def build_dreamer(model,
X,
batch_optimizer=None,
loss_function=None,
rng=None,
model_input=None,
parse_grads=None,
input_type=T.matrix):
if batch_optimizer is None:
batch_optimizer = BatchOptimizer()
input_variables = OrderedDict()
input_variables["X"] = dict(tensor_type=input_type)
functions = dict()
inputs = theano.shared(X, borrow=True)
if loss_function is None:
# by default maximize the L2 norm
def loss_function(i, x):
return -(x**2).sum()
if model_input is None:
def model_input(inp):
return inp
if parse_grads is None:
def parse_grads(grads):
return [grad for grad in grads]
def loss(m, tensors):
#inputs_ = T.exp(inputs)
#inputs__ = inputs_ / inputs_.sum(axis=1).dimshuffle(0, 'x')
return loss_function(inputs, model.get_output(model_input(inputs))[0])
dreamer = Capsule(
input_variables,
DummyModel([inputs]),
loss,
functions=functions,
batch_optimizer=batch_optimizer,
parse_grads=parse_grads
)
dreamer.inputs = inputs
return dreamer
def test():
state = 10
rng = rng_mrg.MRG_RandomStreams(seed=state)
x_dim = 64
h_dim = 100
x_in = layers.InputLayer((None, x_dim))
l_out = Layer(x_in, num_units=h_dim)
model_encoder = LightweightModel([x_in], [l_out])
h_in = layers.InputLayer((None, h_dim))
l_out = Layer(h_in, num_units=x_dim)
model_decoder = LightweightModel([h_in], [l_out])
h_in = layers.InputLayer((None, 1))
l_out = Layer(h_in, num_units=h_dim)
model_prior = LightweightModel([h_in], [l_out])
def loss_function(model, tensors):
X = tensors["X"]
(h, u), = model_encoder.get_output(X, mode="sample", rng=rng)
#print(X.ndim, h.ndim)
(q_h_given_x, _), = model_encoder.get_output(X, mode="evaluate", on=h)
(p_x_given_h, _), = model_decoder.get_output(h, mode="evaluate", on=X)
ones = T.alloc(np.cast[theano.config.floatX](1.), *h.shape)
zeros = T.alloc(np.cast[theano.config.floatX](0.), *h.shape)
(p_h, _), = model_prior.get_output(zeros[:, 0:1], mode="evaluate", on=ones)
L = -((T.log(p_x_given_h).sum(axis=1) + T.log(p_h).sum(axis=1) - T.log(q_h_given_x).sum(axis=1)))
return (L.mean()), u
input_variables = OrderedDict(
X=dict(tensor_type=T.matrix),
)
functions = dict(
)
# sample function
nb = T.iscalar()
sample_input = T.alloc(np.cast[theano.config.floatX](0.), nb, 1)
from theano.updates import OrderedUpdates
u = OrderedUpdates()
(s_h, u_h), = model_prior.get_output(sample_input, mode="sample", rng=rng)
ones = T.alloc(np.cast[theano.config.floatX](1.), s_h.shape[0], x_dim)
(s_x, u_x), = model_decoder.get_output(s_h, mode="evaluate", on=ones)
sample = theano.function([nb], s_x, updates=u_x)
batch_optimizer = BatchOptimizer(verbose=1, max_nb_epochs=100,
batch_size=256,
optimization_procedure=(updates.rmsprop, {"learning_rate": 0.0001})
)
class Container(object):
def __init__(self, models):
self.models = models
def get_all_params(self):
return [p for m in self.models for p in m.get_all_params()]
models = [model_encoder, model_decoder]
models = Container(models)
darn = Capsule(input_variables,
models,
loss_function,
functions=functions,
batch_optimizer=batch_optimizer)
from sklearn.datasets import load_digits
def build_digits():
digits = load_digits()
X = digits.data
X = X.astype(np.float32) / 16.
return X, digits.images.shape[1:]
X, imshape = build_digits()
darn.fit(X=X)
s=(sample(20))
s = get_2d_square_image_view(s)
grid_plot(s, imshow_options={"cmap": "gray"})
plt.savefig("out.png")
input_variables = OrderedDict(
X=dict(tensor_type=T.matrix),
)
functions = dict(
sample=dict(
get_output=lambda model, X: model.get_output(X, sampler=True)[0],
params=["X"]
),
log_likelihood=dict(
get_output=lambda model, X: T.log(model.get_output(X)[0]).sum(axis=1),
params=["X"]
),
)
batch_optimizer = BatchOptimizer(verbose=1, max_nb_epochs=2,
batch_size=256,
optimization_procedure=(updates.rmsprop, {"learning_rate": 0.001})
)
nade = Capsule(input_variables, model,
loss_function,
functions=functions,
batch_optimizer=batch_optimizer)
nade.fit(X=X)
T = np.ones((100, x_dim)).astype(np.float32)
T = T[:, order]
s = get_2d_square_image_view(nade.sample(T))
grid_plot(s, imshow_options={"cmap": "gray"})
plt.savefig("out.png")
