def measure_marginal_log_likelihood(model, dataset, subdataset, seed=123, minibatch_size=20, num_samples=50):
print "Measuring {} log likelihood".format(subdataset)
srng = utils.srng(seed)
test_x = dataset.data[subdataset]
n_examples = test_x.get_value(borrow=True).shape[0]
if n_examples % minibatch_size == 0:
num_minibatches = n_examples // minibatch_size
else:
num_minibatches = n_examples // minibatch_size + 1
index = T.lscalar('i')
minibatch = dataset.minibatchIindex_minibatch_size(index, minibatch_size, subdataset=subdataset, srng=srng)
log_marginal_likelihood_estimate = model.log_marginal_likelihood_estimate(minibatch, num_samples, srng)
get_log_marginal_likelihood = theano.function([index], T.sum(log_marginal_likelihood_estimate))
pbar = progressbar.ProgressBar(maxval=num_minibatches).start()
sum_of_log_likelihoods = 0.
for i in xrange(num_minibatches):
summand = get_log_marginal_likelihood(i)
sum_of_log_likelihoods += summand
pbar.update(i)
pbar.finish()
marginal_log_likelihood = sum_of_log_likelihoods/n_examples
return marginal_log_likelihood
def get_samples(model, num_samples, seed=123):
srng = utils.srng(seed)
prior_samples = model.prior.samplesIshape_srng((num_samples, model.first_p_layer_weights_np().shape[0]), srng)
samples = [prior_samples]
for layer in model.p_layers[:-1]:
samples.append(layer.samplesIx_srng(samples[-1], srng))
samples_function = theano.function([], model.p_layers[-1].meanIx(samples[-1]))
return reshape_and_tile_images(samples_function())
def measure_marginal_log_likelihood(self,
dataset,
subdataset,
batch_size,
num_samples,
seed=123,
z_mu=None):
'''
method used from iwae code, name of the method to be used to keep track of
what is being implemented.
The method in iwae takes in:
model: we have direct access to model here as this method is being made part
of the same class
dataset: the input dataset
subdataset: this is to identify whether it is training dataset or testing one,
and we will do it accordingly.
batch_size: this will help us determine the num of batches that has to be used
num_samples
'''
print("Measuring {} log likelihood".format(subdataset))
srng = utils.srng(seed)
if subdataset == 'train':
n_examples = mnist.train.num_examples
else:
n_examples = mnist.test.num_examples
if n_examples % batch_size == 0:
num_minibatches = n_examples // batch_size
else:
num_minibatches = n_examples // batch_size + 1
index = T.lscalar('i')
minibatch = dataset.minibatchIindex_minibatch_size(
index, batch_size, subdataset=subdataset, srng=srng)
log_marginal_likelihood_estimate = self.log_marginal_likelihood_estimate(
minibatch, num_samples, srng)
get_log_marginal_likelihood = theano.function(
[index], T.sum(log_marginal_likelihood_estimate))
sum_of_log_likelihoods = 0.
for i in range(num_minibatches):
summand = get_log_marginal_likelihood(i)
sum_of_log_likelihoods += summand
marginal_log_likelihood = sum_of_log_likelihoods / n_examples
return marginal_log_likelihood
def checkpoint0(dataset):
data_dimension = dataset.get_data_dim()
model = iwae.random_iwae(latent_units=[data_dimension] + latent_units,
hidden_units_q=hidden_units_q,
hidden_units_p=hidden_units_p,
dataset=dataset
)
srng = utils.srng()
optimizer = optimizers.Adam(model=model, learning_rate=1e-3)
return model, optimizer, srng
def get_samples(model, num_samples, seed=123):
srng = utils.srng(seed)
prior_samples = model.prior.samplesIshape_srng(
(num_samples, model.first_p_layer_weights_np().shape[0]), srng)
samples = [prior_samples]
for layer in model.p_layers[:-1]:
samples.append(layer.samplesIx_srng(samples[-1], srng))
samples_function = theano.function([],
model.p_layers[-1].meanIx(samples[-1]))
return reshape_and_tile_images(samples_function())
def load_checkpoint(directory_name, i):
'''Loads model, optimizer, and random number generator from a pickle file named training_state[i].pkl
Returns -1, None, None, None if loading failed
Returns i, model, optimizer, random number generator if loading succeedeed'''
try:
load_from_filename = os.path.join(directory_name, "training_state{}.pkl".format(i))
with open(load_from_filename, "rb") as f:
model, optimizer, rstate = pkl.load(f)
srng = utils.srng()
srng.rstate = rstate
loaded_checkpoint = i
except:
loaded_checkpoint = -1
model, optimizer, srng = None, None, None
return loaded_checkpoint, model, optimizer, srng
def get_units_variances(model, dataset):
srng = utils.srng()
x = dataset.minibatchIindex_minibatch_size(0, 500, subdataset='train', srng=srng)
samples = model.q_samplesIx_srng(x, srng)
means = []
for layer, x in zip(model.q_layers, samples):
mean, _ = layer.mean_sigmaIx(x)
means.append(mean)
mean_fun = theano.function([], means)
mean_vals = mean_fun()
vars_of_means = [np.var(mean_val, axis=0) for mean_val in mean_vals]
return vars_of_means
def get_units_variances(model, dataset):
srng = utils.srng()
x = dataset.minibatchIindex_minibatch_size(0,
500,
subdataset='train',
srng=srng)
samples = model.q_samplesIx_srng(x, srng)
means = []
for layer, x in zip(model.q_layers, samples):
mean, _ = layer.mean_sigmaIx(x)
means.append(mean)
mean_fun = theano.function([], means)
mean_vals = mean_fun()
vars_of_means = [np.var(mean_val, axis=0) for mean_val in mean_vals]
return vars_of_means
def measure_marginal_log_likelihood(model,
dataset,
subdataset,
seed=123,
minibatch_size=20,
num_samples=50):
print "Measuring {} log likelihood".format(subdataset)
srng = utils.srng(seed)
test_x = dataset.data[subdataset]
n_examples = test_x.get_value(borrow=True).shape[0]
if n_examples % minibatch_size == 0:
num_minibatches = n_examples // minibatch_size
else:
num_minibatches = n_examples // minibatch_size + 1
index = T.lscalar('i')
minibatch = dataset.minibatchIindex_minibatch_size(index,
minibatch_size,
subdataset=subdataset,
srng=srng)
log_marginal_likelihood_estimate = model.log_marginal_likelihood_estimate(
minibatch, num_samples, srng)
get_log_marginal_likelihood = theano.function(
[index], T.sum(log_marginal_likelihood_estimate))
pbar = progressbar.ProgressBar(maxval=num_minibatches).start()
sum_of_log_likelihoods = 0.
for i in xrange(num_minibatches):
summand = get_log_marginal_likelihood(i)
sum_of_log_likelihoods += summand
pbar.update(i)
pbar.finish()
marginal_log_likelihood = sum_of_log_likelihoods / n_examples
return marginal_log_likelihood
def checkpoint0(dataset):
if dataset == 'MNIST':
continuous=True
else:
RuntimeError('Case not implemented')
load_dataset_from_name(dataset)
model = VA(
numberOfInducingPoints, # Number of inducing ponts in sparse GP
batchSize, # Size of mini batch
dimX, # Dimensionality of the latent co-ordinates
dimZ, # Dimensionality of the latent variables
x_train, # [NxP] matrix of observations
kernelType=kernelType,
encoderType_qX=encoderType_qX, # 'FreeForm', 'MLP', 'Kernel'.
encoderType_rX=encoderType_rX, # 'FreeForm', 'MLP', 'Kernel', 'NoEncoding'.
encoderType_ru=encoderType_ru, # 'FreeForm', 'MLP', 'NoEncoding'
Xu_optimise=Xu_optimise,
numHiddenUnits_encoder=numHiddenUnits_encoder,
numHiddentUnits_decoder=numHiddentUnits_decoder,
continuous=continuous
)
model.construct_L_using_r()
model.setKernelParameters(0.01, 5*np.ones((2,)),
1e-100, 0.5,
[1e-10,1e-10], [10,10] )
model.randomise()
model.constructUpdateFunction()
model = model.randomise()
srng = utils.srng()
return model, srng
