def f(log_probs, advantages, old_log_probs, mask):
if reweight: # Use new policy weights for sampled actions instead.
mask *= jnp.exp(math.stop_gradient(log_probs) - old_log_probs)
weights = jnp.minimum(awr_weights(advantages, beta), w_max)
return -jnp.sum(log_probs * weights * mask) / jnp.sum(mask)
def _WeightedMean(inputs, **unused_kwargs): """Returns a layer to compute weighted mean over all values in the input.""" values, weights = inputs return np.sum(values * weights) / np.sum(weights)
def MaskedL2Loss(inputs, **unused_kwargs): y_hat, y, mask = inputs l2 = mask * (y_hat - y)**2 return np.sum(l2) / np.sum(mask)
def _L2(inputs, **unused_kwargs): """Returns a layer to compute L2 norms of predicted minus target vectors.""" y_hat, y = inputs return np.sum((y_hat - y)**2, axis=-1)
def _CrossEntropy(inputs, **unused_kwargs):
"""Returns a layer to compute prediction-target cross entropies."""
y_hat, target_category = inputs
return -1.0 * np.sum(y_hat * one_hot(target_category, y_hat.shape[-1]),
axis=-1)
def f(log_probs, advantages, old_log_probs, mask): del old_log_probs # Not used in AWR. weights = jnp.minimum(awr_weights(advantages, beta), w_max) return -jnp.sum(log_probs * weights * mask) / jnp.sum(mask)
def L2Loss(inputs):
y_hat, y, mask = inputs
shapes.assert_same_shape(y_hat, y)
shapes.assert_same_shape(y, mask)
l2 = mask * (y_hat - y)**2
return np.sum(l2) / np.sum(mask)
def Sum(x, axis=-1, keepdims=False):
return np.sum(x, axis=axis, keepdims=keepdims)
def CrossEntropy(x, axis=-1, **kw):
del kw
prediction, target = x
return np.sum(prediction * core.one_hot(target, prediction.shape[-1]),
axis=axis)
def Sum(axis=-1, keepdims=False):
return Fn('Sum', lambda x: jnp.sum(x, axis=axis, keepdims=keepdims))
def entropy(self, log_probs): probs = jnp.exp(log_probs) return -jnp.sum(probs * log_probs, axis=-1)
def Sum(x, axis=-1, keepdims=False, **unused_kwargs):
return np.sum(x, axis=axis, keepdims=keepdims)
def forward_unbatched(self, x, *, weights, state, rng, update_state):
attend_rng, output_rng = jax.random.split(rng)
w_q, w_v, w_o = weights
q = np.matmul(x, w_q)
v = np.matmul(x, w_v)
if update_state:
_, old_hash_rng = state
hash_rng, hash_subrng = jax.random.split(old_hash_rng)
buckets = self.hash_vectors(q, hash_subrng)
state = (buckets, hash_rng)
else:
buckets, _ = state
seqlen = x.shape[0]
assert int(buckets.shape[0]) == self.n_hashes * seqlen
ticker = jax.lax.tie_in(x, np.arange(self.n_hashes * seqlen))
buckets_and_t = seqlen * buckets + (ticker % seqlen)
buckets_and_t = jax.lax.stop_gradient(buckets_and_t)
# Hash-based sort ("s" at the start of variable names means "sorted")
sbuckets_and_t, sticker = jax.lax.sort_key_val(
buckets_and_t, ticker, dimension=-1)
_, undo_sort = jax.lax.sort_key_val(sticker, ticker, dimension=-1)
sbuckets_and_t = jax.lax.stop_gradient(sbuckets_and_t)
sticker = jax.lax.stop_gradient(sticker)
undo_sort = jax.lax.stop_gradient(undo_sort)
st = (sticker % seqlen)
sq = np.take(q, st, axis=0)
sv = np.take(v, st, axis=0)
mask_fn = functools.partial(
mask_self_attention, causal=self.causal, exclude_self=True)
q_info = st
so, slogits = attend(
sq, k=None, v=sv,
q_chunk_len=self.chunk_len,
n_chunks_before=self.n_chunks_before,
n_chunks_after=self.n_chunks_after,
mask_fn=mask_fn, q_info=q_info,
dropout=self.attention_dropout, rng=attend_rng,
)
# np.take(so, undo_sort, axis=0); np.take(slogits, undo_sort, axis=0) would
# also work, but these helpers include performance optimizations for TPU.
o = permute_via_gather(so, undo_sort, sticker, axis=0)
logits = permute_via_sort(slogits, sticker, buckets_and_t, axis=-1)
if self.n_hashes > 1:
o = np.reshape(o, (self.n_hashes, seqlen, o.shape[-1]))
logits = np.reshape(logits, (self.n_hashes, seqlen, 1))
probs = np.exp(logits - logsumexp(logits, axis=0, keepdims=True))
o = np.sum(o * probs, axis=0)
assert o.shape == (seqlen, w_v.shape[-1])
out = np.matmul(o, w_o)
out = apply_broadcasted_dropout(out, self.output_dropout, output_rng)
return out, state
def WeightedSum(inputs, **unused_kwargs): """Returns a layer to compute weighted sum over all values in the input.""" values, weights = inputs return np.sum(values * weights)
def train_epoch(self, evaluate=True):
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.evaluate()
policy_eval_time = policy_based_utils.get_time(policy_eval_start_time)
def write_metric(key, value):
self._train_sw.scalar(key, value, step=self.epoch)
self._history.append('train', key, self.epoch, value)
# Get fresh trajectories every time.
self._should_reset_train_env = True
trajectory_collection_start_time = time.time()
logging.vlog(1, 'AWR epoch [% 6d]: collecting trajectories.',
self._epoch)
trajs, _, timing_info, self._model_state = self.collect_trajectories(
train=True, temperature=1.0, raw_trajectory=True)
del timing_info
trajectory_collection_time = policy_based_utils.get_time(
trajectory_collection_start_time)
logging.vlog(1, 'AWR epoch [% 6d]: n_trajectories [%s].', self._epoch,
len(trajs))
# Convert these into numpy now.
def extract_obs_act_rew_dones(traj_np):
return traj_np[0], traj_np[1], traj_np[2], traj_np[4]
trajs_np = [extract_obs_act_rew_dones(traj.as_numpy) for traj in trajs]
# number of new actions.
new_sample_count = sum(traj[1].shape[0] for traj in trajs_np)
self._n_observations_seen += new_sample_count
logging.vlog(1, 'AWR epoch [% 6d]: new_sample_count [%d].',
self._epoch, new_sample_count)
if self._should_write_summaries:
write_metric('trajs/batch', len(trajs))
write_metric('trajs/new_sample_count', new_sample_count)
# The number of trajectories, i.e. `B`can keep changing from iteration to
# iteration, since we are capped on the number of observations requested.
# So let's operate on each trajectory on this own?
# TODO(afrozm): So should our batches look like (B, T+1, *OBS) or B
# different examples of (T+1, *OBS) each. Since B can keep changing?
# Add these to the replay buffer.
for traj in trajs:
_ = self._replay_buffer.store(traj)
rewards = np.array([np.sum(traj[2]) for traj in trajs_np])
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 and self._should_write_summaries:
metrics = {'raw': {1.0: {'mean': avg_reward, 'std': std_reward}}}
policy_based_utils.write_eval_reward_summaries(
metrics, self._log, self.epoch)
logging.vlog(
1, 'AWR epoch [% 6d]: Rewards avg=[%0.2f], max=[%0.2f], '
'min=[%0.2f].', self.epoch, avg_reward, max_reward, min_reward)
if self._should_write_summaries:
write_metric('reward/avg', avg_reward)
write_metric('reward/std', std_reward)
write_metric('reward/max', max_reward)
write_metric('reward/min', min_reward)
# Wrap these observations/rewards inside ReplayBuffer.
idx, valid_mask, valid_idx = self._replay_buffer.get_valid_indices()
# pylint: disable=g-complex-comprehension
observations = [
self._replay_buffer.get(
replay_buffer.ReplayBuffer.OBSERVATIONS_KEY,
idx[start_idx:end_plus_1_idx])
for (start_idx,
end_plus_1_idx) in self._replay_buffer.iterate_over_paths(idx)
]
rewards = [
self._replay_buffer.get(replay_buffer.ReplayBuffer.REWARDS_KEY,
idx[start_idx:end_plus_1_idx][:-1])
for (start_idx,
end_plus_1_idx) in self._replay_buffer.iterate_over_paths(idx)
]
# pylint: enable=g-complex-comprehension
t_final = awr_utils.padding_length(rewards, boundary=self._boundary)
logging.vlog(1, 'AWR epoch [% 6d]: t_final [%s].', self._epoch,
t_final)
if self._should_write_summaries:
write_metric('trajs/t_final', t_final)
# These padded observations are over *all* the non-final observations in
# the entire replay buffer.
# Shapes:
# padded_observations = (B, T + 1, *OBS)
# padded_observations_mask = (B, T + 1)
padded_observations, padded_observations_mask = (
awr_utils.pad_array_to_length(observations, t_final + 1))
batch = len(observations)
self._check_shapes('padded_observations',
'(batch, t_final + 1)',
padded_observations, (batch, t_final + 1),
array_prefix=2)
self._check_shapes('padded_observations_mask', '(batch, t_final + 1)',
padded_observations_mask, (batch, t_final + 1))
# Shapes:
# padded_rewards = (B, T)
# padded_rewards_mask = (B, T)
padded_rewards, padded_rewards_mask = awr_utils.pad_array_to_length(
rewards, t_final)
self._check_shapes('padded_rewards', '(batch, t_final)',
padded_rewards, (batch, t_final))
self._check_shapes('padded_rewards_mask', '(batch, t_final)',
padded_rewards_mask, (batch, t_final))
# Shapes:
# lengths = (B,)
lengths = np.sum(padded_rewards_mask, axis=1, dtype=np.int32)
self._check_shapes('lengths', '(batch,)', lengths, (batch, ))
# TODO(pkozakowski): Pass the actual actions here, to enable autoregressive
# action sampling.
dummy_actions = np.zeros(
(batch, t_final + 1) + self._action_shape,
self._action_dtype,
)
# Shapes:
# log_probabs_traj = (B, T + 1, #controls, #actions)
# value_predictions_traj = (B, T + 1)
log_probabs_traj, value_predictions_traj, self._model_state, unused_rng = (
self._policy_fun_all_timesteps(padded_observations, lengths,
self._model_state, self._get_rng()))
self._check_shapes(
'log_probabs_traj', '(batch, t_final + 1, n_controls, n_actions)',
log_probabs_traj,
(batch, t_final + 1, self._n_controls, self._n_actions))
self._check_shapes('value_predictions_traj', '(batch, t_final + 1)',
value_predictions_traj, (batch, t_final + 1))
# Zero out the padding's value predictions, since the net may give some
# prediction to the padding observations.
value_predictions_traj *= padded_observations_mask
# Compute td-lambda returns, and reshape to match value_predictions_traj.
list_td_lambda_returns = awr_utils.batched_compute_td_lambda_return(
padded_rewards, padded_rewards_mask, value_predictions_traj,
padded_observations_mask, self._gamma, self._td_lambda)
if logging.vlog_is_on(1) and list_td_lambda_returns:
l = len(list_td_lambda_returns)
logging.vlog(1, f'Len of list_td_lambda_returns: {l}.')
self._log_shape('td_lambda_returns[0]', list_td_lambda_returns[0])
# pad an extra 0 for each to match lengths of value predictions.
list_target_values = [
onp.pad(l, (0, 1), 'constant') for l in list_td_lambda_returns
]
if batch != len(list_target_values):
raise ValueError(f'batch != len(list_target_values) : '
f'{batch} vs {len(list_target_values)}')
# Shape: (len(idx),)
target_values = onp.concatenate(list_target_values)
self._check_shapes('target_values', '(len(idx),)', target_values,
(len(idx), ))
# Shape: (len(idx),)
vals = self.flatten_vals(value_predictions_traj,
padded_observations_mask)
self._check_shapes('vals', '(len(idx),)', vals, (len(idx), ))
# Calculate advantages.
adv, norm_adv, adv_mean, adv_std = self._calc_adv(
target_values, vals, valid_mask)
self._check_shapes('norm_adv', '(len(idx),)', norm_adv, (len(idx), ))
adv_weights, adv_weights_mean, adv_weights_min, adv_weights_max = (
self._calc_adv_weights(norm_adv, valid_mask))
self._check_shapes('adv_weights', '(len(idx),)', adv_weights,
(len(idx), ))
del adv, adv_mean, adv_std
del adv_weights_min, adv_weights_max, adv_weights_mean
combined_steps = int(
np.ceil(self._optimization_steps * new_sample_count /
self._num_samples_to_collect))
optimization_start_time = time.time()
combined_losses = self._update_combined(combined_steps, valid_idx,
target_values, adv_weights)
optimization_time = policy_based_utils.get_time(
optimization_start_time)
self._epoch += 1
if self._should_write_summaries:
write_metric('combined/optimization_steps', combined_steps)
epoch_time = policy_based_utils.get_time(epoch_start_time)
timing_dict = {
'epoch': epoch_time,
'trajectory_collection': trajectory_collection_time,
'optimization': optimization_time,
'policy_eval': policy_eval_time,
}
if self._should_write_summaries:
for k, v in timing_dict.items():
write_metric('timing/{}'.format(k), v)
# Only dump the average post losses.
if combined_losses:
for k, v in combined_losses.items():
if 'post_entropy' in k:
write_metric(k.replace('post_entropy', 'entropy'), v)
if 'post_loss' in k:
write_metric(k.replace('post_loss', 'loss'), v)
self.flush_summaries()
def L2(x, axis=-1, **kw):
del kw
prediction, target = x
return np.sum((prediction - target)**2, axis=axis)
def flatten_vals(self, value_predictions_traj, padded_observations_mask):
batch = len(padded_observations_mask)
lens = np.sum(padded_observations_mask, axis=1)
return np.concatenate(
[value_predictions_traj[b][:int(lens[b])] for b in range(batch)])
def WeightedMean(x, **kw):
del kw
metric, weights = x
weights_sum = np.sum(weights)
return np.sum(metric * weights) / weights_sum
def train_epoch(self, evaluate=True):
def write_metric(key, value):
self._train_sw.scalar(key, value, step=self.epoch)
self._history.append('train', key, self.epoch, value)
# Get fresh trajectories every time.
self._should_reset_train_env = True
trajectory_collection_start_time = time.time()
logging.vlog(1, 'AWR epoch [% 6d]: collecting trajectories.', self._epoch)
trajs, _, timing_info, self._model_state = self.collect_trajectories(
train=True, temperature=1.0, raw_trajectory=True)
del timing_info
trajectory_collection_time = ppo.get_time(trajectory_collection_start_time)
# Convert these into numpy now.
def extract_obs_act_rew_dones(traj_np):
return traj_np[0], traj_np[1], traj_np[2], traj_np[4]
trajs_np = [extract_obs_act_rew_dones(traj.as_numpy) for traj in trajs]
# number of new actions.
new_sample_count = sum(traj[1].shape[0] for traj in trajs_np)
if self._should_write_summaries:
write_metric('trajs/batch', len(trajs))
write_metric('trajs/new_sample_count', new_sample_count)
# The number of trajectories, i.e. `B`can keep changing from iteration to
# iteration, since we are capped on the number of observations requested.
# So let's operate on each trajectory on this own?
# TODO(afrozm): So should our batches look like (B, T+1, *OBS) or B
# different examples of (T+1, *OBS) each. Since B can keep changing?
# Add these to the replay buffer.
for traj in trajs:
_ = self._replay_buffer.store(traj)
if self._should_write_summaries:
rewards = np.array([np.sum(traj[2]) for traj in trajs_np])
avg_reward = np.mean(rewards)
std_reward = np.std(rewards)
max_reward = np.max(rewards)
min_reward = np.min(rewards)
write_metric('reward/avg', avg_reward)
write_metric('reward/std', std_reward)
write_metric('reward/max', max_reward)
write_metric('reward/min', min_reward)
# Wrap these observations/rewards inside ReplayBuffer.
idx, valid_mask, valid_idx = self._replay_buffer.get_valid_indices()
# pylint: disable=g-complex-comprehension
observations = [
self._replay_buffer.get(replay_buffer.ReplayBuffer.OBSERVATIONS_KEY,
idx[start_idx:end_plus_1_idx])
for (start_idx,
end_plus_1_idx) in self._replay_buffer.iterate_over_paths(idx)
]
rewards = [
self._replay_buffer.get(replay_buffer.ReplayBuffer.REWARDS_KEY,
idx[start_idx:end_plus_1_idx][:-1])
for (start_idx,
end_plus_1_idx) in self._replay_buffer.iterate_over_paths(idx)
]
# pylint: enable=g-complex-comprehension
t_final = awr_utils.padding_length(rewards, boundary=self._boundary)
if self._should_write_summaries:
write_metric('trajs/t_final', t_final)
# These padded observations are over *all* the non-final observations in
# the entire replay buffer.
# Shapes:
# padded_observations = (B, T + 1, *OBS)
# padded_observations_mask = (B, T + 1)
padded_observations, padded_observations_mask = (
awr_utils.pad_array_to_length(observations, t_final + 1)
)
batch = len(observations)
if ((batch, t_final + 1) != padded_observations.shape[:2] or
(batch, t_final + 1) != padded_observations_mask.shape):
raise ValueError(
f'Shapes mismatch, batch {batch}, t_final {t_final}'
f'padded_observations.shape {padded_observations.shape}'
f'padded_observations_mask.shape {padded_observations_mask.shape}')
# Shapes:
# padded_rewards = (B, T)
# padded_rewards_mask = (B, T)
padded_rewards, padded_rewards_mask = awr_utils.pad_array_to_length(
rewards, t_final)
if ((padded_rewards.shape != (batch, t_final)) or
(padded_rewards_mask.shape != (batch, t_final))):
raise ValueError(
f'Shapes mismatch, batch {batch}, t_final {t_final}'
f'padded_rewards.shape {padded_rewards.shape}')
# Shapes:
# log_probabs_traj = (B, T + 1, #actions)
# value_predictions_traj = (B, T + 1)
(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=self._get_rng(),
))
if ((batch, t_final + 1) != log_probabs_traj.shape[:2] or
(batch, t_final + 1) != value_predictions_traj.shape):
raise ValueError(
f'Shapes mismatch, batch {batch}, t_final {t_final}'
f'log_probabs_traj.shape {log_probabs_traj.shape}'
f'value_predictions_traj.shape {value_predictions_traj.shape}')
# Zero out the padding's value predictions, since the net may give some
# prediction to the padding observations.
value_predictions_traj *= padded_observations_mask
# Compute td-lambda returns, and reshape to match value_predictions_traj.
list_td_lambda_returns = awr_utils.batched_compute_td_lambda_return(
padded_rewards, padded_rewards_mask, value_predictions_traj,
padded_observations_mask, self._gamma, self._td_lambda)
# pad an extra 0 for each to match lengths of value predictions.
list_target_values = [
onp.pad(l, (0, 1), 'constant') for l in list_td_lambda_returns
]
if batch != len(list_target_values):
raise ValueError(f'batch != len(list_target_values) : '
f'{batch} vs {len(list_target_values)}')
# Shape: (len(idx),)
target_values = onp.concatenate(list_target_values)
if target_values.shape != (len(idx),):
raise ValueError(f'target_values.shape != (len(idx),) = '
f'{target_values.shape} != ({len(idx)},)')
# Shape: (len(idx),)
target_values = onp.concatenate(list_target_values)
vals = self.flatten_vals(value_predictions_traj, padded_observations_mask)
if vals.shape != target_values.shape:
raise ValueError(f'vals.shape != target_values.shape : '
f'{vals.shape} vs {target_values.shape}')
# Calculate advantages.
adv, norm_adv, adv_mean, adv_std = self._calc_adv(
target_values, vals, valid_mask)
adv_weights, adv_weights_mean, adv_weights_min, adv_weights_max = (
self._calc_adv_weights(norm_adv, valid_mask)
)
del adv, adv_mean, adv_std
del adv_weights_min, adv_weights_max, adv_weights_mean
combined_steps = int(
np.ceil(self._optimization_steps * new_sample_count /
self._num_samples_to_collect))
combined_losses = self._update_combined(combined_steps, valid_idx,
target_values, adv_weights)
if self._should_write_summaries:
write_metric('combined/optimization_steps', combined_steps)
timing_dict = {
'trajectory_collection': trajectory_collection_time,
# 'epoch': epoch_time,
# 'policy_eval': policy_eval_time,
# 'preprocessing': preprocessing_time,
# 'log_prob_recompute': log_prob_recompute_time,
# 'loss_compute': loss_compute_time,
# 'optimization': optimization_time,
# 'policy_save': policy_save_time,
}
if self._should_write_summaries:
for k, v in timing_dict.items():
write_metric('timing/{}'.format(k), v)
# Only dump the average post losses.
if combined_losses:
for k, v in combined_losses.items():
if 'post_entropy' in k:
write_metric(k.replace('post_entropy', 'entropy'), v)
if 'post_loss' in k:
write_metric(k.replace('post_loss', 'loss'), v)
self._epoch += 1
self.flush_summaries()
def forward_unbatched(self, x, *, weights, state, update_state):
w_q, w_v, w_o = weights
q = np.matmul(x, w_q)
v = np.matmul(x, w_v)
if update_state:
_, old_rng = state
rng = jax.random.fold_in(old_rng, 0)
hash_rng = jax.random.fold_in(rng, 1)
buckets = self.hash_vectors(q, hash_rng)
state = (buckets, rng)
else:
buckets, rng = state
rng = jax.random.fold_in(rng, 2)
seqlen = x.shape[0]
assert int(buckets.shape[0]) == self.n_hashes * seqlen
ticker = jax.lax.tie_in(x, np.arange(self.n_hashes * seqlen))
buckets_and_t = seqlen * buckets + (ticker % seqlen)
buckets_and_t = jax.lax.stop_gradient(buckets_and_t)
# Hash-based sort ("s" at the start of variable names means "sorted")
sbuckets_and_t, sticker = jax.lax.sort_key_val(buckets_and_t,
ticker,
dimension=-1)
_, undo_sort = jax.lax.sort_key_val(sticker, ticker, dimension=-1)
sbuckets_and_t = jax.lax.stop_gradient(sbuckets_and_t)
sticker = jax.lax.stop_gradient(sticker)
undo_sort = jax.lax.stop_gradient(undo_sort)
st = (sticker % seqlen)
sq = np.take(q, st, axis=0)
sv = np.take(v, st, axis=0)
mask_fn = functools.partial(mask_self_attention,
causal=self.causal,
exclude_self=True)
q_info = st
so, slogits = attend(
sq,
k=None,
v=sv,
q_chunk_len=self.chunk_len,
n_chunks_before=self.n_chunks_before,
n_chunks_after=self.n_chunks_after,
mask_fn=mask_fn,
q_info=q_info,
dropout=self.attention_dropout,
rng=rng,
)
def unsort_for_output_impl(so, slogits):
o = np.take(so, undo_sort, axis=0)
# Sorting is considerably faster than gather, but first we need to get the
# XLA compiler to abandon the idea of fusing this sort with the input sort
# (which introduces a computation cycle and leads to a crash).
# TODO(kitaev): remove "sticker_" variable if XLA is fixed.
sticker_ = sticker + jax.lax.convert_element_type(
slogits[0] > 0, sticker.dtype)
_, logits = jax.lax.sort_key_val(sticker_, slogits, dimension=-1)
return o, logits
def unsort_for_output_vjp(so, slogits):
"""Custom gradient for unsort_for_output."""
so = jax.lax.stop_gradient(so)
slogits = jax.lax.stop_gradient(slogits)
o, logits = unsort_for_output_impl(so, slogits)
def vjpfun(o_logits_grads):
so_grad = np.take(o_logits_grads[0], sticker, axis=0)
# TODO(kitaev): this exists to match the forward pass, but I'm not sure
# if it's actually required.
buckets_and_t_ = buckets_and_t + jax.lax.convert_element_type(
o_logits_grads[1][0] > 0, buckets_and_t.dtype)
_, slogits_grad = jax.lax.sort_key_val(buckets_and_t_,
o_logits_grads[1],
dimension=-1)
return (so_grad, slogits_grad)
return (o, logits), vjpfun
unsort_for_output = jax.custom_transforms(unsort_for_output_impl)
jax.defvjp_all(unsort_for_output, unsort_for_output_vjp)
o, logits = unsort_for_output_impl(so, slogits)
if self.n_hashes > 1:
o = np.reshape(o, (self.n_hashes, seqlen, o.shape[-1]))
logits = np.reshape(logits, (self.n_hashes, seqlen, 1))
probs = np.exp(logits - logsumexp(logits, axis=0, keepdims=True))
o = np.sum(o * probs, axis=0)
assert o.shape == (seqlen, w_v.shape[-1])
out = np.matmul(o, w_o)
return out, state
