N_weights_enc, encoder, encoder_log_like = make_gaussian_nn(enc_layers)
N_weights_dec, decoder, decoder_log_like = make_binary_nn(dec_layers)
# run_aevb(train_images)
# with open('parameters.pkl') as f:
# parameters = pickle.load(f)
with open('mnist_models.pkl') as f:
trained_weights, all_mean, all_cov = pickle.load(f)
banded_cov = create_banded_cov(all_cov.shape[0], 100)
# Build likelihood model.
L2_reg = 1
layer_sizes = [784, 200, 100, 10]
num_weights, make_predictions, likelihood = make_classification_nn(
layer_sizes)
classifier_loglik = lambda image, c: make_predictions(
trained_weights, np.atleast_2d(image))[:, c]
image_prior = build_logprob_mvn(all_mean, banded_cov)
# Combine prior and likelihood.
model_ll = lambda image, c: image_prior(image) + classifier_loglik(
image, c)
def model_nll(image, c):
return -1 * model_ll(image, c)
model_nll_with_grad = value_and_grad(model_nll)
# Optimize a random image to maximize this likelihood.
cur_class = 1
if __name__ == '__main__':
# load_and_pickle_binary_mnist()
with open('../../../autopaint/mnist_binary_data.pkl') as f:
N_data, train_images, train_labels, test_images, test_labels = pickle.load(
f)
# Create aevb function
# Training parameters
D = train_images.shape[1]
enc_layers = [D, hidden_units, 2 * latent_dimensions]
dec_layers = [latent_dimensions, hidden_units, D]
N_weights_enc, encoder, encoder_log_like = make_gaussian_nn(enc_layers)
with open('mnist_models.pkl') as f:
trained_weights, all_mean, all_cov = pickle.load(f)
# Build likelihood model.
L2_reg = 1
layer_sizes = [784, 200, 100, 10]
N_weights_dec, decoder, decoder_log_like = make_classification_nn(
layer_sizes)
decoder = run_variational_network(train_labels, N_weights_dec, decoder,
decoder_log_like, trained_weights,
all_mean)
