(test_set_x, k=1, plot_n=20)
def test_digits(self):
tr = rbm_config.TrainParam(learning_rate=0.05,
momentum_type=rbm_config.CLASSICAL,
momentum=0.5,
weight_decay=0,
sparsity_constraint=False,
sparsity_target=0.1**9,
sparsity_cost=10**8,
sparsity_decay=0.9,
epochs=5)
config = rbm_config.RBMConfig()
config.v_n = 784
config.h_n = 100
config.v_unit = rbm_units.GaussianVisibleUnit
config.h_unit = rbm_units.ReLUnit
config.progress_logger = rbm_logger.ProgressLogger()
config.train_params = tr
np_rand = np.random.RandomState(123)
# Weights
W = np_rand.uniform(low=-1. / 10, high=1. / 10,
size=(784, 100)).astype(np.float32)
vb = np.zeros(784, dtype=np.float32)
hb = np.array(100, dtype=np.float32)
Wt = theano.shared(W, name='W')
vbt = theano.shared(vb, name='vbias')
hbt = theano.shared(hb, name='hbias')
g_rbm = rbm.RBM(config, W=Wt, h_bias=hbt, v_bias=vbt)
self.assertTrue(g_rbm)
self.assertTrue(isinstance(g_rbm.v_unit,
rbm_units.GaussianVisibleUnit))
self.assertTrue(isinstance(g_rbm.h_unit, rbm_units.RBMUnit))
self.assertTrue(
np.count_nonzero(g_rbm.W.get_value(borrow=True) - W) == 0)
self.assertTrue(
np.count_nonzero(g_rbm.v_bias.get_value(borrow=True) - vb) == 0)
self.assertTrue(
np.count_nonzero(g_rbm.h_bias.get_value(borrow=True) - hb) == 0)
tr, vl, te = mnist_loader.load_digits(n=[5, 10, 10],
pre={'scale': True})
v = tr[0]
# print 'inputs:'
# table = ss.itemfreq(v.get_value(borrow=True))
# x = [pt[0] for pt in table]
# y = [pt[1] for pt in table]
# plt.plot(x, y)
# plt.show()
# v = theano.shared(x)
_, _, h = g_rbm.sample_h_given_v(v)
_, _, vs = g_rbm.sample_v_given_h(h)
_, _, hs = g_rbm.sample_h_given_v(vs)
dw = T.dot(v.T, h) - T.dot(vs.T, hs)
dv = T.sum(v - vs, axis=0)
dh = T.sum(h - hs, axis=0)
gr = g_rbm.get_partial_derivatives(v, None)['gradients']
gdw, gdv, gdh = gr[0], gr[1], gr[2]
print gdw, gdv, gdh
compute_derivative = theano.function([], [dw, dv, dh, gdw, gdv, gdh])
for i in xrange(1):
a, b, c, d, e, f = compute_derivative()
# print a, b, c
print 'unfold'
print a[0], b[1:5], c[1:5]
print 'rbm'
print d[0], e[1:5], f[1:5]
def test_kanades_grrbm(self):
nvis = 625
nhid = 1000
train_n = 10000
batch_n = 20
tr = rbm_config.TrainParam(learning_rate=0.05,
momentum_type=rbm_config.CLASSICAL,
momentum=0.5,
weight_decay=0,
sparsity_constraint=False,
sparsity_target=0.1**9,
sparsity_cost=10**8,
sparsity_decay=0.9,
epochs=5,
batch_size=batch_n)
config = rbm_config.RBMConfig()
config.v_n = nvis
config.h_n = nhid
config.v_unit = rbm_units.GaussianVisibleUnit
config.h_unit = rbm_units.ReLUnit
config.progress_logger = rbm_logger.ProgressLogger()
config.train_params = tr
np_rand = np.random.RandomState(123)
# Weights
W = np_rand.normal(0, 0.01, size=(nvis, nhid)).astype(np.float32)
vb = np.zeros(nvis, dtype=np.float32)
hb = np.zeros(nhid, dtype=np.float32)
Wt = theano.shared(W, name='W')
vbt = theano.shared(vb, name='vbias')
hbt = theano.shared(hb, name='hbias')
g_rbm = rbm.RBM(config, W=Wt, h_bias=hbt, v_bias=vbt)
self.assertTrue(g_rbm)
self.assertTrue(isinstance(g_rbm.v_unit,
rbm_units.GaussianVisibleUnit))
self.assertTrue(isinstance(g_rbm.h_unit, rbm_units.RBMUnit))
self.assertTrue(
np.count_nonzero(g_rbm.W.get_value(borrow=True) - W) == 0)
self.assertTrue(
np.count_nonzero(g_rbm.v_bias.get_value(borrow=True) - vb) == 0)
self.assertTrue(
np.count_nonzero(g_rbm.h_bias.get_value(borrow=True) - hb) == 0)
tr, vl, te = k_loader.load_kanade(n=batch_n, pre={'scale': True})
v = tr[0]
vv = v.get_value(borrow=True).ravel()
table = ss.itemfreq(vv)
print table
x = [pt[0] for pt in table]
y = [pt[1] for pt in table]
plt.plot(x, y)
plt.show()
# v = theano.shared(x)
h = g_rbm.h_unit.activate(g_rbm.h_unit.scale(T.dot(v, W) + hb))
_, _, h = g_rbm.sample_h_given_v(v)
_, _, vs = g_rbm.sample_v_given_h(h)
_, _, hs = g_rbm.sample_h_given_v(vs)
dw = (T.dot(v.T, h) - T.dot(vs.T, hs)) / batch_n
dv = T.mean(v - vs, axis=0)
dh = T.mean(h - hs, axis=0)
gr = g_rbm.get_partial_derivatives(v, None)['gradients']
gdw, gdv, gdh = gr[0], gr[1], gr[2]
print gdw, gdv, gdh
compute_derivative = theano.function([], [dw, dv, dh, gdw, gdv, gdh])
for i in xrange(1):
a, b, c, d, e, f = compute_derivative()
# print a, b, c
print 'unfold'
print a[0][0:5], b[1:5], c[1:5]
print 'rbm'
print d[0][0:5], e[1:5], f[1:5]
def KanadeAssociativeRBM(cache=False, train_further=False):
print "Testing Associative RBM which tries to learn the ID map "
# print "Testing Associative RBM which tries to learn the following mapping: {anger, saddness, disgust} -> {sadness}, {contempt, happy, surprise} -> {happy}"
# project set-up
data_manager = store.StorageManager('Kanade/OptMFSparse0.01RBMTest',
log=True)
# data_manager = store.StorageManager('Kanade/OptAssociativeRBMTest', log=True)
shape = 25
dataset_name = 'sharp_equi{}_{}'.format(shape, shape)
# Load kanade database
mapping = None # id map
# mapping = {'anger': 'sadness', 'contempt': 'happy', 'disgust': 'sadness', 'fear': 'sadness', 'happy': 'happy',
# 'sadness': 'sadness', 'surprise': 'happy'}
train, valid, test = loader.load_kanade(pre={'scale': True},
set_name=dataset_name)
train_x, train_y = train
test_x, test_y = test
# Sample associated image
train_x_mapped, train_y_mapped = loader.sample_image(train_y,
mapping=mapping,
pre={'scale': True},
set_name=dataset_name)
test_x_mapped, test_y_mapped = loader.sample_image(test_y,
mapping=mapping,
pre={'scale': True},
set_name=dataset_name)
# Concatenate images
concat1 = T.concatenate([train_x, train_x_mapped], axis=1)
# concat2 = T.concatenate([train_x_mapped, train_x], axis=1)
# concat = T.concatenate([concat1, concat2], axis=0)
# train_tX = theano.function([], concat)()
train_tX = theano.function([], concat1)()
train_X = theano.shared(train_tX)
# Train classifier to be used for classifying reconstruction associated image layer
# mapped_data = loader.load_kanade(#emotions=['sadness', 'happy'],
# pre={'scale': True},
# set_name=dataset_name) # Target Image
# clf_orig = SimpleClassifier('logistic', mapped_data[0][0], mapped_data[0][1])
clf_orig = SimpleClassifier('logistic', train_x, train_y)
# Initialise RBM
tr = rbm_config.TrainParam(learning_rate=0.0001,
momentum_type=rbm_config.NESTEROV,
momentum=0.9,
weight_decay=0.0001,
sparsity_constraint=True,
sparsity_target=0.01,
sparsity_cost=100,
sparsity_decay=0.9,
batch_size=10,
epochs=10)
n_visible = shape * shape * 2
n_hidden = 500
config = rbm_config.RBMConfig()
config.v_n = n_visible
config.h_n = n_hidden
config.v_unit = rbm_units.GaussianVisibleUnit
# config.h_unit = rbm_units.ReLUnit
config.progress_logger = rbm_logger.ProgressLogger(img_shape=(shape * 2,
shape))
config.train_params = tr
rbm = RBM(config)
print "... initialised RBM"
# Load RBM (test)
loaded = data_manager.retrieve(str(rbm))
if loaded:
rbm = loaded
else:
rbm.set_initial_hidden_bias()
rbm.set_hidden_mean_activity(train_X)
# Train RBM - learn joint distribution
# rbm.pretrain_lr(train_x, train_x01)
for i in xrange(0, 10):
if not cache or train_further:
rbm.train(train_X)
data_manager.persist(rbm)
print "... reconstruction of associated images"
# Get reconstruction with train data to get 'mapped' images to train classifiers on
reconstruction = rbm.reconstruct(train_X,
1,
plot_n=100,
plot_every=1,
img_name='recon_train')
reconstruct_assoc_part = reconstruction[:, (shape**2):]
# Get associated images of test data
nsamples = np.random.normal(0, 1,
test_x.get_value(True).shape).astype(
np.float32)
initial_y = theano.shared(nsamples, name='initial_y')
utils.save_images(nsamples[0:100], 'initialisation.png', (10, 10),
(25, 25))
test_x_associated = rbm.reconstruct_association_opt(
test_x,
initial_y,
5,
0.,
plot_n=100,
plot_every=1,
img_name='recon_test_gibbs')
mf_recon = rbm.mean_field_inference_opt(test_x,
y=initial_y,
sample=False,
k=10,
img_name='recon_test_mf_raw')
# Concatenate images
test_MFX = theano.function([], T.concatenate([test_x, mf_recon],
axis=1))()
test_MF = theano.shared(test_MFX)
reconstruction = rbm.reconstruct(test_MF,
1,
plot_n=100,
plot_every=1,
img_name='recon_test_mf_recon')
mf_recon = reconstruction[:, (shape**2):]
print "... reconstructed"
# Classify the reconstructions
# 1. Train classifier on original images
score_orig = clf_orig.get_score(test_x_associated,
test_y_mapped.eval())
score_orig_mf = clf_orig.get_score(test_x_associated,
test_y_mapped.eval())
# 2. Train classifier on reconstructed images
clf_recon = SimpleClassifier('logistic', reconstruct_assoc_part,
train_y_mapped.eval())
score_retrain = clf_recon.get_score(test_x_associated,
test_y_mapped.eval())
score_retrain_mf = clf_recon.get_score(mf_recon, test_y_mapped.eval())
out_msg = '{} (orig, retrain):{},{}'.format(rbm, score_orig,
score_retrain)
out_msg2 = '{} (orig, retrain):{},{}'.format(rbm, score_orig_mf,
score_retrain_mf)
print out_msg
print out_msg2
def KanadeJointDBN(cache=False):
print "Testing JointDBN which tries to learn id map association"
# project set-up
data_manager = store.StorageManager('Kanade/JointDBN', log=True)
shape = 25
dataset_name = 'sharp_equi{}_{}'.format(shape, shape)
preprocessing = {'scale': True}
# Load kanade database
mapping = None
# mapping = {'anger': 'sadness',
# 'contempt': 'happy',
# 'disgust': 'sadness',
# 'fear': 'sadness',
# 'happy': 'happy',
# 'sadness': 'sadness',
# 'surprise': 'happy'}
dataset = loader.load_kanade( # n=3000,
pre=preprocessing, set_name=dataset_name)
mapped_dataset = loader.load_kanade( # n=3000,
# emotions=['sadness', 'happy'],
pre=preprocessing,
set_name=dataset_name) # Target Image
train, valid, test = dataset
train_x, train_y = train
test_x, test_y = test
# Sample associated image
train_x_ass, train_y_ass = loader.sample_image(train_y,
mapping=mapping,
pre=preprocessing,
set_name=dataset_name)
test_x_ass, test_y_ass = loader.sample_image(test_y,
mapping=mapping,
pre=preprocessing,
set_name=dataset_name)
# Initialise RBM parameters
base_tr = rbm_config.TrainParam(learning_rate=0.0001,
momentum_type=rbm_config.NESTEROV,
momentum=0.9,
weight_decay=0.0001,
sparsity_constraint=False,
sparsity_target=0.00001,
sparsity_decay=0.9,
sparsity_cost=10000,
epochs=100,
batch_size=10)
rest_tr = rbm_config.TrainParam(learning_rate=0.0001,
momentum_type=rbm_config.CLASSICAL,
momentum=0.5,
weight_decay=0.01,
epochs=100,
batch_size=10)
# Layer 1
# Layer 2
# Layer 3
topology = [2 * (shape**2), 100, 100]
# batch_size = 10
first_progress_logger = rbm_logger.ProgressLogger(img_shape=(shape * 2,
shape))
rest_progress_logger = rbm_logger.ProgressLogger()
first_rbm_config = rbm_config.RBMConfig(
train_params=base_tr, progress_logger=first_progress_logger)
first_rbm_config.v_unit = rbm_units.GaussianVisibleUnit
rest_rbm_config = rbm_config.RBMConfig(
train_params=rest_tr, progress_logger=rest_progress_logger)
rbm_configs = [first_rbm_config, rest_rbm_config, rest_rbm_config]
config = DBN.DBNConfig(topology=topology,
training_parameters=base_tr,
rbm_configs=rbm_configs,
data_manager=data_manager)
# construct the Deep Belief Network
dbn = DBN.DBN(config)
# Train DBN on concatenated images
train_tX = theano.function([], T.concatenate([train_x, train_x_ass],
axis=1))()
train_X = theano.shared(train_tX)
test_tX = theano.function([], T.concatenate([test_x, test_x_ass],
axis=1))()
test_X = theano.shared(test_tX)
test_tX2 = theano.function([],
T.concatenate(
[test_x, T.zeros_like(test_x)], axis=1))()
test_X2 = theano.shared(test_tX2)
origs = []
recons = []
recons2 = []
# Train DBN
dbn.pretrain(train_X,
cache=[True, True, False],
train_further=[True, True, True])
recon = dbn.reconstruct(train_X,
k=1,
plot_n=20,
img_name='stackedRBM_train_recon_{}_{}'.format(
topology, 0))
train_x_ass_recon = recon[:, shape**2:]
recon = dbn.reconstruct(test_X,
k=1,
plot_n=20,
img_name='stackedRBM_test_recon_{}_{}'.format(
topology, 0))
test_x_ass_recon = recon[:, shape**2:]
recon = dbn.reconstruct(test_X2,
k=2,
plot_n=20,
img_name='stackedRBM_test_zero_recon_{}_{}'.format(
topology, 0))
test_x_ass_recon2 = recon[:, shape**2:]
clf_recon = SimpleClassifier('logistic', train_x, train_y)
score_orig = clf_recon.get_score(test_x_ass_recon, test_y_ass.eval())
clf_recon.retrain(train_x_ass_recon, train_y_ass.eval())
score_recon = clf_recon.get_score(test_x_ass_recon, test_y_ass.eval())
score_recon2 = clf_recon.get_score(test_x_ass_recon2, test_y_ass.eval())
print 'classification rate: {}, {}, {}'.format(score_orig, score_recon,
score_recon2)
origs.append(score_orig)
recons.append(score_recon)
recons2.append(score_recon2)
def KanadeAssociativeDBN(cache=False):
print "Testing Associative RBM which tries to learn the following mapping: " \
"ID"
# "{anger, saddness, disgust} -> {sadness}, {contempt, happy, surprise} -> {happy}"
# project set-up
data_manager = store.StorageManager('Kanade/AssociativeDBNTest', log=True)
shape = 25
dataset_name = 'sharp_equi{}_{}'.format(shape, shape)
preprocessing = {'scale': True}
# Load kanade database
mapping = None
# mapping = {'anger': 'sadness',
# 'contempt': 'happy',
# 'disgust': 'sadness',
# 'fear': 'sadness',
# 'happy': 'happy',
# 'sadness': 'sadness',
# 'surprise': 'happy'}
dataset = loader.load_kanade(n=100,
pre=preprocessing,
set_name=dataset_name)
mapped_dataset = loader.load_kanade(
n=100,
# emotions=['sadness', 'happy'],
pre=preprocessing,
set_name=dataset_name) # Target Image
train, valid, test = dataset
train_x, train_y = train
test_x, test_y = test
# Sample associated image
train_x_ass, train_y_ass = loader.sample_image(train_y,
mapping=mapping,
pre=preprocessing,
set_name=dataset_name)
test_x_ass, test_y_ass = loader.sample_image(test_y,
mapping=mapping,
pre=preprocessing,
set_name=dataset_name)
# initialise AssociativeDBN
config = associative_dbn.DefaultADBNConfig()
# Gaussian Input Layer
bottom_tr = rbm_config.TrainParam(learning_rate=0.0001,
momentum_type=rbm_config.NESTEROV,
momentum=0.9,
weight_decay=0.0001,
epochs=20,
batch_size=10)
h_n = 150
bottom_logger = rbm_logger.ProgressLogger(img_shape=(shape, shape))
bottom_rbm = rbm_config.RBMConfig(v_unit=rbm_units.GaussianVisibleUnit,
v_n=shape**2,
h_n=h_n,
progress_logger=bottom_logger,
train_params=bottom_tr)
config.left_dbn.rbm_configs[0] = bottom_rbm
config.right_dbn.rbm_configs[0] = bottom_rbm
config.left_dbn.topology = [shape**2, h_n]
config.right_dbn.topology = [shape**2, h_n]
config.top_rbm.train_params.epochs = 20
config.top_rbm.train_params.batch_size = 10
config.n_association = 1000
config.reuse_dbn = True
adbn = associative_dbn.AssociativeDBN(config=config,
data_manager=data_manager)
# Plot sample
loader.save_faces(
train_x.get_value(borrow=True)[1:50],
tile=(10, 10),
img_name='n_orig.png',
)
loader.save_faces(train_x_ass.get_value(borrow=True)[1:50],
tile=(10, 10),
img_name='n_ass.png')
# Train classifier to be used for classifying reconstruction associated image layer
clf_orig = SimpleClassifier('knn', mapped_dataset[0][0],
mapped_dataset[0][1])
# Test DBN Performance
for i in xrange(0, 5):
# Train DBN - learn joint distribution
cache_left = [True]
cache_right = [True]
cache_top = False
cache = [cache_left, cache_right, cache_top]
adbn.train(train_x, train_x_ass, cache=cache)
print "... trained associative DBN"
# Reconstruct images
test_x_recon = adbn.recall(test_x, associate_steps=500, recall_steps=0)
print "... reconstructed images"
# Classify the reconstructions
# 1. Train classifier on original images
score_orig = clf_orig.get_score(test_x_recon, test_y_ass.eval())
# 2. Train classifier on reconstructed images - reconstruction obtained by right dbn
right_dbn = adbn.dbn_right
mapped_train_recon = right_dbn.reconstruct(
mapped_dataset[0][0],
k=1,
plot_n=100,
plot_every=1,
img_name='right_dbn_reconstruction')
clf_recon = SimpleClassifier('knn', mapped_train_recon,
mapped_dataset[0][1].eval())
score_retrain = clf_recon.get_score(test_x_recon, test_y_ass.eval())
out_msg = '{} (orig, retrain):{},{}'.format(adbn, score_orig,
score_retrain)
print out_msg