torch.tensor(np.array([done])).to(device).float()))
state = next_state
episode_rewards += reward
if done:
break
rewards.append(episode_rewards)
# Train the model if memory is sufficient
if len(memory) + len(episode_transitions) >= min_buffer:
for i in range(train_steps):
loss = optimize(model, target, memory, episode_transitions, optimizer)
losses.append(loss.item())
memory.extend(episode_transitions)
# Update target network every once in a while
if episode % target_update == 0:
target.load_state_dict(model.state_dict())
target.eval()
if episode % print_interval == 0 and episode > 0:
print("[Episode {}] | avg rewards : {:.3f} | s.d. rewards: {:.3f} | avg loss : {:.10f} | buffer size : {} | epsilon : {:.1f}%".format(
episode, np.mean(rewards), np.std(rewards), np.mean(losses), len(memory), epsilon*100))
rewards = []
losses = []
if episode % SAVE_INTERVAL == 0 and episode > 0:
torch.save(model.state_dict(), SAVE_PATH)
episode_rewards += reward
if done:
break
rewards.append(episode_rewards)
# Train the model if memory is sufficient
if len(memory_success) >= min_buffer and len(
memory_failure) >= min_buffer:
for i in range(train_steps):
loss = optimize(model, target, memory_success, memory_failure,
episode_transitions, optimizer)
losses.append(loss.item())
# Update target network every once in a while
if episode % target_update == 0:
target.load_state_dict(model.state_dict())
target.eval()
if episode_rewards > 0:
memory_success.extend(episode_transitions)
else:
memory_failure.extend(episode_transitions)
if episode % print_interval == 0 and episode > 0:
print(
"[Episode {}] | avg rewards : {:.3f} | s.d. rewards: {:.3f} | avg loss : {:.10f} | succ. buffer : {} | fail. buffer : {}"
.format(episode, np.mean(rewards), np.std(rewards),
np.mean(losses), len(memory_success),
len(memory_failure)))
rewards = []
losses = []