def learn(policy,
env,
seed,
ob_space,
ac_space,
save_name,
nsteps=5,
total_timesteps=int(80e6),
vf_coef=0.5,
ent_coef=0.01,
max_grad_norm=0.5,
lr=7e-4,
lrschedule='linear',
epsilon=1e-5,
alpha=0.99,
gamma=0.99,
log_interval=100):
set_global_seeds(seed)
nenvs = env.num_envs
#ob_space = env.observation_space
#ac_space = env.action_space
save_dir = './model/' + save_name + '.ckt'
summary_dir = './summary/' + save_name
model = Model(policy=policy,
ob_space=ob_space,
ac_space=ac_space,
nenvs=nenvs,
nsteps=nsteps,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
lr=lr,
alpha=alpha,
epsilon=epsilon,
total_timesteps=total_timesteps,
lrschedule=lrschedule,
summary_dir=summary_dir)
runner = Runner(env, model, ob_space=ob_space, nsteps=nsteps, gamma=gamma)
nbatch = nenvs * nsteps
tstart = time.time()
train_writer = model.train_writer
episode_stats = EpisodeStats(nsteps, nenvs)
for update in range(1, total_timesteps // nbatch + 1):
obs, states, rewards, masks, actions, values, raw_rewards = runner.run(
)
episode_stats.feed(raw_rewards, masks)
mean_reward = episode_stats.mean_reward()
mean_reward = np.asarray(mean_reward, dtype=np.float32)
policy_loss, value_loss, policy_entropy, summary = model.train(
obs, states, mean_reward, rewards, masks, actions, values)
train_writer.add_summary(summary, update * nbatch)
nseconds = time.time() - tstart
fps = int((update * nbatch) / nseconds)
if update % log_interval == 0 or update == 1:
ev = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", update * nbatch)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy", float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("explained_variance", float(ev))
logger.record_tabular("episode_reward",
episode_stats.mean_reward())
logger.record_tabular("episode_length",
episode_stats.mean_length())
logger.dump_tabular()
model.save(save_dir)
env.close()
return model
def learn(self,
total_timesteps,
callback=None,
log_interval=1,
tb_log_name="PPO2",
reset_num_timesteps=True):
# Transform to callable if needed
self.learning_rate = get_schedule_fn(self.learning_rate)
self.cliprange = get_schedule_fn(self.cliprange)
cliprange_vf = get_schedule_fn(self.cliprange_vf)
bestscore = 0
new_tb_log = self._init_num_timesteps(reset_num_timesteps)
callback = self._init_callback(callback)
fig = plt.figure()
ax = fig.add_subplot(111)
x, y = [0], [0]
with SetVerbosity(self.verbose), TensorboardWriter(self.graph, self.tensorboard_log, tb_log_name, new_tb_log) \
as writer:
self._setup_learn()
episode_stats = EpisodeStats(self.n_steps, self.n_envs)
t_first_start = time.time()
n_updates = total_timesteps // self.n_batch
callback.on_training_start(locals(), globals())
for update in range(1, n_updates + 1):
assert self.n_batch % self.nminibatches == 0, (
"The number of minibatches (`nminibatches`) "
"is not a factor of the total number of samples "
"collected per rollout (`n_batch`), "
"some samples won't be used.")
batch_size = self.n_batch // self.nminibatches
t_start = time.time()
frac = 1.0 - (update - 1.0) / n_updates
lr_now = self.learning_rate(frac)
cliprange_now = self.cliprange(frac)
cliprange_vf_now = cliprange_vf(frac)
callback.on_rollout_start()
# true_reward is the reward without discount
rollout = self.runner.run(callback)
# Unpack
obs, returns, masks, actions, values, neglogpacs, states, ep_infos, true_reward = rollout
callback.update_locals(locals())
callback.on_rollout_end()
# Early stopping due to the callback
if not self.runner.continue_training:
break
episode_stats.feed(true_reward, masks)
self.ep_info_buf.extend(ep_infos)
mb_loss_vals = []
if states is None: # nonrecurrent version
update_fac = max(
self.n_batch // self.nminibatches // self.noptepochs,
1)
inds = np.arange(self.n_batch)
for epoch_num in range(self.noptepochs):
np.random.shuffle(inds)
for start in range(0, self.n_batch, batch_size):
timestep = self.num_timesteps // update_fac + (
(epoch_num * self.n_batch + start) //
batch_size)
end = start + batch_size
mbinds = inds[start:end]
slices = (arr[mbinds]
for arr in (obs, returns, masks, actions,
values, neglogpacs))
mb_loss_vals.append(
self._train_step(
lr_now,
cliprange_now,
*slices,
writer=writer,
update=timestep,
cliprange_vf=cliprange_vf_now))
else: # recurrent version
update_fac = max(
self.n_batch // self.nminibatches // self.noptepochs //
self.n_steps, 1)
assert self.n_envs % self.nminibatches == 0
env_indices = np.arange(self.n_envs)
flat_indices = np.arange(self.n_envs *
self.n_steps).reshape(
self.n_envs, self.n_steps)
envs_per_batch = batch_size // self.n_steps
for epoch_num in range(self.noptepochs):
np.random.shuffle(env_indices)
for start in range(0, self.n_envs, envs_per_batch):
timestep = self.num_timesteps // update_fac + (
(epoch_num * self.n_envs + start) //
envs_per_batch)
end = start + envs_per_batch
mb_env_inds = env_indices[start:end]
mb_flat_inds = flat_indices[mb_env_inds].ravel()
slices = (arr[mb_flat_inds]
for arr in (obs, returns, masks, actions,
values, neglogpacs))
mb_states = states[mb_env_inds]
mb_loss_vals.append(
self._train_step(
lr_now,
cliprange_now,
*slices,
update=timestep,
writer=writer,
states=mb_states,
cliprange_vf=cliprange_vf_now))
loss_vals = np.mean(mb_loss_vals, axis=0)
t_now = time.time()
fps = int(self.n_batch / (t_now - t_start))
if writer is not None:
total_episode_reward_logger(
self.episode_reward,
true_reward.reshape((self.n_envs, self.n_steps)),
masks.reshape((self.n_envs, self.n_steps)), writer,
self.num_timesteps)
if self.verbose == 1 and (update % log_interval == 0
or update == 1):
explained_var = explained_variance(values, returns)
logger.logkv("serial_timesteps", update * self.n_steps)
logger.logkv("n_updates", update)
logger.logkv("total_timesteps", self.num_timesteps)
logger.logkv("fps", fps)
logger.logkv("explained_variance", float(explained_var))
#if len(self.ep_info_buf) > 0 and len(self.ep_info_buf[0]) > 0:
#logger.logkv('ep_reward_mean', safe_mean([ep_info['r'] for ep_info in self.ep_info_buf]))
#logger.logkv('ep_len_mean', safe_mean([ep_info['l'] for ep_info in self.ep_info_buf]))
logger.logkv("mean_episode_length",
episode_stats.mean_length())
logger.logkv("mean_episode_reward",
episode_stats.mean_reward())
logger.logkv('time_elapsed', t_start - t_first_start)
for (loss_val, loss_name) in zip(loss_vals,
self.loss_names):
logger.logkv(loss_name, loss_val)
logger.dumpkvs()
if self.verbose == 2 and (
update % log_interval == 0 or update
== 1) and episode_stats.mean_reward() > bestscore:
bestscore = episode_stats.mean_reward()
logger.logkv('time_elapsed', t_start - t_first_start)
logger.logkv("mean_episode_reward", bestscore)
logger.dumpkvs()
x.append(self.num_timesteps)
y.append(bestscore)
ax.plot(x, y, marker='.', color='b')
fig.canvas.draw()
ax.set_xlim(left=0, right=total_timesteps)
ax.set(title='Street Fighter 2 AI - PPO2 Algorithm',
ylabel='Fitness score',
xlabel='Timesteps')
fig.show()
plt.pause(0.001)
callback.on_training_end()
return self
def learn(policy,
env,
seed,
nsteps=5,
nstack=4,
total_timesteps=int(80e6),
vf_coef=0.5,
ent_coef=0.01,
max_grad_norm=0.5,
lr=7e-4,
lrschedule='linear',
epsilon=1e-5,
alpha=0.99,
gamma=0.99,
log_interval=100,
max_episode_length=None,
optimizer=None):
tf.reset_default_graph()
set_global_seeds(seed)
nenvs = env.num_envs
ob_space = env.observation_space
ac_space = env.action_space
num_procs = len(env.remotes) # HACK
model = Model(policy=policy,
ob_space=ob_space,
ac_space=ac_space,
nenvs=nenvs,
nsteps=nsteps,
nstack=nstack,
num_procs=num_procs,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
lr=lr,
alpha=alpha,
epsilon=epsilon,
total_timesteps=total_timesteps,
lrschedule=lrschedule,
optimizer=optimizer)
runner = Runner(env, model, nsteps=nsteps, nstack=nstack, gamma=gamma)
stats = EpisodeStats(nsteps, nenvs, maxlen=100)
nbatch = nenvs * nsteps
tstart = time.time()
for update in itertools.count():
obs, states, rewards, masks, actions, values = runner.run()
total_loss, policy_loss, value_loss, policy_entropy = model.train(
obs, states, rewards, masks, actions, values)
nseconds = time.time() - tstart
fps = int((update * nbatch) / nseconds)
stats.feed(rewards, masks)
if update % log_interval == 0 or update == 1:
ev = explained_variance(values, rewards)
logger.record_tabular("nupdates", update)
logger.record_tabular("total_timesteps", update * nbatch)
logger.record_tabular("fps", fps)
logger.record_tabular("policy_entropy", float(policy_entropy))
logger.record_tabular("value_loss", float(value_loss))
logger.record_tabular("total_loss", float(total_loss))
logger.record_tabular("explained_variance", float(ev))
logger.record_tabular("mean_episode_length", stats.mean_length())
logger.record_tabular("mean_episode_reward", stats.mean_reward())
logger.dump_tabular()
if max_episode_length and stats.mean_length(
) >= max_episode_length:
break
env.close()