valid_costs.append(ev_total_cost)
# Plot and summarize
values = { 'nepochs':count,
'has_any_valid_set': True,
'tr_total_cost':tr_total_cost,
'ev_total_cost':ev_total_cost,
'tr_recon_cost':tr_recon_cost,
'ev_recon_cost':ev_recon_cost,
'tr_kl_cost':tr_kl_cost,
'ev_kl_cost':ev_kl_cost,
'l2_weight':l2_weight,
'kl_weight':kl_weight,
'l2_cost':l2_cost
}
model.summarize_all(datasets, values)
# Manage learning rate.
n_lr = hps.learning_rate_n_to_compare
if len(train_costs) > n_lr and tr_total_cost > np.max(train_costs[-n_lr:]):
lr = session.run(model.learning_rate_decay_op)
print(" Decreasing learning rate to %f." % lr)
# Force the system to run n_lr times while at this lr.
train_costs.append(np.inf)
else:
train_costs.append(tr_total_cost)
if lr < lr_stop:
print("Stopping optimization based on learning rate criteria.")
break
valid_costs.append(ev_total_cost)
# Plot and summarize
values = { 'nepochs':count,
'has_any_valid_set': True,
'tr_total_cost':tr_total_cost,
'ev_total_cost':ev_total_cost,
'tr_recon_cost':tr_recon_cost,
'ev_recon_cost':ev_recon_cost,
'tr_kl_cost':tr_kl_cost,
'ev_kl_cost':ev_kl_cost,
'l2_weight':l2_weight,
'kl_weight':kl_weight,
'l2_cost':l2_cost
}
model.summarize_all(data, values)
# Manage learning rate.
n_lr = hps.learning_rate_n_to_compare
if len(train_costs) > n_lr and tr_total_cost > np.max(train_costs[-n_lr:]):
lr = session.run(model.learning_rate_decay_op)
print(" Decreasing learning rate to %f." % lr)
# Force the system to run n_lr times while at this lr.
train_costs.append(np.inf)
else:
train_costs.append(tr_total_cost)
if lr < lr_stop:
print("Stopping optimization based on learning rate criteria.")
break
