episode_count += 1
reward_history.append(episode_reward)
pb.set_description(
f"episode: {episode_count}, reward: {episode_reward}, eps: {eps_schedule.get(i)*100:.2f}%"
)
plotter.plot('episode reward', 'episode return', "Episode Return",
episode_count, episode_reward)
plotter.plot('episode reward', 'average return', "Episode Return",
episode_count,
sum(reward_history) / len(reward_history))
episode_reward = 0
if episode_count > 0 and episode_count % params.save_frequency == 0:
agent.save(chk_dir / f"checkpoint-episode-{episode_count}.pt")
obs = next_obs
if i < 0: continue
if i % params.target_sync == 0:
agent.sync_target()
if i % params.train_frequency == 0:
opt.zero_grad()
*batch, _ = memory.sample(params.batch_size)
loss = agent.calculate_loss(*batch)
loss.mean().backward()
opt.step()
pb.update(1)
memory.store_effect(idx, action, np.sign(reward), done)
if done:
next_obs = env.reset()
episode_count += 1
reward_history.append(episode_reward)
pb.set_description(f"episode: {episode_count}, reward: {episode_reward}, eps: {eps_schedule.get(i)*100:.2f}%")
plotter.plot('episode reward', 'episode return', "Episode Return", episode_count, episode_reward)
plotter.plot('episode reward', 'average return', "Episode Return", episode_count, sum(reward_history) / len(reward_history))
episode_reward = 0
if episode_count > 0 and episode_count % params.save_frequency == 0:
agent.save(chk_dir/ f"checkpoint-episode-{episode_count}.pt")
# torch.save(memory, chk_dir / f"memory.pt")
obs = next_obs
if i < 0: continue
if i % params.target_sync == 0:
agent.sync_target()
if i % params.train_frequency == 0:
opt.zero_grad()
*batch, _ = memory.sample(params.batch_size)
loss = agent.calculate_loss(*batch, double=True)
loss.mean().backward()
opt.step()
pb.update(1)