def load_letters(which_letter=None):
datadict = np.load('data/letters.npz')
images, labels = datadict['images'], datadict['labels']
if which_letter is not None:
images = images[labels == string.lowercase.index(which_letter)]
return theano.shared(floatX(images), borrow=True), labels
def load_letters(which_letter=None):
datadict = np.load('data/letters.npz')
images, labels = datadict['images'], datadict['labels']
if which_letter is not None:
images = images[labels == string.lowercase.index(which_letter)]
return theano.shared(floatX(images), borrow=True), labels
def load_mice(N, permute=True, addnoise=True):
data = np.load("data/images_for_vae.npy").astype(theano.config.floatX)
if permute:
data = np.random.permutation(data)
data = data.reshape(data.shape[0], -1)[:N]
data /= data.max()
if addnoise:
data += 1e-3 * np.random.normal(size=data.shape)
return theano.shared(floatX(data), borrow=True)
def load_mice(N, permute=True, addnoise=True):
data = np.load('data/images_for_vae.npy').astype(theano.config.floatX)
print data[0].shape
if permute:
data = np.random.permutation(data)
data = data.reshape(data.shape[0], -1)[:N]
data /= data.max()
if addnoise:
data += 1e-3 * np.random.normal(size=data.shape)
return theano.shared(floatX(data), borrow=True)
def vlb(X, N, M, L):
def sample_z(mu, log_sigmasq):
eps = srng.normal((M, z_dim), dtype=theano.config.floatX)
return mu + T.exp(0.5 * log_sigmasq) * eps
mu, log_sigmasq = encode(X)
logpxz = sum(loglike(X, decode(sample_z(mu, log_sigmasq)))
for l in xrange(L)) / floatX(L)
minibatch_val = -kl_to_prior(mu, log_sigmasq) + logpxz
return minibatch_val / M # NOTE: multiply by N for overall vlb
def vlb(X, N, M, L):
def sample_z(mu, log_sigmasq):
eps = srng.normal((M, z_dim), dtype=theano.config.floatX)
return mu + T.exp(0.5 * log_sigmasq) * eps
mu, log_sigmasq = encode(X)
logpxz = sum(
loglike(X, decode(sample_z(mu, log_sigmasq)))
for l in xrange(L)) / floatX(L)
minibatch_val = -kl_to_prior(mu, log_sigmasq) + logpxz
return minibatch_val / M # NOTE: multiply by N for overall vlb
def adam(cost, params):
grads = T.grad(cost=cost, wrt=params)
ms = [shared_zeros_like(p) for p in params]
vs = [shared_zeros_like(p) for p in params]
t = theano.shared(floatX(1))
def make_update(p, g, m, v, t):
m_new = beta_1*m + (1.-beta_1)*g
v_new = beta_2*v + (1.-beta_2)*g**2
mhat = m / (1.-beta_1**t)
vhat = v / (1.-beta_2**t)
p_new = p - alpha * mhat / (T.sqrt(vhat) + epsilon)
return [(m, m_new), (v, v_new), (p, p_new)]
return [(t, t+1)] + concat(
make_update(p, g, m, v, t) for p,g,m,v in zip(params, grads, ms, vs))
def adam(cost, params):
grads = T.grad(cost=cost, wrt=params)
ms = [shared_zeros_like(p) for p in params]
vs = [shared_zeros_like(p) for p in params]
t = theano.shared(floatX(1))
def make_update(p, g, m, v, t):
m_new = beta_1 * m + (1. - beta_1) * g
v_new = beta_2 * v + (1. - beta_2) * g**2
mhat = m / (1. - beta_1**t)
vhat = v / (1. - beta_2**t)
p_new = p - alpha * mhat / (T.sqrt(vhat) + epsilon)
return [(m, m_new), (v, v_new), (p, p_new)]
return [(t, t + 1)] + concat(
make_update(p, g, m, v, t)
for p, g, m, v in zip(params, grads, ms, vs))
def load_pendulum(N, permute=True, addnoise=True):
with open('data/pendulous.pkl') as infile:
images = pickle.load(infile).astype(theano.config.floatX)
if permute:
images = np.random.permutation(images)
images = images[:N]
images -= images.min()
images /= images.max()
if addnoise:
images += 1e-2*np.random.normal(size=images.shape)
images = np.reshape(images, (images.shape[0], -1))
return theano.shared(floatX(images), borrow=True)
def load_pendulum(N, permute=True, addnoise=True):
with open('data/pendulous.pkl') as infile:
images = pickle.load(infile).astype(theano.config.floatX)
if permute:
images = np.random.permutation(images)
images = images[:N]
images -= images.min()
images /= images.max()
if addnoise:
images += 1e-2 * np.random.normal(size=images.shape)
images = np.reshape(images, (images.shape[0], -1))
return theano.shared(floatX(images), borrow=True)
def init_tensor(shape, name=None):
return theano.shared(floatX(1e-2 * np.random.normal(size=shape)),
borrow=True,
name=name)
def init_tensor(shape, name=None):
return theano.shared(
floatX(1e-2 * np.random.normal(size=shape)),
borrow=True, name=name)