# n_z=n_z) # dimensionality of latent space
#
# jnt_network_architecture = \
# dict(scope='joint',
# hidden_conv=False,
# n_hidden_recog_1=200, # 1st layer encoder neurons
# n_hidden_recog_2=200, # 2nd layer encoder neurons
# n_hidden_gener_1=200, # 1st layer decoder neurons
# n_hidden_gener_2=200, # 2nd layer decoder neurons
# n_input=147, # 21 bases for each function approximator
# n_z=n_z) # dimensionality of latent space
#create a new graph to ensure the resource is released after the training
#change to clear the default graph
tf.reset_default_graph()
vae_assoc_model, cost_hist = vae_assoc.train(data_sets, [img_network_architecture, jnt_network_architecture], binary=[True, False], weights=weights, assoc_lambda = assoc_lambda, learning_rate=0.001,
batch_size=batch_size, training_epochs=1000, display_step=5)
vae_assoc_model.save_model('output/model_batchsize{}_nz{}_lambda{}_weight{}.ckpt'.format(batch_size, n_z, assoc_lambda, weights[0]))
# time.sleep(10)
# #change to clear the default graph
# tf.reset_default_graph()
#
# vae_assoc_model = vae_assoc.AssocVariationalAutoEncoder([img_network_architecture, jnt_network_architecture],
# binary=[True, False],
# transfer_fct=tf.nn.relu,
# assoc_lambda=5,
# learning_rate=0.0001,
# batch_size=batch_size)
# vae_assoc_model.restore_model()