def test_shuffled_index_batches_generates_all_indices(self):
dataset_size = 16
batch_size = 4
stream = ppo.shuffled_index_batches(dataset_size, batch_size)
indices = np.reshape(
list(itertools.islice(stream, dataset_size // batch_size)), -1)
self.assertEqual(set(indices), set(range(dataset_size)))
def test_shuffled_index_batches_generates_valid_batch(self): dataset_size = 16 batch_size = 4 stream = ppo.shuffled_index_batches(dataset_size, batch_size) batch = next(stream) self.assertEqual(batch.shape, (batch_size,)) # Assert that all indices are different. self.assertEqual(len(set(batch)), batch_size)
def test_shuffled_index_batches_gives_different_permutations(self): dataset_size = 256 batch_size = 8 stream1 = ppo.shuffled_index_batches(dataset_size, batch_size) stream2 = ppo.shuffled_index_batches(dataset_size, batch_size) self.assertFalse(np.array_equal(next(stream1), next(stream2)))
def train_epoch(self, evaluate=True):
"""Train one PPO epoch."""
epoch_start_time = time.time()
# Evaluate the policy.
policy_eval_start_time = time.time()
if evaluate and (self._epoch + 1) % self._eval_every_n == 0:
self._rng, key = jax_random.split(self._rng, num=2)
self.evaluate()
policy_eval_time = ppo.get_time(policy_eval_start_time)
trajectory_collection_start_time = time.time()
logging.vlog(1, 'PPO epoch [% 6d]: collecting trajectories.',
self._epoch)
self._rng, key = jax_random.split(self._rng)
trajs, _, timing_info, self._model_state = self.collect_trajectories(
train=True, temperature=1.0)
trajs = [(t[0], t[1], t[2], t[4]) for t in trajs]
self._should_reset = False
trajectory_collection_time = ppo.get_time(
trajectory_collection_start_time)
logging.vlog(1, 'Collecting trajectories took %0.2f msec.',
trajectory_collection_time)
rewards = np.array([np.sum(traj[2]) for traj in trajs])
avg_reward = np.mean(rewards)
std_reward = np.std(rewards)
max_reward = np.max(rewards)
min_reward = np.min(rewards)
self._log('train', 'train/reward_mean_truncated', avg_reward)
if evaluate and not self._separate_eval:
metrics = {'raw': {1.0: {'mean': avg_reward, 'std': std_reward}}}
ppo.write_eval_reward_summaries(metrics, self._log, self._epoch)
logging.vlog(1,
'Rewards avg=[%0.2f], max=[%0.2f], min=[%0.2f], all=%s',
avg_reward, max_reward, min_reward,
[float(np.sum(traj[2])) for traj in trajs])
logging.vlog(
1, 'Trajectory Length average=[%0.2f], max=[%0.2f], min=[%0.2f]',
float(sum(len(traj[0]) for traj in trajs)) / len(trajs),
max(len(traj[0]) for traj in trajs),
min(len(traj[0]) for traj in trajs))
logging.vlog(2, 'Trajectory Lengths: %s',
[len(traj[0]) for traj in trajs])
preprocessing_start_time = time.time()
(padded_observations, padded_actions, padded_rewards, reward_mask,
padded_infos) = self._preprocess_trajectories(trajs)
preprocessing_time = ppo.get_time(preprocessing_start_time)
logging.vlog(1, 'Preprocessing trajectories took %0.2f msec.',
ppo.get_time(preprocessing_start_time))
logging.vlog(1, 'Padded Observations\' shape [%s]',
str(padded_observations.shape))
logging.vlog(1, 'Padded Actions\' shape [%s]',
str(padded_actions.shape))
logging.vlog(1, 'Padded Rewards\' shape [%s]',
str(padded_rewards.shape))
# Some assertions.
B, RT = padded_rewards.shape # pylint: disable=invalid-name
B, AT = padded_actions.shape # pylint: disable=invalid-name
assert (B, RT) == reward_mask.shape
assert B == padded_observations.shape[0]
log_prob_recompute_start_time = time.time()
# TODO(pkozakowski): The following commented out code collects the network
# predictions made while stepping the environment and uses them in PPO
# training, so that we can use non-deterministic networks (e.g. with
# dropout). This does not work well with serialization, so instead we
# recompute all network predictions. Let's figure out a solution that will
# work with both serialized sequences and non-deterministic networks.
# assert ('log_prob_actions' in padded_infos and
# 'value_predictions' in padded_infos)
# These are the actual log-probabs and value predictions seen while picking
# the actions.
# actual_log_probabs_traj = padded_infos['log_prob_actions']
# actual_value_predictions_traj = padded_infos['value_predictions']
# assert (B, T, C) == actual_log_probabs_traj.shape[:3]
# A = actual_log_probabs_traj.shape[3] # pylint: disable=invalid-name
# assert (B, T, 1) == actual_value_predictions_traj.shape
del padded_infos
# TODO(afrozm): log-probabs doesn't need to be (B, T+1, C, A) it can do with
# (B, T, C, A), so make that change throughout.
# NOTE: We don't have the log-probabs and value-predictions for the last
# observation, so we re-calculate for everything, but use the original ones
# for all but the last time-step.
self._rng, key = jax_random.split(self._rng)
(log_probabs_traj,
value_predictions_traj) = (self._policy_and_value_net_apply(
padded_observations,
weights=self._policy_and_value_net_weights,
state=self._model_state,
rng=key,
))
assert (B, AT) == log_probabs_traj.shape[:2]
assert (B, AT) == value_predictions_traj.shape
# TODO(pkozakowski): Commented out for the same reason as before.
# Concatenate the last time-step's log-probabs and value predictions to the
# actual log-probabs and value predictions and use those going forward.
# log_probabs_traj = np.concatenate(
# (actual_log_probabs_traj, log_probabs_traj[:, -1:, :]), axis=1)
# value_predictions_traj = np.concatenate(
# (actual_value_predictions_traj, value_predictions_traj[:, -1:, :]),
# axis=1)
log_prob_recompute_time = ppo.get_time(log_prob_recompute_start_time)
# Compute value and ppo losses.
self._rng, key1 = jax_random.split(self._rng, num=2)
logging.vlog(2, 'Starting to compute P&V loss.')
loss_compute_start_time = time.time()
(cur_combined_loss, component_losses, summaries,
self._model_state) = (ppo.combined_loss(
self._policy_and_value_net_weights,
log_probabs_traj,
value_predictions_traj,
self._policy_and_value_net_apply,
padded_observations,
padded_actions,
self._rewards_to_actions,
padded_rewards,
reward_mask,
nontrainable_params=self._nontrainable_params,
state=self._model_state,
rng=key1))
loss_compute_time = ppo.get_time(loss_compute_start_time)
(cur_ppo_loss, cur_value_loss, cur_entropy_bonus) = component_losses
logging.vlog(
1,
'Calculating P&V loss [%10.2f(%10.2f, %10.2f, %10.2f)] took %0.2f msec.',
cur_combined_loss, cur_ppo_loss, cur_value_loss, cur_entropy_bonus,
ppo.get_time(loss_compute_start_time))
self._rng, key1 = jax_random.split(self._rng, num=2)
logging.vlog(1, 'Policy and Value Optimization')
optimization_start_time = time.time()
keys = jax_random.split(key1, num=self._n_optimizer_steps)
opt_step = 0
opt_batch_size = min(self._optimizer_batch_size, B)
index_batches = ppo.shuffled_index_batches(dataset_size=B,
batch_size=opt_batch_size)
for (index_batch, key) in zip(index_batches, keys):
k1, k2, k3 = jax_random.split(key, num=3)
t = time.time()
# Update the optimizer state on the sampled minibatch.
self._policy_and_value_opt_state, self._model_state = (
ppo.policy_and_value_opt_step(
# We pass the optimizer slots between PPO epochs, so we need to
# pass the optimization step as well, so for example the
# bias-correction in Adam is calculated properly. Alternatively we
# could reset the slots and the step in every PPO epoch, but then
# the moment estimates in adaptive optimizers would never have
# enough time to warm up. So it makes sense to reuse the slots,
# even though we're optimizing a different loss in every new
# epoch.
self._total_opt_step,
self._policy_and_value_opt_state,
self._policy_and_value_opt_update,
self._policy_and_value_get_params,
self._policy_and_value_net_apply,
log_probabs_traj[index_batch],
value_predictions_traj[index_batch],
padded_observations[index_batch],
padded_actions[index_batch],
self._rewards_to_actions,
padded_rewards[index_batch],
reward_mask[index_batch],
nontrainable_params=self._nontrainable_params,
state=self._model_state,
rng=k1))
opt_step += 1
self._total_opt_step += 1
# Compute the approx KL for early stopping. Use the whole dataset - as we
# only do inference, it should fit in the memory.
(log_probab_actions_new, _) = (self._policy_and_value_net_apply(
padded_observations,
weights=self._policy_and_value_net_weights,
state=self._model_state,
rng=k2))
action_mask = np.dot(np.pad(reward_mask, ((0, 0), (0, 1))),
self._rewards_to_actions)
approx_kl = ppo.approximate_kl(log_probab_actions_new,
log_probabs_traj, action_mask)
early_stopping = approx_kl > 1.5 * self._target_kl
if early_stopping:
logging.vlog(
1,
'Early stopping policy and value optimization after %d steps, '
'with approx_kl: %0.2f', opt_step, approx_kl)
# We don't return right-away, we want the below to execute on the last
# iteration.
t2 = time.time()
if (opt_step % self._print_every_optimizer_steps == 0
or opt_step == self._n_optimizer_steps or early_stopping):
# Compute and log the loss.
(combined_loss, component_losses, _,
self._model_state) = (ppo.combined_loss(
self._policy_and_value_net_weights,
log_probabs_traj,
value_predictions_traj,
self._policy_and_value_net_apply,
padded_observations,
padded_actions,
self._rewards_to_actions,
padded_rewards,
reward_mask,
nontrainable_params=self._nontrainable_params,
state=self._model_state,
rng=k3))
logging.vlog(
1, 'One Policy and Value grad desc took: %0.2f msec',
ppo.get_time(t, t2))
(ppo_loss, value_loss, entropy_bonus) = component_losses
logging.vlog(
1, 'Combined Loss(value, ppo, entropy_bonus) [%10.2f] ->'
' [%10.2f(%10.2f,%10.2f,%10.2f)]', cur_combined_loss,
combined_loss, ppo_loss, value_loss, entropy_bonus)
if early_stopping:
break
optimization_time = ppo.get_time(optimization_start_time)
logging.vlog(
1, 'Total Combined Loss reduction [%0.2f]%%',
(100 *
(cur_combined_loss - combined_loss) / np.abs(cur_combined_loss)))
summaries.update({
'n_optimizer_steps': opt_step,
'approx_kl': approx_kl,
})
for (name, value) in summaries.items():
self._log('train', 'train/{}'.format(name), value)
logging.info(
'PPO epoch [% 6d], Reward[min, max, avg] [%5.2f,%5.2f,%5.2f], Combined'
' Loss(ppo, value, entropy) [%2.5f(%2.5f,%2.5f,%2.5f)]',
self._epoch, min_reward, max_reward, avg_reward, combined_loss,
ppo_loss, value_loss, entropy_bonus)
# Bump the epoch counter before saving a checkpoint, so that a call to
# save() after the training loop is a no-op if a checkpoint was saved last
# epoch - otherwise it would bump the epoch counter on the checkpoint.
last_epoch = self._epoch
self._epoch += 1
# Save parameters every time we see the end of at least a fraction of batch
# number of trajectories that are done (not completed -- completed includes
# truncated and done).
# Also don't save too frequently, enforce a minimum gap.
policy_save_start_time = time.time()
# TODO(afrozm): Refactor to trax.save_trainer_state.
if (self._n_trajectories_done >=
self._done_frac_for_policy_save * self.train_env.batch_size
and self._epoch % self._save_every_n == 0) or self._async_mode:
self.save()
policy_save_time = ppo.get_time(policy_save_start_time)
epoch_time = ppo.get_time(epoch_start_time)
timing_dict = {
'epoch': epoch_time,
'policy_eval': policy_eval_time,
'trajectory_collection': trajectory_collection_time,
'preprocessing': preprocessing_time,
'log_prob_recompute': log_prob_recompute_time,
'loss_compute': loss_compute_time,
'optimization': optimization_time,
'policy_save': policy_save_time,
}
timing_dict.update(timing_info)
if self._should_write_summaries:
for k, v in timing_dict.items():
self._timing_sw.scalar('timing/%s' % k, v, step=last_epoch)
max_key_len = max(len(k) for k in timing_dict)
timing_info_list = [
'%s : % 10.2f' % (k.rjust(max_key_len + 1), v)
for k, v in sorted(timing_dict.items())
]
logging.info('PPO epoch [% 6d], Timings: \n%s', last_epoch,
'\n'.join(timing_info_list))
# Flush summary writers once in a while.
if self._epoch % 1000 == 0:
self.flush_summaries()
