"""
Proximal Policy Optimization with KL divergence (PPO-KL)
Paper link: https://arxiv.org/pdf/1707.06347.pdf
Implementation: TensorFlow2
"""
import numpy as np
from argparse import Namespace
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.learners import Learner
from xuance.tensorflow.utils import merge_distributions
[docs]
class PPOKL_Learner(Learner):
def __init__(self,
config: Namespace,
policy: Module,
callback):
super(PPOKL_Learner, self).__init__(config, policy, callback)
if ("macOS" in self.os_name) and ("arm" in self.os_name): # For macOS with Apple's M-series chips.
if self.distributed_training:
with self.policy.mirrored_strategy.scope():
self.optimizer = tk.optimizers.legacy.Adam(config.learning_rate)
else:
self.optimizer = tk.optimizers.legacy.Adam(config.learning_rate)
else:
if self.distributed_training:
with self.policy.mirrored_strategy.scope():
self.optimizer = tk.optimizers.Adam(config.learning_rate)
else:
self.optimizer = tk.optimizers.Adam(config.learning_rate)
self.vf_coef = config.vf_coef
self.ent_coef = config.ent_coef
self.target_kl = config.target_kl
self.kl_coef = config.kl_coef
self.mse_loss = tk.losses.MeanSquaredError()
self.is_continuous = self.policy.is_continuous
@tf.function
def forward_fn(self, *args):
with tf.GradientTape() as tape:
if self.is_continuous:
obs_batch, act_batch, ret_batch, adv_batch, old_mu, old_std = args
outputs, mu, std, v_pred = self.policy(obs_batch)
# calculate log prob
log_2pi = tf.math.log(2.0 * np.pi)
log_std, old_log_std = tf.math.log(std + 1e-8), tf.math.log(old_std + 1e-8)
log_prob = -0.5 * (((act_batch - mu) / (std + 1e-8)) ** 2 + 2.0 * log_std + log_2pi)
old_log_prob = -0.5 * (((act_batch - old_mu) / (old_std + 1e-8)) ** 2 + 2.0 * old_log_std + log_2pi)
log_prob_a = tf.reduce_sum(log_prob, axis=-1, keepdims=True)
old_log_prob_a = tf.reduce_sum(old_log_prob, axis=-1, keepdims=True)
# calculate entropy
entropy = tf.reduce_sum(0.5 + 0.5 * tf.math.log(2.0 * np.pi) + log_std, axis=-1, keepdims=True)
# calculate kl divergence
var1, var2 = tf.square(std), tf.square(old_std)
kl = tf.math.log(old_std / std) + (var1 + tf.square(mu - old_mu)) / (2.0 * var2) - 0.5
else:
obs_batch, act_batch, ret_batch, adv_batch, old_logits = args
outputs, logits, v_pred = self.policy(obs_batch)
# calculate log prob
log_prob = tf.nn.log_softmax(logits, axis=-1)
log_prob_a = tf.gather(log_prob, act_batch, axis=-1, batch_dims=-1)
# calculate entropy
probs = tf.exp(log_prob)
entropy = -tf.reduce_sum(probs * log_prob, axis=-1, keepdims=True)
# calculate kl divergence
old_log_prob_batch = tf.nn.log_softmax(old_logits, axis=-1) # log Q(a)
old_log_prob_a = tf.gather(old_log_prob_batch, act_batch, axis=-1, batch_dims=-1)
kl = tf.reduce_sum(probs * (log_prob - old_log_prob_batch), axis=-1, keepdims=True)
# ppo-clip core implementations
ratio = tf.math.exp(log_prob_a - old_log_prob_a)
kl = tf.reduce_mean(kl)
a_loss = -tf.reduce_mean(ratio * adv_batch) + self.kl_coef * kl
c_loss = self.mse_loss(ret_batch, v_pred)
e_loss = tf.reduce_mean(entropy)
loss = a_loss - self.ent_coef * e_loss + self.vf_coef * c_loss
gradients = tape.gradient(loss, self.policy.trainable_variables)
if self.use_grad_clip:
gradients, _ = tf.clip_by_global_norm(gradients, clip_norm=self.grad_clip_norm)
self.optimizer.apply_gradients(zip(gradients, self.policy.trainable_variables))
else:
self.optimizer.apply_gradients(zip(gradients, self.policy.trainable_variables))
return a_loss, c_loss, e_loss, kl, v_pred
@tf.function
def learn(self, *inputs):
if self.distributed_training:
a_loss, c_loss, e_loss, kl, v_pred = self.policy.mirrored_strategy.run(self.forward_fn, args=inputs)
return (self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, a_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, c_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, e_loss, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, kl, axis=None),
self.policy.mirrored_strategy.reduce(tf.distribute.ReduceOp.SUM, v_pred, axis=None))
else:
return self.forward_fn(*inputs)
[docs]
def update(self, **samples):
self.iterations += 1
obs_batch = tf.convert_to_tensor(samples["obs"], dtype=tf.float32)
ret_batch = tf.convert_to_tensor(samples["returns"], dtype=tf.float32)
adv_batch = tf.convert_to_tensor(samples['advantages'][:, None], dtype=tf.float32)
if self.is_continuous:
act_batch = tf.convert_to_tensor(samples["actions"], dtype=tf.float32)
else:
act_batch = tf.convert_to_tensor(samples["actions"][:, None], dtype=tf.int32)
old_dists = merge_distributions(samples['aux_batch']['old_dist'])
info = self.callback.on_update_start(self.iterations,
policy=self.policy, obs=obs_batch, act=act_batch,
returns=ret_batch, advantages=adv_batch, old_dists=old_dists)
if self.is_continuous:
old_mu = old_dists.mu
old_std = old_dists.std
a_loss, c_loss, e_loss, kl, v_pred = self.learn(obs_batch, act_batch, ret_batch, adv_batch, old_mu, old_std)
else:
old_logits = old_dists.logits
a_loss, c_loss, e_loss, kl, v_pred = self.learn(obs_batch, act_batch, ret_batch, adv_batch, old_logits)
if kl > self.target_kl * 1.5:
self.kl_coef = self.kl_coef * 2.
elif kl < self.target_kl * 0.5:
self.kl_coef = self.kl_coef / 2.
self.kl_coef = tf.clip_by_value(self.kl_coef, 0.1, 20)
info.update({
"actor-loss": a_loss.numpy(),
"critic-loss": c_loss.numpy(),
"entropy": e_loss.numpy(),
"kl": kl.numpy(),
"predict_value": tf.math.reduce_mean(v_pred).numpy()
})
info.update(self.callback.on_update_end(self.iterations,
policy=self.policy, info=info,
v_pred=v_pred, kl=kl, a_loss=a_loss, c_loss=c_loss, e_loss=e_loss))
return info