class DDPG:
def __init__(self,
env=gym.make('Pendulum-v0'),
s_dim=2,
a_dim=1,
gamma=0.99,
episodes=100,
tau=0.001,
buffer_size=1e06,
minibatch_size=64,
actor_lr=0.001,
critic_lr=0.001,
save_name='final_weights',
render=False):
self.save_name = save_name
self.render = render
self.env = env
self.upper_bound = env.action_space.high[0]
self.lower_bound = env.action_space.low[0]
self.EPISODES = episodes
self.MAX_TIME_STEPS = 200
self.s_dim = s_dim
self.a_dim = a_dim
self.GAMMA = gamma
self.TAU = tau
self.buffer_size = buffer_size
self.minibatch_size = minibatch_size
self.actor_lr = actor_lr
self.critic_lr = critic_lr
self.ou_noise = OUNoise(mean=np.zeros(1))
self.actor = Actor(self.s_dim, self.a_dim).model()
self.target_actor = Actor(self.s_dim, self.a_dim).model()
self.actor_opt = tf.keras.optimizers.Adam(learning_rate=self.actor_lr)
self.target_actor.set_weights(self.actor.get_weights())
self.critic = Critic(self.s_dim, self.a_dim).model()
self.critic_opt = tf.keras.optimizers.Adam(
learning_rate=self.critic_lr)
self.target_critic = Critic(self.s_dim, self.a_dim).model()
self.target_critic.set_weights(self.critic.get_weights())
self.replay_buffer = ReplayBuffer(self.buffer_size)
def update_target(self):
# Two methods to update the target actor
# Method 1:
self.target_actor.set_weights(
np.array(self.actor.get_weights()) * self.TAU +
np.array(self.target_actor.get_weights()) * (1 - self.TAU))
self.target_critic.set_weights(
np.array(self.critic.get_weights()) * self.TAU +
np.array(self.target_critic.get_weights()) * (1 - self.TAU))
"""
# Method 2:
new_weights = []
target_variables = self.target_critic.weights
for i, variable in enumerate(self.critic.weights):
new_weights.append(variable * self.TAU + target_variables[i] * (1 - self.TAU))
self.target_critic.set_weights(new_weights)
new_weights = []
target_variables = self.target_actor.weights
for i, variable in enumerate(self.actor.weights):
new_weights.append(variable * self.TAU + target_variables[i] * (1 - self.TAU))
self.target_actor.set_weights(new_weights)
"""
def train_step(self):
s_batch, a_batch, r_batch, d_batch, s2_batch = self.replay_buffer.sample_batch(
self.minibatch_size)
"""
mu_prime = self.target_actor(s2_batch) # predictions by target actor
Q_prime = self.target_critic([s2_batch, mu_prime]) # predictions by target critic
y = np.zeros_like(Q_prime)
for k in range(self.minibatch_size):
if d_batch[k]:
y[k] = r_batch[k]
else:
y[k] = r_batch[k] + self.GAMMA * Q_prime[k]
# y = r_batch + gamma * Q_prime
checkpoint_path = "training/cp_critic.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)
# Create a callback that saves the model's weights
cp_callback1 = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_dir,
save_weights_only=True,
verbose=1)
self.critic.train_on_batch([s_batch, a_batch], y)
# self.critic.fit([s_batch, a_batch], y, verbose=0, steps_per_epoch=8, callbacks=[cp_callback1])
with tf.GradientTape(persistent=True) as tape:
a = self.actor(s_batch)
tape.watch(a)
theta = self.actor.trainable_variables
q = self.critic([s_batch, a])
dq_da = tape.gradient(q, a)
da_dtheta = tape.gradient(a, theta, output_gradients=-dq_da)
self.actor_opt.apply_gradients(zip(da_dtheta, self.actor.trainable_variables))
"""
with tf.GradientTape() as tape:
target_actions = self.target_actor(s2_batch)
y = r_batch + self.GAMMA * self.target_critic(
[s2_batch, target_actions])
critic_value = self.critic([s_batch, a_batch])
critic_loss = tf.math.reduce_mean(tf.math.square(y - critic_value))
critic_grad = tape.gradient(critic_loss,
self.critic.trainable_variables)
self.critic_opt.apply_gradients(
zip(critic_grad, self.critic.trainable_variables))
with tf.GradientTape() as tape:
actions = self.actor(s_batch)
q = self.critic([s_batch, actions]) # critic_value
# Used `-value` as we want to maximize the value given
# by the critic for our actions
actor_loss = -tf.math.reduce_mean(q)
actor_grad = tape.gradient(actor_loss, self.actor.trainable_variables)
self.actor_opt.apply_gradients(
zip(actor_grad, self.actor.trainable_variables))
self.update_target()
return np.mean(q)
def policy(self, s):
# since batch normalization is done on self.actor, it is multiplied with upper_bound
if s.ndim == 1:
s = s[None, :]
action = self.actor(s) * self.upper_bound + self.ou_noise()
action = np.clip(action, self.lower_bound, self.upper_bound)
return action
def train(self):
# To store reward history of each episode
ep_reward_list = []
# To store average reward history of last few episodes
avg_reward_list = []
monitor = Monitor([1, 1], titles=['Reward', 'Loss'], log=2)
with Loop_handler(
) as interruption: # to properly save even if ctrl+C is pressed
for eps in range(self.EPISODES):
episode_reward = 0
s = self.env.reset()
"""
if an env is created using the "gym.make" method, it will terminate after 200 steps
"""
for t in range(self.MAX_TIME_STEPS):
# done = False
# while not done:
if self.render:
self.env.render()
a = self.policy(s)
s_, r, done, _ = self.env.step(a)
self.replay_buffer.add(np.reshape(s, (self.s_dim, )),
np.reshape(a, (self.a_dim, )),
r, done,
np.reshape(s_, (self.s_dim, )))
episode_reward += r
if self.replay_buffer.size() > self.minibatch_size:
q = self.train_step()
s = s_.reshape(1, -1)
if interruption():
break
ep_reward_list.append(episode_reward)
# Mean of last 40 episodes
avg_reward = np.mean(ep_reward_list[-40:])
print("Episode * {} * Avg Reward is ==> {}".format(
eps, avg_reward))
avg_reward_list.append(avg_reward)
monitor.add_data(avg_reward, q)
self.save_weights(
save_name=self.save_name) # if you want to save weights
self.plot_results(avg_reward=avg_reward_list, train=True)
def save_weights(self, save_name='final_weights'):
self.actor.save_weights("training/%s_actor.h5" % save_name)
self.critic.save_weights("training/%s_critic.h5" % save_name)
self.target_actor.save_weights("training/%s_target_actor.h5" %
save_name)
self.target_critic.save_weights("training/%s_target_critic.h5" %
save_name)
# to save in other format
self.target_actor.save_weights('training/%s_actor_weights' % save_name,
save_format='tf')
self.target_critic.save_weights('training/%s_critic_weights' %
save_name,
save_format='tf')
print('Training completed and network weights saved')
# For evaluation of the policy learned
def collect_data(self, act_net, iterations=1000):
a_all, states_all = [], []
obs = self.env.reset()
for t in range(iterations):
obs = np.squeeze(obs)
if obs.ndim == 1:
a = act_net(obs[None, :])
else:
a = act_net(obs)
obs, _, done, _ = self.env.step(a)
states_all.append(obs)
a_all.append(a)
# self.env.render() # Uncomment this to see the actor in action (But not in python notebook)
# if done:
# break
states = np.squeeze(
np.array(states_all)) # cos(theta), sin(theta), theta_dot
a_all = np.squeeze(np.array(a_all))
return states, a_all
def plot_results(self,
avg_reward=None,
actions=None,
states=None,
train=False,
title=None):
# An additional way to visualize the avg episode rewards
if train:
plt.figure()
plt.plot(avg_reward)
plt.xlabel("Episode")
plt.ylabel("Avg. Epsiodic Reward")
plt.show()
else: # work only for Pendulum-v0 environment
fig, ax = plt.subplots(3, sharex=True)
theta = np.arctan2(states[:, 1], states[:, 0])
ax[0].set_ylabel('u')
ax[0].plot(np.squeeze(actions))
ax[1].set_ylabel(u'$\\theta$')
ax[1].plot(theta)
# ax[1].plot(states[:, 0])
ax[2].set_ylabel(u'$\omega$')
ax[2].plot(states[:, 2]) # ang velocity
fig.canvas.set_window_title(title)
def ddpg(env_fn, ac_kwargs=dict(), seed=0, steps_per_epoch=5000, epochs=100,
replay_size=int(1e6), discount=0.99, polyak=0.995, pi_lr=1e-3, q_lr=1e-3,
batch_size=100, start_steps=10000, act_noise=0.1, max_ep_len=1000,
logger_kwargs=dict(), save_freq=1):
"""
Implements the deep deterministic policy gradient algorithm.
Performance statistics are logged to stdout and to file in
CSV format, and models are saved regularly during training.
Args:
env_fn: callable. Must load an instance of an environment
that implements the OpenAI Gym API.
ac_kwargs: dict. Additional keyword arguments to be passed
to the Actor and Critic constructors.
seed: int. Random seed.
steps_per_epoch: int. Number of training steps or
environment interactions that make up one epoch.
epochs: int. Number of epochs for training.
replay_size: int. Maximum number of transitions that
can be stored in the replay buffer.
discount: float. Rate of discounting on future reward,
usually denoted with the Greek letter gamma. Normally
between 0 and 1.
polyak: float. Weighting of target estimator parameters
in the target update (which is a "polayk" average).
pi_lr: float. Learning rate for the policy or actor estimator.
q_lr: float. Learning rate for the Q or critic estimator.
batch_size: int. Number of transitions to sample from the
replay buffer per gradient update of the estimators.
start_steps: int. Number of initial training steps where
actions are chosen at random instead of the policy,
as a means of increasing exploration.
act_noise: float. Scale (standard deviation) of the Gaussian
noise added to the policy for exploration during training.
max_ep_len: int. Maximum number of steps for one episode in
the environment. Episode length may be shorter if there
are terminal states.
logger_kwargs: dict. Keyword arguments to be passed to the
logger. Can be set up using utils.setup_logger_kwargs().
save_freq: int. Models are saved per this number of epochs.
"""
# Set up logging
logger = EpochLogger(**logger_kwargs)
logger.save_config(locals())
# Set random seed for relevant modules
tf.random.set_seed(seed)
np.random.seed(seed)
# Create environment
env, test_env = env_fn(), env_fn()
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
if env._max_episode_steps < max_ep_len:
max_ep_len = env._max_episode_steps
if steps_per_epoch % max_ep_len != 0:
"""
Training steps are batched at the end of a trajectory, so if
episode length does not divide steps per epoch, the size of
training step log arrays can be inconsistent. This takes the
upper bound on size, which wastes some memory but is easy.
"""
max_logger_steps = steps_per_epoch + max_ep_len - (steps_per_epoch % max_ep_len)
else:
max_logger_steps = steps_per_epoch
# Action limit for clipping
# Assumes all dimensions have the same limit
act_limit = env.action_space.high[0]
# Give actor-critic model access to action space
ac_kwargs['action_space'] = env.action_space
# Randomly initialise critic and actor networks
critic = Critic(input_shape=(batch_size, obs_dim + act_dim), lr=q_lr, **ac_kwargs)
actor = Actor(input_shape=(batch_size, obs_dim), lr=pi_lr, **ac_kwargs)
# Initialise target networks with the same weights as main networks
critic_target = Critic(input_shape=(batch_size, obs_dim + act_dim), **ac_kwargs)
actor_target = Actor(input_shape=(batch_size, obs_dim), **ac_kwargs)
critic_target.set_weights(critic.get_weights())
actor_target.set_weights(actor.get_weights())
# Initialise replay buffer for storing and getting batches of transitions
replay_buffer = ReplayBuffer(obs_dim, act_dim, size=replay_size)
# Set up model checkpointing so we can resume training or test separately
checkpoint_dir = os.path.join(logger.output_dir, 'training_checkpoints')
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt")
checkpoint = tf.train.Checkpoint(critic=critic, actor=actor)
def get_action(o, noise_scale):
"""
Computes an action from the policy (as a function of the
observation `o`) with added noise (scaled by `noise_scale`),
clipped within the bounds of the action space.
"""
a = actor(o.reshape(1, -1))
a += noise_scale * np.random.randn(act_dim)
return np.clip(a, -act_limit, act_limit)
@tf.function
def train_step(batch):
"""
Performs a gradient update on the actor and critic estimators
from the given batch of transitions.
Args:
batch: dict. A batch of transitions. Must store valid
values for 'obs1', 'acts', 'obs2', 'rwds', and 'done'.
Obtained from ReplayBuffer.sample_batch().
Returns:
A tuple of the Q values, critic loss, and actor loss.
"""
with tf.GradientTape(persistent=True) as tape:
# Critic loss
q = critic(batch['obs1'], batch['acts'])
q_pi_targ = critic_target(batch['obs2'], actor_target(batch['obs2']))
backup = tf.stop_gradient(batch['rwds'] + discount * (1 - batch['done']) * q_pi_targ)
q_loss = tf.reduce_mean((q - backup)**2)
# Actor loss
pi = actor(batch['obs1'])
q_pi = critic(batch['obs1'], pi)
pi_loss = -tf.reduce_mean(q_pi)
# Q learning update
critic_gradients = tape.gradient(q_loss, critic.trainable_variables)
critic.optimizer.apply_gradients(zip(critic_gradients, critic.trainable_variables))
# Policy update
actor_gradients = tape.gradient(pi_loss, actor.trainable_variables)
actor.optimizer.apply_gradients(zip(actor_gradients, actor.trainable_variables))
return q, q_loss, pi_loss
def test_agent(n=10):
"""
Evaluates the deterministic (noise-free) policy with a sample
of `n` trajectories.
"""
for _ in range(n):
o, r, d, ep_ret, ep_len = test_env.reset(), 0, False, 0, 0
while not(d or (ep_len == max_ep_len)):
# Take deterministic actions at test time (noise_scale=0)
o, r, d, _ = test_env.step(get_action(o, 0))
ep_ret += r
ep_len += 1
logger.store(n, TestEpRet=ep_ret, TestEpLen=ep_len)
start_time = time.time()
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
total_steps = steps_per_epoch * epochs
for t in range(total_steps):
"""
Start with `start_steps` number of steps with random actions,
to improve exploration. Then use the learned policy with some
noise added to keep up exploration (but less so).
"""
if t > start_steps:
a = get_action(o, act_noise)
else:
a = env.action_space.sample()
# Execute a step in the environment
o2, r, d, _ = env.step(a)
o2 = np.squeeze(o2) # bug fix for Pendulum-v0 environment, where act_dim == 1
ep_ret += r
ep_len += 1
"""
Ignore the "done" signal if it comes from hitting the time
horizon (that is, when it's an artificial terminal signal
that isn't based on the agent's state)
"""
d = False if ep_len==max_ep_len else d
# Store transition in replay buffer
replay_buffer.store(o, a, r, o2, d)
# Advance the stored state
o = o2
if d or (ep_len == max_ep_len):
"""
Perform all DDPG updates at the end of the trajectory,
in accordance with tuning done by TD3 paper authors.
"""
for _ in range(ep_len):
batch = replay_buffer.sample_batch(batch_size)
# Actor-critic update
q, q_loss, pi_loss = train_step(batch)
logger.store((max_logger_steps, batch_size), QVals=q.numpy())
logger.store(max_logger_steps, LossQ=q_loss.numpy(), LossPi=pi_loss.numpy())
# Target update
critic_target.polyak_update(critic, polyak)
actor_target.polyak_update(actor, polyak)
logger.store(max_logger_steps // max_ep_len, EpRet=ep_ret, EpLen=ep_len)
o, r, d, ep_ret, ep_len = env.reset(), 0, False, 0, 0
# Post-training for this epoch: save, test and write logs
if t > 0 and (t+1) % steps_per_epoch == 0:
epoch = (t+1) // steps_per_epoch
# Save the model
if (epoch % save_freq == 0) or (epoch == epochs - 1):
checkpoint.save(file_prefix=checkpoint_prefix)
# Test the performance of the deterministic policy
test_agent()
# Log info about the epoch
logger.log_tabular('Epoch', epoch)
logger.log_tabular('EpRet', with_min_and_max=True)
logger.log_tabular('TestEpRet', with_min_and_max=True)
logger.log_tabular('EpLen', average_only=True)
logger.log_tabular('TestEpLen', average_only=True)
logger.log_tabular('TotalEnvInteracts', t+1)
logger.log_tabular('QVals', with_min_and_max=True)
logger.log_tabular('LossPi', average_only=True)
logger.log_tabular('LossQ', average_only=True)
logger.log_tabular('Time', time.time() - start_time)
logger.dump_tabular()