def apply_adam_updates(self,
variables_server,
gradients,
learning_rate,
epsilon=1e-6):
update_steps = hlp.load_object(
variables_server.get('update_steps')) + 1
variables_server.set('update_steps', hlp.dump_object(update_steps))
learning_rate = learning_rate * (
(1 - 0.999**update_steps)**0.5) / (1 - 0.9**update_steps)
for i, gradient in enumerate(gradients):
momentum = hlp.load_object(
variables_server.get('momentum_{}'.format(i)))
momentum = 0.999 * momentum + (1 - 0.999) * gradient * gradient
variables_server.set('momentum_{}'.format(i),
hlp.dump_object(momentum))
velocity = hlp.load_object(
variables_server.get('velocity_{}'.format(i)))
velocity = 0.9 * velocity + (1 - 0.9) * gradient
variables_server.set('velocity_{}'.format(i),
hlp.dump_object(velocity))
weight = hlp.load_object(
variables_server.get('weight_{}'.format(i)))
new_weight = weight - velocity * learning_rate / (
(momentum**0.5) + epsilon)
variables_server.set('weight_{}'.format(i),
hlp.dump_object(new_weight))
return update_steps
def make_rollout(self):
variables_server = Redis(port=12000)
if self.scale != 'off':
try:
means = hlp.load_object(variables_server.get("means"))
stds = hlp.load_object(variables_server.get("stds"))
self.sess.run(self.norm_set_op, feed_dict=dict(zip(self.norm_phs, [means, stds])))
except:
pass
try:
weights = [hlp.load_object(variables_server.get("weight_{}".format(i))) for i in
range(len(self.weights))]
self.set_weights(weights)
except:
pass
env = self.env
n_tasks = self.n_tests
timestep = 0
i_task = 0
paths = []
while i_task < n_tasks:
path = {}
observations, action_tuples, rewards, dist_tuples, timestamps = [], [], [], [], []
sums = np.zeros((1, env.get_observation_space()))
sumsqrs = np.zeros(sums.shape)
env.reset()
while not env.done and env.timestamp < self.timesteps_per_launch:
sums += env.features
sumsqrs += np.square(env.features)
observations.append(env.features[0])
timestamps.append(env.timestamp)
actions = self.act(env.features)
env.step(actions)
timestep += 1
action_tuples.append(actions)
rewards.append(env.reward)
path["observations"] = np.array(observations)
path["action_tuples"] = np.array(action_tuples)
path["rewards"] = np.array(rewards)
if not self.test_mode:
path["dist_tuples"] = np.array(dist_tuples)
path["timestamps"] = np.array(timestamps)
path["sumobs"] = sums
path["sumsqrobs"] = sumsqrs
path["terminated"] = env.done
path["total"] = env.get_total_reward()
paths.append(path)
i_task += 1
if self.distributed:
variables_server.set("paths_{}".format(self.id_worker), hlp.dump_object(paths))
else:
self.paths = paths
def work(self):
variables_server = Redis(port=12000)
if self.scale != 'off':
try:
means = hlp.load_object(variables_server.get("means"))
stds = hlp.load_object(variables_server.get("stds"))
self.sess.run(self.norm_set_op, feed_dict=dict(zip(self.norm_phs, [means, stds])))
except:
pass
try:
weights = [hlp.load_object(variables_server.get("weight_{}".format(i))) for i in
range(len(self.weights))]
self.set_weights(weights)
except:
pass
env = self.env
while True:
observations, action_tuples, rewards, dist_tuples, timestamps = [], [], [], [], []
for _ in range(self.n_steps):
observations.append(env.features[0])
timestamps.append(env.timestamp)
actions = self.act(env.features)
env.step(actions)
action_tuples.append(actions)
rewards.append(env.reward)
if env.done or env.timestamp > self.timesteps_per_launch:
variables_server.lpush('results', hlp.dump_object(env.get_total_reward()))
print("Episode reward: {}".format(env.get_total_reward()), "Length: {}".format(env.timestamp))
break
timestamps.append(env.timestamp)
observations_batch = np.array(observations)
actions_batch = np.array(action_tuples)
feed_dict = {self.state_input: observations_batch,
self.targets["action"]: actions_batch}
if env.done or env.timestamp > self.timesteps_per_launch:
rewards.append(0)
env.reset()
else:
obs = observations[-1]
rewards.append(self.sess.run(self.value, feed_dict={self.state_input: obs.reshape((1,) + obs.shape)}))
returns_batch = hlp.discount(np.array(rewards), self.gamma, np.array(timestamps))[:-1]
values = self.sess.run(self.value, feed_dict)
feed_dict[self.targets["advantage"]] = returns_batch - values
feed_dict[self.targets["return"]] = returns_batch
gradients = self.sess.run(self.gradients, feed_dict)
self.apply_adam_updates(variables_server, gradients, self.learning_rate)
weights = [hlp.load_object(variables_server.get("weight_{}".format(i))) for i in
range(len(self.weights))]
self.set_weights(weights)
def train(self):
cmd_server = 'redis-server --port 12000'
p = subprocess.Popen(cmd_server, shell=True, preexec_fn=os.setsid)
self.variables_server = Redis(port=12000)
means = "-"
stds = "-"
if self.scale != 'off':
if self.timestep == 0:
print("Time to measure features!")
if self.distributed:
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = False
self.make_rollout()
paths = self.paths
for path in paths:
self.sums += path["sumobs"]
self.sumsqrs += path["sumsqrobs"]
self.sumtime += len(path["rewards"])
stds = np.sqrt((self.sumsqrs - np.square(self.sums) / self.sumtime) / (self.sumtime - 1))
means = self.sums / self.sumtime
print("Init means: {}".format(means))
print("Init stds: {}".format(stds))
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op, feed_dict=dict(zip(self.norm_phs, [means, stds])))
print("Let's go!")
self.update_target_weights(alpha=1.0)
index_replay = 0
iteration = 0
episode = 0
idxs_range = np.arange(self.xp_size)
xp_replay_state = np.zeros(shape=(self.xp_size, self.env.get_observation_space()))
xp_replay_next_state = np.zeros(shape=(self.xp_size, self.env.get_observation_space()))
xp_replay_reward = np.zeros(shape=(self.xp_size,))
xp_replay_action = np.zeros(shape=(self.xp_size,))
xp_replay_terminal = np.zeros(shape=(self.xp_size,))
if self.prioritized:
xp_replay_priority = np.zeros(shape=(self.xp_size,))
self.max_prior = 1
start_time = time.time()
self.last_state = self.env.reset()
discounts = self.gamma ** np.arange(self.n_steps)
self.last_rewards = np.zeros(shape=(self.n_steps,))
self.last_states = np.zeros(shape=(self.n_steps, self.n_features))
self.last_actions = np.zeros(shape=(self.n_steps, ))
buffer_index = 0
env = self.env
while True:
if iteration <= self.random_steps:
actions = env.env.action_space.sample()
else:
actions = self.act(env.features, exploration=True)
self.last_states[buffer_index] = env.features.reshape(-1)
self.last_actions[buffer_index] = actions
env.step([actions])
self.last_rewards[buffer_index] = env.reward
buffer_index = (buffer_index + 1) % self.n_steps
if env.timestamp >= self.n_steps:
xp_replay_state[index_replay] = np.copy(self.last_states[buffer_index])
xp_replay_next_state[index_replay] = env.features.reshape(-1)
discounted_return = np.sum(discounts*self.last_rewards[np.roll(np.arange(self.n_steps), -(buffer_index))])
xp_replay_reward[index_replay] = discounted_return
xp_replay_action[index_replay] = self.last_actions[buffer_index]
xp_replay_terminal[index_replay] = env.done
if self.prioritized:
xp_replay_priority[index_replay] = self.max_prior
index_replay = (index_replay + 1) % self.xp_size
if env.done or env.timestamp > self.timesteps_per_launch:
episode += 1
print("Episode #{}".format(episode), env.get_total_reward())
self.train_scores.append(env.get_total_reward())
for i in range(1, self.n_steps):
buffer_index = (buffer_index + 1) % self.n_steps
xp_replay_state[index_replay] = np.copy(self.last_states[buffer_index])
xp_replay_next_state[index_replay] = env.features.reshape(-1)
discounted_return = np.sum(
discounts[:self.n_steps-i] * self.last_rewards[np.roll(np.arange(self.n_steps), -(buffer_index))[:self.n_steps-i]])
xp_replay_reward[index_replay] = discounted_return
xp_replay_action[index_replay] = self.last_actions[buffer_index]
xp_replay_terminal[index_replay] = env.done
index_replay = (index_replay + 1) % self.xp_size
env.reset()
self.last_rewards = np.zeros(shape=(self.n_steps,))
self.last_states = np.zeros(shape=(self.n_steps, self.n_features))
self.last_actions = np.zeros(shape=(self.n_steps,))
buffer_index = 0
self.last_state = env.features
if iteration % 1000 == 0:
print("Iteration #{}".format(iteration))
self.save(self.config[:-5])
if iteration > self.random_steps:
if self.prioritized:
max_id = np.min([xp_replay_state.shape[0], iteration])
probs = xp_replay_priority[:max_id]/np.sum(xp_replay_priority[:max_id])
idxs = np.random.choice(idxs_range[:max_id], size=self.batch_size, p=probs)
importance_weights = (1/(max_id*probs[idxs]))**self.prior_beta
else:
idxs = np.random.randint(np.min([xp_replay_state.shape[0], iteration]), size=self.batch_size)
importance_weights = np.ones(shape=(self.batch_size,))
state_batch = xp_replay_state[idxs]
next_state_batch = xp_replay_next_state[idxs]
action_batch = xp_replay_action[idxs]
reward_batch = xp_replay_reward[idxs]
done_batch = xp_replay_terminal[idxs]
feed_dict = {
self.state_input: state_batch,
self.next_state_input: next_state_batch,
self.action_input: action_batch,
self.importance_weights: importance_weights
}
target_atom_probs = self.sess.run(self.target_atom_probs, feed_dict)
target_atom_probs = np.exp(target_atom_probs)
if not self.double:
target_q_values = target_atom_probs * np.tile(np.arange(self.n_atoms).reshape((1, 1, self.n_atoms)), [self.batch_size, 1, 1])
target_q_values = np.sum(target_q_values, axis=2)
target_greedy_actions = np.argmax(target_q_values, axis=1).astype(np.int32).reshape((-1, 1))
target_probs = target_atom_probs[np.arange(self.batch_size).reshape((-1, 1)), target_greedy_actions]
else:
feed_dict[self.state_input] = next_state_batch
atom_probs = self.sess.run(self.atom_probs, feed_dict)
atom_probs = np.exp(atom_probs)
q_values = atom_probs * np.tile(np.arange(self.n_atoms).reshape((1, 1, self.n_atoms)),
[self.batch_size, 1, 1])
q_values = np.sum(q_values, axis=2)
greedy_actions = np.argmax(q_values, axis=1).astype(np.int32).reshape((-1, 1))
target_probs = target_atom_probs[np.arange(self.batch_size).reshape((-1, 1)), greedy_actions]
feed_dict[self.state_input] = state_batch
atom_values = np.arange(self.n_atoms, dtype=np.float32).reshape((-1, self.n_atoms))
atom_values = 2 * self.max_q_magnitude * (
np.tile(atom_values, [self.batch_size, 1]) / (self.n_atoms - 1) - 0.5)
atom_new_values = np.clip((self.gamma**self.n_steps) * atom_values * (1-done_batch).reshape(-1, 1) + reward_batch.reshape((-1, 1)),
- self.max_q_magnitude, self.max_q_magnitude)
new_positions = ((atom_new_values / (2 * self.max_q_magnitude) + 0.5) * (self.n_atoms - 1)).reshape((-1))
lower = np.floor(new_positions).astype(np.int32).reshape(-1)
upper = np.floor(new_positions).astype(np.int32).reshape(-1) + 1
final_target_probs = np.zeros(shape=(self.batch_size, self.n_atoms+1, self.n_atoms))
final_target_probs[np.sort(np.tile(np.arange(self.batch_size), [self.n_atoms])), lower, np.tile(np.arange(self.n_atoms), [self.batch_size])] += (upper-new_positions) * target_probs.reshape((-1))
final_target_probs[np.sort(np.tile(np.arange(self.batch_size), [self.n_atoms])), upper, np.tile(np.arange(self.n_atoms), [self.batch_size])] += (new_positions-lower) * target_probs.reshape((-1))
final_target_probs = np.sum(final_target_probs, axis=2)[:, :-1]
feed_dict[self.target_probs] = final_target_probs
KLs = self.sess.run([self.loss, self.train_op], feed_dict)[0]
if self.prioritized:
xp_replay_priority[idxs] = KLs ** self.prior_alpha
self.update_target_weights()
if iteration % self.test_every == 0:
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i), hlp.dump_object(weight))
print("Time to test!")
if self.distributed:
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i), hlp.dump_object(weight))
worker_args = \
{
'config': self.config,
'test_mode': True,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = True
self.make_rollout()
paths = self.paths
total_rewards = np.array([path["total"] for path in paths])
eplens = np.array([len(path["rewards"]) for path in paths])
print("""
-------------------------------------------------------------
Mean test score: {test_scores}
Mean test episode length: {test_eplengths}
Max test score: {max_test}
Time for iteration: {tt}
Mean of features: {means}
Std of features: {stds}
-------------------------------------------------------------
""".format(
means=means,
stds=stds,
test_scores=np.mean(total_rewards),
test_eplengths=np.mean(eplens),
max_test=np.max(total_rewards),
tt=time.time() - start_time
))
start_time = time.time()
self.test_scores.append(np.mean(total_rewards))
iteration += 1
self.timestep += 1
def train(self):
cmd_server = 'redis-server --port 12000'
p = subprocess.Popen(cmd_server, shell=True, preexec_fn=os.setsid)
self.variables_server = Redis(port=12000)
means = "-"
stds = "-"
if self.scale != 'off':
if self.timestep == 0:
print("Time to measure features!")
if self.distributed:
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = False
self.make_rollout()
paths = self.paths
for path in paths:
self.sums += path["sumobs"]
self.sumsqrs += path["sumsqrobs"]
self.sumtime += len(path["rewards"])
stds = np.sqrt(
(self.sumsqrs - np.square(self.sums) / self.sumtime) /
(self.sumtime - 1))
means = self.sums / self.sumtime
print("Init means: {}".format(means))
print("Init stds: {}".format(stds))
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs,
[means, stds])))
while True:
print("Iteration {}".format(self.timestep))
start_time = time.time()
weight_noises = []
random.seed()
seed_for_random = random.randint(0, np.iinfo(np.int32).max)
np.random.seed(seed_for_random)
seeds = np.random.randint(-np.iinfo(np.int32).min +
np.iinfo(np.int32).max,
size=self.n_tasks_all)
self.variables_server.set("seeds", hlp.dump_object(seeds))
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
weight_noises.append(
np.empty((self.n_tasks_all, ) + weight.shape))
for index in range(self.n_tasks_all):
np.random.seed(seeds[index])
for i, weight in enumerate(weights):
weight_noises[i][index] = np.random.normal(
size=weight.shape)
if self.distributed:
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args,
self.n_workers,
command='rollout_with_noise')
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = False
self.make_rollout()
paths = self.paths
scores = []
train_lengths = []
for i in range(self.n_tasks_all):
scores.append(
hlp.load_object(
self.variables_server.get("scores_" + str(i))))
train_lengths.append(
hlp.load_object(
self.variables_server.get("eplen_" + str(i))))
scores.append(
hlp.load_object(
self.variables_server.get("scores_" + str(-i))))
train_lengths.append(
hlp.load_object(
self.variables_server.get("eplen_" + str(-i))))
scores = np.array(scores)
train_mean_score = np.mean(scores)
ranks = np.zeros(shape=scores.shape)
if self.normalize == 'ranks':
ranks[np.argsort(scores)] = np.arange(
ranks.shape[0], dtype=np.float32) / (ranks.shape[0] - 1)
ranks -= 0.5
elif self.normalize == 'center':
ranks = scores[:]
ranks -= train_mean_score
ranks /= (np.std(ranks, ddof=1) + 0.001)
gradients = [np.zeros(w.get_shape()) for w in self.weights]
for i, weight in enumerate(weights):
for index in 2 * np.arange(seeds.shape[0]):
gradients[i] += weight_noises[i][index // 2] * (
ranks[index] - ranks[index + 1]) / self.n_tasks_all
gradients[i] -= self.l1_reg * weights[i]
if self.adam:
self.apply_adam_updates(gradients)
else:
for i, weight in enumerate(weights):
weights[i] += self.learning_rate * gradients[i]
self.sess.run(self.set_op,
feed_dict=dict(zip(self.weights_phs, weights)))
print("Time to testing!")
if self.distributed:
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
worker_args = \
{
'config': self.config,
'test_mode': True,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = True
self.make_rollout()
paths = self.paths
total_rewards = np.array([path["total"] for path in paths])
eplens = np.array([len(path["rewards"]) for path in paths])
if self.scale:
for i in range(self.n_tasks_all):
self.sums += hlp.load_object(
self.variables_server.get("sum_{}".format(i)))
self.sumsqrs += hlp.load_object(
self.variables_server.get("sumsqr_{}".format(i)))
self.sumtime += np.sum(train_lengths)
stds = np.sqrt(
(self.sumsqrs - np.square(self.sums) / self.sumtime) /
(self.sumtime - 1))
means = self.sums / self.sumtime
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs,
[means, stds])))
print("""
-------------------------------------------------------------
Mean test score: {test_scores}
Mean train score: {train_scores}
Mean test episode length: {test_eplengths}
Mean train episode length: {train_eplengths}
Max test score: {max_test}
Max train score: {max_train}
Mean of features: {means}
Std of features: {stds}
Time for iteration: {tt}
-------------------------------------------------------------
""".format(means=means,
stds=stds,
test_scores=np.mean(total_rewards),
test_eplengths=np.mean(eplens),
train_scores=train_mean_score,
train_eplengths=np.mean(train_lengths),
max_test=np.max(total_rewards),
max_train=np.max(scores),
tt=time.time() - start_time))
self.train_scores.append(train_mean_score)
self.test_scores.append(np.mean(total_rewards))
if self.timestep % self.save_every == 0:
self.save(self.config[:-5])
def rollout_with_noise(self):
variables_server = Redis(port=12000)
if self.scale != 'off':
means = hlp.load_object(variables_server.get("means"))
stds = hlp.load_object(variables_server.get("stds"))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs, [means, stds])))
weights = [
hlp.load_object(variables_server.get("weight_{}".format(i)))
for i in range(len(self.weights))
]
self.set_weights(weights)
env = self.env
seeds = hlp.load_object(variables_server.get("seeds"))
for id_task_for_worker in range(self.n_tasks):
id_task = id_task_for_worker * self.n_workers + self.id_worker
if id_task % self.report_every == 0:
print("Rollout # {} of {}".format(id_task, self.n_tasks_all))
seed = seeds[id_task]
np.random.seed(seed)
noises = []
for i, weight in enumerate(weights):
noise = np.random.normal(size=weight.shape)
noises.append(noise)
weights[i] += self.noise_scale * noise
self.set_weights(weights)
env.reset()
sums = np.zeros((1, env.get_observation_space()))
sumsqrs = np.zeros(sums.shape)
while not env.done and env.timestamp < self.timesteps_per_launch:
sums += env.features
sumsqrs += np.square(env.features)
actions = self.act(env.features)
env.step(actions)
variables_server.set("scores_{}".format(id_task),
hlp.dump_object(env.get_total_reward()))
variables_server.set("eplen_{}".format(id_task),
hlp.dump_object(env.timestamp))
for i, weight in enumerate(weights):
noise = noises[i]
weights[i] -= 2 * self.noise_scale * noise
self.set_weights(weights)
env.reset()
while not env.done and env.timestamp < self.timesteps_per_launch:
sums += env.features
sumsqrs += np.square(env.features)
actions = self.act(env.features)
env.step(actions)
variables_server.set("scores_{}".format(-id_task),
hlp.dump_object(env.get_total_reward()))
variables_server.set("eplen_{}".format(-id_task),
hlp.dump_object(env.timestamp))
variables_server.set("sum_{}".format(id_task),
hlp.dump_object(sums))
variables_server.set("sumsqr_{}".format(id_task),
hlp.dump_object(sumsqrs))
for i, weight in enumerate(weights):
noise = noises[i]
weights[i] += self.noise_scale * noise
self.set_weights(weights)
def train(self):
cmd_server = 'redis-server --port 12000'
p = subprocess.Popen(cmd_server, shell=True, preexec_fn=os.setsid)
self.variables_server = Redis(port=12000)
means = "-"
stds = "-"
if self.scale != 'off':
if self.timestep == 0:
print("Time to measure features!")
if self.distributed:
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = False
self.make_rollout()
paths = self.paths
for path in paths:
self.sums += path["sumobs"]
self.sumsqrs += path["sumsqrobs"]
self.sumtime += path["observations"].shape[0]
stds = np.sqrt(
(self.sumsqrs - np.square(self.sums) / self.sumtime) /
(self.sumtime - 1))
means = self.sums / self.sumtime
print("Init means: {}".format(means))
print("Init stds: {}".format(stds))
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs, [means, stds])))
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
self.variables_server.set('momentum_{}'.format(i),
hlp.dump_object(np.zeros(weight.shape)))
self.variables_server.set('velocity_{}'.format(i),
hlp.dump_object(np.zeros(weight.shape)))
self.variables_server.set('update_steps', hlp.dump_object(0))
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args,
self.n_workers,
command='work',
wait=False)
while True:
time.sleep(self.test_every)
print("Time for testing!")
if self.distributed:
worker_args = \
{
'config': self.config,
'test_mode': True,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = True
self.make_rollout()
paths = self.paths
total_rewards = np.array([path["total"] for path in paths])
eplens = np.array([len(path["rewards"]) for path in paths])
print("""
-------------------------------------------------------------
Mean test score: {test_scores}
Mean test episode length: {test_eplengths}
Max test score: {max_test}
Number of train episodes: {number}
Mean of features: {means}
Std of features: {stds}
-------------------------------------------------------------
""".format(means=means,
stds=stds,
test_scores=np.mean(total_rewards),
test_eplengths=np.mean(eplens),
max_test=np.max(total_rewards),
number=self.variables_server.llen('results')))
self.timestep += 1
self.train_scores = [
hlp.load_object(res)
for res in self.variables_server.lrange('results', 0, -1)
][::-1]
self.test_scores.append(np.mean(total_rewards))
if self.timestep % self.save_every == 0:
self.save(self.config[:-5])
def train(self):
cmd_server = 'redis-server --port 12000'
p = subprocess.Popen(cmd_server, shell=True, preexec_fn=os.setsid)
self.variables_server = Redis(port=12000)
if self.scale:
if self.timestep == 0:
self.test_mode = False
self.make_rollout()
paths = self.paths
for path in paths:
self.sums += path["sumobs"]
self.sumsqrs += path["sumsqrobs"]
self.sumtime += len(path["rewards"])
stds = np.sqrt(
(self.sumsqrs - np.square(self.sums) / self.sumtime) /
(self.sumtime - 1))
means = self.sums / self.sumtime
print("Init means: {}".format(means))
print("Init stds: {}".format(stds))
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs, [means, stds])))
print("Let's go!")
self.update_target_weights(alpha=1.0)
index_replay = 0
iteration = 0
episode = 0
xp_replay_state = np.zeros(shape=(self.xp_size,
self.env.get_observation_space()))
xp_replay_next_state = np.zeros(
shape=(self.xp_size, self.env.get_observation_space()))
xp_replay_reward = np.zeros(shape=(self.xp_size, ))
xp_replay_action = np.zeros(shape=(self.xp_size, len(self.n_actions)))
xp_replay_terminal = np.zeros(shape=(self.xp_size, ))
start_time = time.time()
self.last_state = self.env.reset()
env = self.env
while True:
if iteration <= self.random_steps:
actions = env.env.action_space.sample()
else:
actions = self.act(env.features)
actions += np.random.normal(0,
scale=self.action_noise,
size=actions.shape)
env.step(actions)
xp_replay_state[index_replay] = self.last_state.reshape(-1)
xp_replay_next_state[index_replay] = env.features.reshape(-1)
xp_replay_reward[index_replay] = env.reward
xp_replay_action[index_replay] = actions.reshape(-1)
xp_replay_terminal[index_replay] = env.done
index_replay = (index_replay + 1) % self.xp_size
if env.done or env.timestamp > self.timesteps_per_launch:
episode += 1
print("Episode #{}".format(episode), env.get_total_reward())
self.train_scores.append(env.get_total_reward())
env.reset()
self.last_state = env.features
if iteration % 1000 == 0:
print("Iteration #{}".format(iteration))
self.save(self.config[:-5])
if iteration > self.random_steps:
idxs = np.random.randint(np.min(
[xp_replay_state.shape[0], iteration]),
size=self.batch_size)
feed_dict = {
self.state_input: xp_replay_state[idxs],
self.next_state_input: xp_replay_next_state[idxs],
self.action_input: xp_replay_action[idxs],
self.reward_input: xp_replay_reward[idxs],
self.done_input: xp_replay_terminal[idxs]
}
self.sess.run([self.value_train_op], feed_dict)
self.sess.run(self.train_actor_op, feed_dict)
self.update_target_weights()
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
if iteration % self.test_every == 0:
print("Time to test!")
self.test_mode = True
self.make_rollout()
paths = self.paths
total_rewards = np.array([path["total"] for path in paths])
eplens = np.array([len(path["rewards"]) for path in paths])
print("""
-------------------------------------------------------------
Mean test score: {test_scores}
Mean test episode length: {test_eplengths}
Max test score: {max_test}
Time for iteration: {tt}
-------------------------------------------------------------
""".format(
test_scores=np.mean(total_rewards),
test_eplengths=np.mean(eplens),
max_test=np.max(total_rewards),
tt=time.time() - start_time))
start_time = time.time()
self.test_scores.append(np.mean(total_rewards))
self.timestep += 1
iteration += 1
def train(self):
cmd_server = 'redis-server --port 12000'
p = subprocess.Popen(cmd_server, shell=True, preexec_fn=os.setsid)
self.variables_server = Redis(port=12000)
means = "-"
stds = "-"
if self.scale != 'off':
if self.timestep == 0:
print("Time to measure features!")
if self.distributed:
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = False
self.make_rollout()
paths = self.paths
for path in paths:
self.sums += path["sumobs"]
self.sumsqrs += path["sumsqrobs"]
self.sumtime += path["observations"].shape[0]
stds = np.sqrt(
(self.sumsqrs - np.square(self.sums) / self.sumtime) /
(self.sumtime - 1))
means = self.sums / self.sumtime
print("Init means: {}".format(means))
print("Init stds: {}".format(stds))
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs, [means, stds])))
while True:
print("Iteration {}".format(self.timestep))
start_time = time.time()
if self.distributed:
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
worker_args = \
{
'config': self.config,
'test_mode': False,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = False
self.make_rollout()
paths = self.paths
observations = np.concatenate(
[path["observations"] for path in paths])
actions = np.concatenate([path["action_tuples"] for path in paths])
action_dists = []
for _ in range(len(self.n_actions)):
action_dists.append([])
returns = []
advantages = []
for path in paths:
self.sums += path["sumobs"]
self.sumsqrs += path["sumsqrobs"]
self.sumtime += path["rewards"].shape[0]
dists = path["dist_tuples"]
for i in range(len(self.n_actions)):
action_dists[i] += [dist[i][0] for dist in dists]
returns += hlp.discount(path["rewards"], self.gamma,
path["timestamps"]).tolist()
values = self.sess.run(
self.value,
feed_dict={self.state_input: path["observations"]})
values = np.append(values,
0 if path["terminated"] else values[-1])
deltas = (path["rewards"] + self.gamma * values[1:] -
values[:-1])
advantages += hlp.discount(deltas, self.gamma,
path["timestamps"]).tolist()
returns = np.array(returns)
advantages = np.array(advantages)
if self.normalize == 'ranks':
ranks = np.zeros_like(advantages)
ranks[np.argsort(advantages)] = np.arange(
ranks.shape[0], dtype=np.float32) / (ranks.shape[0] - 1)
ranks -= 0.5
advantages = ranks[:]
elif self.normalize == 'center':
advantages -= np.mean(advantages)
advantages /= (np.std(advantages, ddof=1) + 0.001)
feed_dict = {
self.state_input: observations,
self.targets["return"]: returns,
self.targets["advantage"]: advantages
}
for i in range(len(self.n_actions)):
feed_dict[self.targets["old_dist_{}".format(i)]] = np.array(
action_dists[i])
feed_dict[self.targets["action_{}".format(i)]] = actions[:, i]
for i in range(self.value_updates):
self.sess.run(self.value_train_op, feed_dict)
train_rewards = np.array([path["rewards"].sum() for path in paths])
train_lengths = np.array([len(path["rewards"]) for path in paths])
thprev = self.get_flat()
def fisher_vector_product(p):
feed_dict[self.targets["flat_tangent"]] = p
return self.sess.run(self.fisher_vector_product,
feed_dict) + 0.1 * p
g = self.sess.run(self.policy_grad, feed_dict)
stepdir = hlp.conjugate_gradient(fisher_vector_product, -g)
shs = .5 * stepdir.dot(fisher_vector_product(stepdir))
lm = np.sqrt(shs / self.max_kl)
fullstep = stepdir / (lm + 1e-18)
def loss_kl(th):
self.set_from_flat(th)
return self.sess.run([self.loss, self.KL], feed_dict=feed_dict)
theta = hlp.linesearch(loss_kl, thprev, fullstep, self.max_kl)
self.set_from_flat(theta)
lossafter, kloldnew = self.sess.run([self.loss, self.KL],
feed_dict=feed_dict)
print("Time for testing!")
if self.distributed:
weights = self.get_weights()
for i, weight in enumerate(weights):
self.variables_server.set("weight_" + str(i),
hlp.dump_object(weight))
worker_args = \
{
'config': self.config,
'test_mode': True,
}
hlp.launch_workers(worker_args, self.n_workers)
paths = []
for i in range(self.n_workers):
paths += hlp.load_object(
self.variables_server.get("paths_{}".format(i)))
else:
self.test_mode = True
self.make_rollout()
paths = self.paths
total_rewards = np.array([path["total"] for path in paths])
eplens = np.array([len(path["rewards"]) for path in paths])
if self.scale != 'full':
stds = np.sqrt(
(self.sumsqrs - np.square(self.sums) / self.sumtime) /
(self.sumtime - 1))
means = self.sums / self.sumtime
self.variables_server.set("means", hlp.dump_object(means))
self.variables_server.set("stds", hlp.dump_object(stds))
self.sess.run(self.norm_set_op,
feed_dict=dict(zip(self.norm_phs,
[means, stds])))
print("""
-------------------------------------------------------------
Mean test score: {test_scores}
Mean train score: {train_scores}
Mean test episode length: {test_eplengths}
Mean train episode length: {train_eplengths}
Max test score: {max_test}
Max train score: {max_train}
KL between old and new {kl}
Loss after update {loss}
Mean of features: {means}
Std of features: {stds}
-------------------------------------------------------------
""".format(means=means,
stds=stds,
test_scores=np.mean(total_rewards),
test_eplengths=np.mean(eplens),
train_scores=np.mean(train_rewards),
train_eplengths=np.mean(train_lengths),
max_test=np.max(total_rewards),
max_train=np.max(train_rewards),
kl=kloldnew,
loss=lossafter))
self.timestep += 1
self.train_scores.append(np.mean(train_rewards))
self.test_scores.append(np.mean(total_rewards))
if self.timestep % self.save_every == 0:
self.save(self.config[:-5])