# -*- encoding: utf-8 -*-
import time
import cPickle
import numpy as np
import utils.logger
import tensorflow as tf
from skimage.transform import resize
from collections import deque
from utils import checkpoint_utils
from actor_learner import ONE_LIFE_GAMES
from utils.decorators import Experimental
from utils.fast_cts import CTSDensityModel
from utils.replay_memory import ReplayMemory
from policy_based_actor_learner import A3CLearner, A3CLSTMLearner
from value_based_actor_learner import ValueBasedLearner
logger = utils.logger.getLogger('intrinsic_motivation_actor_learner')
class PixelCNNDensityModel(object):
pass
class PerPixelDensityModel(object):
"""
Calculates image probability according to per-pixel counts: P(X) = ∏ p(x_ij)
Mostly here for debugging purposes as CTSDensityModel is much more expressive
"""
def __init__(self, height=42, width=42, num_bins=8, beta=0.05):
self.counts = np.zeros((width, height, num_bins))
self.height = height
self.width = width
self.beta = beta
self.num_bins = num_bins
def update(self, obs):
obs = resize(obs, (self.height, self.width), preserve_range=True)
obs = np.floor((obs*self.num_bins)).astype(np.int32)
log_prob, log_recoding_prob = self._update(obs)
return self.exploration_bonus(log_prob, log_recoding_prob)
def _update(self, obs):
log_prob = 0.0
log_recoding_prob = 0.0
for i in range(self.height):
for j in range(self.height):
self.counts[i, j, obs[i, j]] += 1
bin_count = self.counts[i, j, obs[i, j]]
pixel_mass = self.counts[i, j].sum()
log_prob += np.log(bin_count / pixel_mass)
log_recoding_prob += np.log((bin_count + 1) / (pixel_mass + 1))
return log_prob, log_recoding_prob
def exploration_bonus(self, log_prob, log_recoding_prob):
recoding_prob = np.exp(log_recoding_prob)
prob_ratio = np.exp(log_recoding_prob - log_prob)
pseudocount = (1 - recoding_prob) / np.maximum(prob_ratio - 1, 1e-10)
return self.beta / np.sqrt(pseudocount + .01)
def get_state(self):
return self.num_bins, self.height, self.width, self.beta, self.counts
def set_state(self, state):
self.num_bins, self.height, self.width, self.beta, self.counts = state
class DensityModelMixin(object):
"""
Mixin to provide initialization and synchronization methods for density models
"""
def _init_density_model(self, args):
self.density_model_update_steps = 20*args.q_target_update_steps
self.density_model_update_flags = args.density_model_update_flags
model_args = {
'height': args.cts_rescale_dim,
'width': args.cts_rescale_dim,
'num_bins': args.cts_bins,
'beta': args.cts_beta
}
if args.density_model == 'cts':
self.density_model = CTSDensityModel(**model_args)
else:
self.density_model = PerPixelDensityModel(**model_args)
def write_density_model(self):
logger.info('T{} Writing Pickled Density Model to File...'.format(self.actor_id))
raw_data = cPickle.dumps(self.density_model.get_state(), protocol=2)
with self.barrier.counter.lock, open('/tmp/density_model.pkl', 'wb') as f:
f.write(raw_data)
for i in xrange(len(self.density_model_update_flags.updated)):
self.density_model_update_flags.updated[i] = 1
def read_density_model(self):
logger.info('T{} Synchronizing Density Model...'.format(self.actor_id))
with self.barrier.counter.lock, open('/tmp/density_model.pkl', 'rb') as f:
raw_data = f.read()
self.density_model.set_state(cPickle.loads(raw_data))
class A3CDensityModelMixin(DensityModelMixin):
"""
Mixin to share _train method between A3C and A3C-LSTM models
"""
def _train(self):
""" Main actor learner loop for advantage actor critic learning. """
logger.debug("Actor {} resuming at Step {}".format(self.actor_id,
self.global_step.value()))
bonuses = deque(maxlen=100)
while (self.global_step.value() < self.max_global_steps):
# Sync local learning net with shared mem
s = self.emulator.get_initial_state()
self.reset_hidden_state()
self.local_episode += 1
episode_over = False
total_episode_reward = 0.0
episode_start_step = self.local_step
while not episode_over:
self.sync_net_with_shared_memory(self.local_network, self.learning_vars)
self.save_vars()
rewards = list()
states = list()
actions = list()
values = list()
local_step_start = self.local_step
self.set_local_lstm_state()
while self.local_step - local_step_start < self.max_local_steps and not episode_over:
# Choose next action and execute it
a, readout_v_t, readout_pi_t = self.choose_next_action(s)
new_s, reward, episode_over = self.emulator.next(a)
total_episode_reward += reward
# Update density model
current_frame = new_s[...,-1]
bonus = self.density_model.update(current_frame)
bonuses.append(bonus)
if self.is_master() and (self.local_step % 400 == 0):
bonus_array = np.array(bonuses)
logger.debug('π_a={:.4f} / V={:.4f} / Mean Bonus={:.4f} / Max Bonus={:.4f}'.format(
readout_pi_t[a.argmax()], readout_v_t, bonus_array.mean(), bonus_array.max()))
# Rescale or clip immediate reward
reward = self.rescale_reward(self.rescale_reward(reward) + bonus)
rewards.append(reward)
states.append(s)
actions.append(a)
values.append(readout_v_t)
s = new_s
self.local_step += 1
global_step, _ = self.global_step.increment()
if global_step % self.density_model_update_steps == 0:
self.write_density_model()
if self.density_model_update_flags.updated[self.actor_id] == 1:
self.read_density_model()
self.density_model_update_flags.updated[self.actor_id] = 0
next_val = self.bootstrap_value(new_s, episode_over)
advantages = self.compute_gae(rewards, values, next_val)
targets = self.compute_targets(rewards, next_val)
# Compute gradients on the local policy/V network and apply them to shared memory
entropy = self.apply_update(states, actions, targets, advantages)
elapsed_time = time.time() - self.start_time
steps_per_sec = self.global_step.value() / elapsed_time
perf = "{:.0f}".format(steps_per_sec)
logger.info("T{} / EPISODE {} / STEP {}k / REWARD {} / {} STEPS/s".format(
self.actor_id,
self.local_episode,
self.global_step.value()/1000,
total_episode_reward,
perf))
self.log_summary(total_episode_reward, np.array(values).mean(), entropy)
@Experimental
class PseudoCountA3CLearner(A3CLearner, A3CDensityModelMixin):
"""
Attempt at replicating the A3C+ model from the paper 'Unifying Count-Based Exploration and Intrinsic Motivation' (https://arxiv.org/abs/1606.01868)
"""
def __init__(self, args):
super(PseudoCountA3CLearner, self).__init__(args)
self._init_density_model(args)
def train(self):
self._train()
@Experimental
class PseudoCountA3CLSTMLearner(A3CLSTMLearner, A3CDensityModelMixin):
def __init__(self, args):
super(PseudoCountA3CLSTMLearner, self).__init__(args)
self._init_density_model(args)
def train(self):
self._train()
class PseudoCountQLearner(ValueBasedLearner, DensityModelMixin):
"""
Based on DQN+CTS model from the paper 'Unifying Count-Based Exploration and Intrinsic Motivation' (https://arxiv.org/abs/1606.01868)
Presently the implementation differs from the paper in that the novelty bonuses are computed online rather than by computing the
prediction gains after the model has been updated with all frames from the episode. Async training with different final epsilon values
tends to produce better results than just using a single actor-learner.
"""
def __init__(self, args):
self.args = args
super(PseudoCountQLearner, self).__init__(args)
self.cts_eta = args.cts_eta
self.cts_beta = args.cts_beta
self.batch_size = args.batch_update_size
self.replay_memory = ReplayMemory(
args.replay_size,
self.local_network.get_input_shape(),
self.num_actions)
self._init_density_model(args)
self._double_dqn_op()
def generate_final_epsilon(self):
if self.num_actor_learners == 1:
return self.args.final_epsilon
else:
return super(PseudoCountQLearner, self).generate_final_epsilon()
def _get_summary_vars(self):
q_vars = super(PseudoCountQLearner, self)._get_summary_vars()
bonus_q05 = tf.Variable(0., name='novelty_bonus_q05')
s1 = tf.summary.scalar('Novelty_Bonus_q05_{}'.format(self.actor_id), bonus_q05)
bonus_q50 = tf.Variable(0., name='novelty_bonus_q50')
s2 = tf.summary.scalar('Novelty_Bonus_q50_{}'.format(self.actor_id), bonus_q50)
bonus_q95 = tf.Variable(0., name='novelty_bonus_q95')
s3 = tf.summary.scalar('Novelty_Bonus_q95_{}'.format(self.actor_id), bonus_q95)
augmented_reward = tf.Variable(0., name='augmented_episode_reward')
s4 = tf.summary.scalar('Augmented_Episode_Reward_{}'.format(self.actor_id), augmented_reward)
return q_vars + [bonus_q05, bonus_q50, bonus_q95, augmented_reward]
#TODO: refactor to make this cleaner
def prepare_state(self, state, total_episode_reward, steps_at_last_reward,
ep_t, episode_ave_max_q, episode_over, bonuses, total_augmented_reward):
# Start a new game on reaching terminal state
if episode_over:
T = self.global_step.value() * self.max_local_steps
t = self.local_step
e_prog = float(t)/self.epsilon_annealing_steps
episode_ave_max_q = episode_ave_max_q/float(ep_t)
s1 = "Q_MAX {0:.4f}".format(episode_ave_max_q)
s2 = "EPS {0:.4f}".format(self.epsilon)
self.scores.insert(0, total_episode_reward)
if len(self.scores) > 100:
self.scores.pop()
logger.info('T{0} / STEP {1} / REWARD {2} / {3} / {4}'.format(
self.actor_id, T, total_episode_reward, s1, s2))
logger.info('ID: {0} -- RUNNING AVG: {1:.0f} ± {2:.0f} -- BEST: {3:.0f}'.format(
self.actor_id,
np.array(self.scores).mean(),
2*np.array(self.scores).std(),
max(self.scores),
))
self.log_summary(
total_episode_reward,
episode_ave_max_q,
self.epsilon,
np.percentile(bonuses, 5),
np.percentile(bonuses, 50),
np.percentile(bonuses, 95),
total_augmented_reward,
)
state = self.emulator.get_initial_state()
ep_t = 0
total_episode_reward = 0
episode_ave_max_q = 0
episode_over = False
return (
state,
total_episode_reward,
steps_at_last_reward,
ep_t,
episode_ave_max_q,
episode_over
)
def _double_dqn_op(self):
q_local_action = tf.cast(tf.argmax(
self.local_network.output_layer, axis=1), tf.int32)
q_target_max = utils.ops.slice_2d(
self.target_network.output_layer,
tf.range(0, self.batch_size),
q_local_action,
)
self.one_step_reward = tf.placeholder(tf.float32, self.batch_size, name='one_step_reward')
self.is_terminal = tf.placeholder(tf.bool, self.batch_size, name='is_terminal')
self.y_target = self.one_step_reward + self.cts_eta*self.gamma*q_target_max \
* (1 - tf.cast(self.is_terminal, tf.float32))
self.double_dqn_loss = self.local_network._value_function_loss(
self.local_network.q_selected_action
- tf.stop_gradient(self.y_target))
self.double_dqn_grads = tf.gradients(self.double_dqn_loss, self.local_network.params)
# def batch_update(self):
# if len(self.replay_memory) < self.replay_memory.maxlen//10:
# return
# s_i, a_i, r_i, s_f, is_terminal = self.replay_memory.sample_batch(self.batch_size)
# feed_dict={
# self.one_step_reward: r_i,
# self.target_network.input_ph: s_f,
# self.local_network.input_ph: np.vstack([s_i, s_f]),
# self.local_network.selected_action_ph: np.vstack([a_i, a_i]),
# self.is_terminal: is_terminal
# }
# grads = self.session.run(self.double_dqn_grads, feed_dict=feed_dict)
# self.apply_gradients_to_shared_memory_vars(grads)
def batch_update(self):
if len(self.replay_memory) < self.replay_memory.maxlen//10:
return
s_i, a_i, r_i, s_f, is_terminal = self.replay_memory.sample_batch(self.batch_size)
feed_dict={
self.local_network.input_ph: s_f,
self.target_network.input_ph: s_f,
self.is_terminal: is_terminal,
self.one_step_reward: r_i,
}
y_target = self.session.run(self.y_target, feed_dict=feed_dict)
feed_dict={
self.local_network.input_ph: s_i,
self.local_network.target_ph: y_target,
self.local_network.selected_action_ph: a_i
}
grads = self.session.run(self.local_network.get_gradients,
feed_dict=feed_dict)
self.apply_gradients_to_shared_memory_vars(grads)
def train(self):
""" Main actor learner loop for n-step Q learning. """
logger.debug("Actor {} resuming at Step {}, {}".format(self.actor_id,
self.global_step.value(), time.ctime()))
s = self.emulator.get_initial_state()
s_batch = list()
a_batch = list()
y_batch = list()
bonuses = deque(maxlen=1000)
episode_over = False
t0 = time.time()
global_steps_at_last_record = self.global_step.value()
while (self.global_step.value() < self.max_global_steps):
# # Sync local learning net with shared mem
# self.sync_net_with_shared_memory(self.local_network, self.learning_vars)
# self.save_vars()
rewards = list()
states = list()
actions = list()
max_q_values = list()
local_step_start = self.local_step
total_episode_reward = 0
total_augmented_reward = 0
episode_ave_max_q = 0
ep_t = 0
while not episode_over:
# Sync local learning net with shared mem
self.sync_net_with_shared_memory(self.local_network, self.learning_vars)
self.save_vars()
# Choose next action and execute it
a, q_values = self.choose_next_action(s)
new_s, reward, episode_over = self.emulator.next(a)
total_episode_reward += reward
max_q = np.max(q_values)
current_frame = new_s[...,-1]
bonus = self.density_model.update(current_frame)
bonuses.append(bonus)
# Rescale or clip immediate reward
reward = self.rescale_reward(self.rescale_reward(reward) + bonus)
total_augmented_reward += reward
ep_t += 1
rewards.append(reward)
states.append(s)
actions.append(a)
max_q_values.append(max_q)
s = new_s
self.local_step += 1
episode_ave_max_q += max_q
global_step, _ = self.global_step.increment()
if global_step % self.q_target_update_steps == 0:
self.update_target()
if global_step % self.density_model_update_steps == 0:
self.write_density_model()
# Sync local tensorflow target network params with shared target network params
if self.target_update_flags.updated[self.actor_id] == 1:
self.sync_net_with_shared_memory(self.target_network, self.target_vars)
self.target_update_flags.updated[self.actor_id] = 0
if self.density_model_update_flags.updated[self.actor_id] == 1:
self.read_density_model()
self.density_model_update_flags.updated[self.actor_id] = 0
if self.local_step % self.q_update_interval == 0:
self.batch_update()
if self.is_master() and (self.local_step % 500 == 0):
bonus_array = np.array(bonuses)
steps = global_step - global_steps_at_last_record
global_steps_at_last_record = global_step
logger.debug('Mean Bonus={:.4f} / Max Bonus={:.4f} / STEPS/s={}'.format(
bonus_array.mean(), bonus_array.max(), steps/float(time.time()-t0)))
t0 = time.time()
else:
#compute monte carlo return
mc_returns = np.zeros((len(rewards),), dtype=np.float32)
running_total = 0.0
for i, r in enumerate(reversed(rewards)):
running_total = r + self.gamma*running_total
mc_returns[len(rewards)-i-1] = running_total
mixed_returns = self.cts_eta*np.asarray(rewards) + (1-self.cts_eta)*mc_returns
#update replay memory
states.append(new_s)
episode_length = len(rewards)
for i in range(episode_length):
self.replay_memory.append(
states[i],
actions[i],
mixed_returns[i],
i+1 == episode_length)
s, total_episode_reward, _, ep_t, episode_ave_max_q, episode_over = \
self.prepare_state(s, total_episode_reward, self.local_step, ep_t, episode_ave_max_q, episode_over, bonuses, total_augmented_reward)
