training_info["epoch mean rewards"].append(epoch_rewards[-1])
training_info["value net loss"].append(critic_loss_total)
training_info["AE Hash loss"].append(ae_hash_loss)
if (i_epoch + 1) % num_avg_epoch:
training_info["past %d epochs mean reward" % (num_avg_epoch)] = \
(sum(training_info["epoch mean rewards"][-num_avg_epoch:]) / num_avg_epoch) \
if len(training_info["epoch mean rewards"]) >= num_avg_epoch else 0
# Print stats
print("\n\n============= Epoch: %d =============" % (i_epoch + 1))
print("epoch mean durations: %f" % (epoch_durations[-1]))
print("epoch mean rewards: %f" % (epoch_rewards[-1]))
print("Max reward achieved: %f" % training_info["max reward achieved"])
print("value net loss: %f" % critic_loss_total)
print("Autoencoder Hashing model loss: %f" % ae_hash_loss)
# Plot stats
if plot:
# plot_durations(training_info["epoch mean rewards"], training_info["value net loss"])
plot_durations(training_info["epoch mean rewards"],
training_info["value net loss"],
training_info["AE Hash loss"])
# Update counter
i_epoch += 1
# Every save_ckpt_interval, save a checkpoint according to current i_episode.
if i_epoch % save_ckpt_interval == 0:
save_checkpoint(ckpt_dir, actor_critic, ae_hash, ae_hash_optim,
i_epoch, **training_info)
training_info["epoch mean durations"].append(epoch_durations[-1])
training_info["epoch mean rewards"].append(epoch_rewards[-1])
training_info["extrinsic value net loss"].append(ex_value_net_mse)
training_info["intrinsic value net loss"].append(in_value_net_mse)
if (i_epoch + 1) % num_avg_epoch:
training_info["past %d epochs mean reward" % (num_avg_epoch)] = \
(sum(training_info["epoch mean rewards"][-num_avg_epoch:]) / num_avg_epoch) \
if len(training_info["epoch mean rewards"]) >= num_avg_epoch else 0
# Print stats
print("\n\n============= Epoch: %d =============" % (i_epoch + 1))
print("epoch mean durations: %f" % (epoch_durations[-1]))
print("epoch mean rewards: %f" % (epoch_rewards[-1]))
print("Max reward achieved: %f" % training_info["max reward achieved"])
print("extrinsic value net loss: %f" % ex_value_net_mse)
print("intrinsic value net loss: %f" % in_value_net_mse)
# Plot stats
if plot:
plot_durations(training_info["epoch mean rewards"],
training_info["extrinsic value net loss"],
training_info["intrinsic value net loss"])
# Update counter
i_epoch += 1
# Every save_ckpt_interval, save a checkpoint according to current i_episode.
# if i_epoch % save_ckpt_interval == 0:
# save_checkpoint(ckpt_dir, policy_net, value_net_in, value_net_ex,
# valuenet_in_optimizer, valuenet_ex_optimizer, simhash, i_epoch, **training_info)
# Record stats
training_info["epoch mean durations"].append(epoch_durations[-1])
training_info["epoch mean rewards"].append(epoch_rewards[-1])
if (i_epoch + 1) % num_avg_epoch:
training_info["past %d epochs mean reward" % (num_avg_epoch)] = \
(sum(training_info["epoch mean rewards"][-num_avg_epoch:]) / num_avg_epoch) \
if len(training_info["epoch mean rewards"]) >= num_avg_epoch else 0
# Print stats
print("\n\n============= Epoch: %d =============" % (i_epoch + 1))
print("epoch mean durations: %f" % (epoch_durations[-1]))
print("epoch mean rewards: %f" % (epoch_rewards[-1]))
print("Max reward achieved: %f" % training_info["max reward achieved"])
# Plot stats
if plot:
plot_durations(training_info["epoch mean rewards"])
# Update counter
i_epoch += 1
# Every save_ckpt_interval, save a checkpoint according to current i_episode.
if i_epoch % save_ckpt_interval == 0:
save_checkpoint(ckpt_dir,
policy_net,
policynet_optimizer,
i_epoch,
policy_lr=policy_lr,
**training_info)