import tensorflow as tf
from architectures.pretrained_encoder import encoder_model
from architectures.latent_space import latent_space
from architectures.decoder import decoder_model
from dataprovider import DataProvider
import os
epochs = 50
batch_size = 20
latent_units = 200
l_rate = 0.0001
# data
data_provider = DataProvider(batch_size, root_folder='../data')
train_num_batches, val_num_batches = data_provider.get_num_batches()
training_dataset_init, val_dataset_init, images, labels = data_provider.get_data(
)
# model
encoder = encoder_model(images)
latent_vector, mean, stddev = latent_space(encoder, latent_units)
predictions = decoder_model(latent_vector)
# losses
generative_loss = tf.reduce_mean(tf.reduce_sum(tf.squared_difference(
predictions, labels),
axis=1),
axis=1)
latent_loss = tf.reduce_mean(0.5 * tf.reduce_sum(
tf.square(mean) + tf.square(stddev) - tf.log(1e-8 + tf.square(stddev)) - 1,
1))
