def iter_update(self, epoch, nb_batches, iter_update_batch):
status = super(MyBatchOptimizer, self).iter_update(epoch, nb_batches, iter_update_batch)
#status["reconstruction_error"] = self.model.reconstruction_error_function(X)
#status["lower_bound_train"] = self.model.get_likelihood_lower_bound(X_train)
status["lower_bound_validation"] = self.model.get_likelihood_lower_bound(X_valid)
status["log_likelihood_validation"] = self.model.log_likelihood_approximation_function(X_valid)
for k, v in status.items():
light.append(k, float(v))
if epoch % 10 == 0:
if save_codes is True:
plt.clf()
fig = plt.gcf()
code, code_sigma = self.model.encode(X_valid)
pca = PCA(n_components=2)
code_2d = pca.fit_transform(code)
if y is None:
plt.scatter(code_2d[:, 0], code_2d[:, 1])
else:
plt.scatter(code_2d[:, 0], code_2d[:, 1], c=y_valid)
fig.canvas.draw()
data = fig_to_list(fig)
light.append("codes", light.insert_blob(data))
if save_samples is True:
fig = plt.gcf()
X_ = self.model.sample(100, only_means=True)
X_ = get_2d_square_image_view(X_)
light.append("samples", light.insert_blob(X_.tolist()))
if learning_trajectory is True:
points, _ = self.model.encode(X[0:200])
points = points.ravel().tolist()
light.append("trajectories", {"points": points, "epoch": epoch, "seed": state})
if save_hidden_activations is True:
samples = range(0, 200)
hidden = []
for layer in self.model.encoder_layers:
h = layer(X[samples])
h = h.tolist()
hidden.append(h)
light.append("hidden_activations_encoder", light.insert_blob(hidden))
return status
optimization_procedure=(updates.adadelta, {"learning_rate" : 1.}),
batch_size=256,
whole_dataset_in_device=True,
verbose=1)
rng = RandomStreams(seed=1000)
#noise_function = lambda X_batch: X_batch * rng.binomial(size=X_batch.shape, p=0.7)
noise_function = None
model = autoencoder.Autoencoder(nnet_x_to_z, nnet_z_to_x, batch_optimizer, noise_function=noise_function, walkback=1)
model.fit(X)
"""
conv_filters = h.W.get_value()
conv_filters = conv_filters.reshape( (conv_filters.shape[0], conv_filters.shape[2], conv_filters.shape[3]))
grid_plot(conv_filters, imshow_options={"cmap": "gray"})
plt.savefig('out-filters-conv.png')
plt.show()
"""
filters = h.W.get_value().T
filters = get_2d_square_image_view(filters)
grid_plot(filters, imshow_options={"cmap":"gray"})
plt.savefig("out-filters.png")
plt.show()
plt.clf()
samples = model.sample(nb=100, nb_iterations=10000)
samples = get_2d_square_image_view(samples)
grid_plot(samples, imshow_options={"cmap": "gray"})
plt.savefig('out-samples.png')
plt.show()
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")