def main(args):
env = gym.make('CartPole-v0')
env.seed(0)
ob_space = env.observation_space
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
D = Discriminator(env)
expert_observations = np.genfromtxt('trajectory/observations.csv')
expert_actions = np.genfromtxt('trajectory/actions.csv', dtype=np.int32)
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
reward = 0 # do NOT use rewards to update policy
success_num = 0
for iteration in range(args.iteration):
observations = []
actions = []
rewards = []
v_preds = []
run_policy_steps = 0
while True:
run_policy_steps += 1
obs = np.stack([obs]).astype(
dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
rewards.append(reward)
v_preds.append(v_pred)
next_obs, reward, done, info = env.step(act)
if done:
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value
obs = env.reset()
reward = -1
break
else:
obs = next_obs
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_length',
simple_value=run_policy_steps)
]), iteration)
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_reward',
simple_value=sum(rewards))
]), iteration)
if sum(rewards) >= 195:
success_num += 1
if success_num >= 100:
saver.save(sess, args.savedir + '/model.ckpt')
print('Clear!! Model saved.')
break
else:
success_num = 0
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations,
newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
# train discriminator
for i in range(2):
D.train(expert_s=expert_observations,
expert_a=expert_actions,
agent_s=observations,
agent_a=actions)
# output of this discriminator is reward
d_rewards = D.get_rewards(agent_s=observations, agent_a=actions)
d_rewards = np.reshape(d_rewards,
newshape=[-1]).astype(dtype=np.float32)
gaes = PPO.get_gaes(rewards=d_rewards,
v_preds=v_preds,
v_preds_next=v_preds_next)
gaes = np.array(gaes).astype(dtype=np.float32)
# gaes = (gaes - gaes.mean()) / gaes.std()
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
# train policy
inp = [observations, actions, gaes, d_rewards, v_preds_next]
PPO.assign_policy_parameters()
for epoch in range(6):
sample_indices = np.random.randint(
low=0, high=observations.shape[0],
size=32) # indices are in [low, high)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
#init directories
if not os.path.isdir(args.logdir):
os.mkdir(args.logdir)
if not os.path.isdir(args.logdir + '/' + args.env):
os.mkdir(args.logdir + '/' + args.env)
if not os.path.isdir(args.logdir + '/' + args.env + '/' + args.optimizer):
os.mkdir(args.logdir + '/' + args.env + '/' + args.optimizer)
if not os.path.isdir(args.logdir + '/' + args.env + '/' + args.optimizer + '/lr_' + str(args.lr)):
os.mkdir(args.logdir + '/' + args.env + '/' + args.optimizer + '/lr_' + str(args.lr))
args.logdir = args.logdir + '/' + args.env + '/' + args.optimizer + '/lr_' + str(args.lr)
if not os.path.isdir(args.savedir):
os.mkdir(args.savedir)
if not os.path.isdir(args.savedir + '/' + args.env):
os.mkdir(args.savedir + '/' + args.env)
if not os.path.isdir(args.savedir + '/' + args.env + '/' + args.optimizer):
os.mkdir(args.savedir + '/' + args.env + '/' + args.optimizer)
args.savedir = args.savedir + '/' + args.env + '/' + args.optimizer
#init classes
env = gym.make(args.env)
env.seed(0)
ob_space = env.observation_space
Policy = Policy_net('policy', env, args.env)
Old_Policy = Policy_net('old_policy', env, args.env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma, _optimizer=args.optimizer, _lr=args.lr)
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
reward = 0
success_num = 0
for iteration in range(args.iteration):
observations = []
actions = []
v_preds = []
rewards = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.stack([obs]).astype(dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
next_obs, reward, done, info = env.step(act)
if done:
v_preds_next = v_preds[1:] + [0] # next state of terminate state has 0 state value
obs = env.reset()
reward = -1
break
else:
obs = next_obs
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_length', simple_value=episode_length)])
, iteration)
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_reward', simple_value=sum(rewards))])
, iteration)
if iteration == (args.iteration-1):
saver.save(sess, args.savedir+'/model'+str(args.lr)+'.ckpt')
print('Clear!! Model saved.')
break
gaes = PPO.get_gaes(rewards=rewards, v_preds=v_preds, v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations, newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, rewards, v_preds_next]
print('iteration:', iteration, ',rewards:', sum(rewards))
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0, high=observations.shape[0], size=32)
sampled_inp = [np.take(a=a, indices=sample_indices, axis=0) for a in inp] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
#init directories
if not os.path.isdir(args.logdir):
os.mkdir(args.logdir)
if not os.path.isdir(args.logdir + '/' + args.env):
os.mkdir(args.logdir + '/' + args.env)
if not os.path.isdir(args.logdir + '/' + args.env + '/' + args.optimizer):
os.mkdir(args.logdir + '/' + args.env + '/' + args.optimizer)
args.logdir = args.logdir + '/' + args.env + '/' + args.optimizer
if not os.path.isdir(args.savedir):
os.mkdir(args.savedir)
if not os.path.isdir(args.savedir + '/' + args.env):
os.mkdir(args.savedir + '/' + args.env)
if not os.path.isdir(args.savedir + '/' + args.env + '/' + args.optimizer):
os.mkdir(args.savedir + '/' + args.env + '/' + args.optimizer)
args.savedir = args.savedir + '/' + args.env + '/' + args.optimizer
#init classes
env = gym.make(args.env)
env.seed(0)
ob_space = env.observation_space
Policy = Policy_net('policy', env, args.env)
Old_Policy = Policy_net('old_policy', env, args.env)
PPO = PPOTrain(Policy,
Old_Policy,
gamma=args.gamma,
_optimizer=args.optimizer,
_lr=args.lr)
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
reward = 0
success_num = 0
for iteration in range(args.iteration):
observations = []
actions = []
v_preds = []
rewards = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.stack([obs]).astype(
dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
next_obs, reward, done, info = env.step(act)
if done:
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value
obs = env.reset()
reward = -1
break
else:
obs = next_obs
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_length',
simple_value=episode_length)
]), iteration)
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_reward',
simple_value=sum(rewards))
]), iteration)
if iteration == (args.iteration - 1):
saver.save(sess, args.savedir + '/model.ckpt')
print('Clear!! Model saved.')
break
gaes = PPO.get_gaes(rewards=rewards,
v_preds=v_preds,
v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations,
newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, rewards, v_preds_next]
print('iteration:', iteration, ',rewards:', sum(rewards))
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0,
high=observations.shape[0],
size=32)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
print(date)
energyPolicy_training_data.append("Energy poilcy training")
energyPolicy_training_data.append(
"Date: " +
str(date))
energyPolicy_training_data.append(
"Noise type: " +
str(args.noise_type))
energyPolicy_training_data.append(
"Policy Training max episodes: " +
str(args.iteration))
energyPolicy_training_data.append(
"Number of iterations the energy model have ben trained: " +
str(args.model))
energyPolicy_training_data.append(
"PPO gamma: " +
str(args.gamma))
energyPolicy_training_data.append(
"Do we add noise to sapair for calculating energy " +
str(args.sanoise))
energyPolicy_training_data.append(
"The noise we add to sapair " +
str(args.noise_sigma))
energyPolicy_training_data.append(
"h(energy) " +
str(args.reward_function))
energyPolicy_training_data.append(" \n\n")
env = gym.make('CartPole-v0')
Energy = Energy_net('energy', 'CartPole-v0')
energy_saver = tf.train.Saver()
sapairs = np.genfromtxt('training_data/sapairs.csv')
noise_sapairs = np.genfromtxt('training_data/noise_sapairs.csv')
with tf.Session() as sess:
# writer = tf.summary.FileWriter(args.logdir+'/'+args.alg, sess.graph)
sess.run(tf.global_variables_initializer())
if args.model == '':
energy_saver.restore(
sess, args.modeldir + '/' + args.alg + '/' + args.noise_type +
'/' + 'model.ckpt')
else:
energy_saver.restore(
sess, args.modeldir + '/' + args.alg + '/' + args.noise_type +
'/' + 'model.ckpt-' + args.model)
print("As for model after ", args.model, "training iterations")
print("Energy for expert sapairs looks like:",
Energy.get_energy(sapairs))
print(
"Energy for noise sapairs (not corresponding to the noise trained for Energy) looks like:",
Energy.get_energy(noise_sapairs))
energyPolicy_training_data.append(
["As for model after ", args.model, "training iterations"])
energyPolicy_training_data.append(
"Energy for expert sapairs looks like:\n" +
str(Energy.get_energy(sapairs)))
energyPolicy_training_data.append(
"Energy for noise sapairs (not corresponding to the noise trained for Energy) looks like:\n"
+ str(Energy.get_energy(noise_sapairs)))
energyPolicy_training_data.append(" \n\n\n\n\n\n\n\n\n")
energyPolicy_training_data.append(
"Done with reloading Energy. Start RL")
# writer.close()
open_file_and_save(
args.logdir + '/' + args.model + "_iter_" + args.noise_type +
'_Policy' + date, energyPolicy_training_data)
print("Done with reloading Energy. Start RL")
# Start RL
env.seed(0)
ob_space = env.observation_space
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
saver = tf.train.Saver()
# writer = tf.summary.FileWriter(args.logdir+'/'+args.noise_type, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
reward = 0
alter_reward = 0
success_num = 0
render = False
#ep_reward = []
# 用于记录每个trajectory的数据最后做总结
Summary_after_max_episodes_training = []
Trajectory_rewards = []
Trajectory_alter_rewards = []
Trajectory_success_num = 0 # 与success_num一样，只不过这个不会清零，这个用于评估这个energy对于训练的效果
plot_rewards = []
plot_alter_rewards = []
plot_iteration = []
for iteration in range(args.iteration):
observations = []
actions = []
v_preds = []
rewards = []
alter_rewards = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.stack([obs]).astype(
dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
alter_rewards.append(alter_reward)
rewards.append(reward)
next_obs, reward, done, info = env.step(act)
# alter reward
sapair = np.append(obs, np.array([[act]]), axis=1)
# print("sapair:",sapair)
energy = Energy.get_energy(sapair)[0][0]
print("Energy for this sapair", energy)
if args.sanoise == True:
# 定义 gauss noise 的均值和方差
mu, sigma = 0, args.noise_sigma
# 一维guass
# saNumber = sapairs.shape[0]
saShape = sapair.shape[1]
# sampleNo = saNumber * saShape # 采样sampleNo个gauss noise
noise = np.random.normal(mu, sigma, saShape)
noise_sapair = sapair + noise
print("noise_sapair:", noise_sapair)
# noise_sapairs = np.reshape(noise_sapairs, newshape=[saNumber, saShape])
noise_energy = Energy.get_energy(noise_sapair)[0][0]
print("Noise Energy for this sapair", noise_energy)
energy = noise_energy
if args.reward_function == "-energy":
alter_reward = -energy
elif args.reward_function == "-energy+1":
alter_reward = -energy + 1
elif args.reward_function == "exp(-energy-1)":
alter_reward = np.exp(-energy - 1)
elif args.reward_function == "exp(-energy)":
alter_reward = np.exp(-energy)
else:
print("No such reward_function")
#alter_reward = np.exp(-energy-1)
#alter_reward = -energy+1
#alter_reward = reward
#alter_reward = -energy
# if render:
# env.render()
# pass
if done:
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value
obs = env.reset()
reward = -1
alter_reward = -1
break
else:
obs = next_obs
# writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_length', simple_value=episode_length)]), iteration)
# writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_reward', simple_value=sum(rewards))]), iteration)
# if sum(rewards) >= 195:
# success_num += 1
# Trajectory_success_num +=1
# render = True
# if success_num >= 100:
# saver.save(sess, args.savedir + '/model.ckpt')
# print('Clear!! Model saved.')
# break
# else:
# success_num = 0
sum_rewards = sum(rewards)
sum_alter_rewards = sum(alter_rewards)
Trajectory_rewards.append(sum_rewards)
Trajectory_alter_rewards.append(sum_alter_rewards)
#画图
plot_rewards.append(sum_rewards)
plot_alter_rewards.append(sum_alter_rewards)
plot_iteration.append(iteration)
#ep_reward.append(sum(rewards))
# print("Sample done in one traj.")
energyPolicy_training_data_for_this_episode = []
energyPolicy_training_data_for_this_episode.append(" ")
energyPolicy_training_data_for_this_episode.append(
"Trajectory: " + str(iteration))
energyPolicy_training_data_for_this_episode.append(
"episode_len: " + str(episode_length))
energyPolicy_training_data_for_this_episode.append(
"True rewards: " + str(sum_rewards))
energyPolicy_training_data_for_this_episode.append(
"alter_rewards: " + str(sum_alter_rewards))
open_file_and_save(
args.logdir + '/' + args.model + "_iter_" + args.noise_type +
'_Policy' + date, energyPolicy_training_data_for_this_episode)
print()
print("Trajectory", iteration, ":")
print("episode_len: ", episode_length)
print("rewards: ", sum(rewards))
print("alter_rewards: ", sum(alter_rewards))
# gaes = PPO.get_gaes(rewards=rewards, v_preds=v_preds, v_preds_next=v_preds_next)
gaes = PPO.get_gaes(rewards=alter_rewards,
v_preds=v_preds,
v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations,
newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
alter_rewards = np.array(alter_rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, alter_rewards, v_preds_next]
# inp = [observations, actions, gaes, rewards, v_preds_next]
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0,
high=observations.shape[0],
size=32)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
# writer.add_summary(summary, iteration)
# writer.close()
#开始画图
plt.title('Noise:' + str(args.sanoise))
plt.plot(plot_iteration,
plot_rewards,
color='red',
label='True_rewards')
plt.plot(plot_iteration,
plot_alter_rewards,
color='green',
label='alter_rewards')
plt.legend() #显示图例
plt.xlabel('Episodes')
plt.ylabel('Rewards')
plt.show()
def main(args):
# init directories
if not os.path.isdir(args.logdir):
os.mkdir(args.logdir)
if not os.path.isdir(args.logdir + '/' + args.env):
os.mkdir(args.logdir + '/' + args.env)
if not os.path.isdir(args.logdir + '/' + args.env + '/' + args.optimizer):
os.mkdir(args.logdir + '/' + args.env + '/' + args.optimizer)
args.logdir = args.logdir + '/' + args.env + '/' + args.optimizer
if not os.path.isdir(args.savedir):
os.mkdir(args.savedir)
if not os.path.isdir(args.savedir + '/' + args.env):
os.mkdir(args.savedir + '/' + args.env)
if not os.path.isdir(args.savedir + '/' + args.env + '/' + args.optimizer):
os.mkdir(args.savedir + '/' + args.env + '/' + args.optimizer)
args.savedir = args.savedir + '/' + args.env + '/' + args.optimizer
args.tradir = args.tradir + '/' + args.env + '/' + args.optimizer
# init classes
env = gym.make(args.env)
env.seed(0)
ob_space = env.observation_space
Policy = Policy_net('policy', env, args.env)
Old_Policy = Policy_net('old_policy', env, args.env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma, _optimizer=args.optimizer)
D = Discriminator(env, args.env, _optimizer=args.optimizer)
expert_observations = np.genfromtxt(args.tradir + '/observations.csv')
expert_actions = np.genfromtxt(args.tradir + '/actions.csv', dtype=np.int32)
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
reward = 0 # do NOT use rewards to update policy
success_num = 0
for iteration in range(args.iteration):
observations = []
actions = []
rewards = []
v_preds = []
run_policy_steps = 0
while True:
run_policy_steps += 1
obs = np.stack([obs]).astype(dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
rewards.append(reward)
v_preds.append(v_pred)
next_obs, reward, done, info = env.step(act)
if done:
v_preds_next = v_preds[1:] + [0] # next state of terminate state has 0 state value
obs = env.reset()
reward = -1
break
else:
obs = next_obs
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_length', simple_value=run_policy_steps)])
, iteration)
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_reward', simple_value=sum(rewards))])
, iteration)
print('iteration:', iteration, ',rewards:', sum(rewards))
if iteration == (args.iteration - 1):
saver.save(sess, args.savedir + '/model.ckpt')
print('Clear!! Model saved.')
break
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations, newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
# train discriminator
for i in range(2):
D.train(expert_s=expert_observations,
expert_a=expert_actions,
agent_s=observations,
agent_a=actions)
# output of this discriminator is reward
d_rewards = D.get_rewards(agent_s=observations, agent_a=actions)
d_rewards = np.reshape(d_rewards, newshape=[-1]).astype(dtype=np.float32)
gaes = PPO.get_gaes(rewards=d_rewards, v_preds=v_preds, v_preds_next=v_preds_next)
gaes = np.array(gaes).astype(dtype=np.float32)
# gaes = (gaes - gaes.mean()) / gaes.std()
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
# train policy
inp = [observations, actions, gaes, d_rewards, v_preds_next]
PPO.assign_policy_parameters()
for epoch in range(6):
sample_indices = np.random.randint(low=0, high=observations.shape[0],
size=32) # indices are in [low, high)
sampled_inp = [np.take(a=a, indices=sample_indices, axis=0) for a in inp] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
# env = gym.make('CartPole-v0')
# env.seed(0)
env = CustomEnv()
ob_space = env.observation_space
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy, Old_Policy, gamma=gamma)
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(log_path, sess.graph)
sess.run(tf.global_variables_initializer())
obs, acs, target_video = env.reset()
success_num = 0
for iteration in range(iterations):
observations = []
actions = []
pred_actions = []
rewards = []
v_preds = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.array([obs]).astype(dtype=np.float32) # prepare to feed placeholder Policy.obs
acs = np.array([acs]).astype(dtype=np.float32)
pred_act, v_pred = Policy.act(obs=obs, acs=acs, stochastic=True)
# act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
next_obs, reward, done, info = env.step(acs)
observations.append(obs)
actions.append(acs)
pred_actions.append(pred_act)
rewards.append(reward)
v_preds.append(v_pred)
if done:
next_obs = np.stack([next_obs]).astype(dtype=np.float32) # prepare to feed placeholder Policy.obs
_, v_pred = Policy.act(obs=next_obs, stochastic=True)
v_preds_next = v_preds[1:] + [np.asscalar(v_pred)]
obs, acs, target_video = env.reset()
break
else:
obs = next_obs
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_length', simple_value=episode_length)])
, iteration)
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_reward', simple_value=sum(rewards))])
, iteration)
if sum(rewards) >= 195:
success_num += 1
if success_num >= 100:
saver.save(sess, weight_path + '/model.ckpt')
print('Clear!! Model saved.')
break
else:
success_num = 0
gaes = PPO.get_gaes(rewards=rewards, v_preds=v_preds, v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
# observations = np.reshape(observations, newshape=[-1,] + list(ob_space.shape))
observations = np.array(observations).astype(dtype=np.float32)
actions = np.array(actions).astype(dtype=np.float32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, rewards, v_preds_next]
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0, high=observations.shape[0], size=32)
sampled_inp = [np.take(a=a, indices=sample_indices, axis=0) for a in inp] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
scene_scope = 'bathroom_02'
task_scope = 26 #26 43 53 32 41
env = Environment({
'scene_name': scene_scope,
'terminal_state_id': int(task_scope)
})
S_Class = SIAMESE() #Creating a siamese class -object
Policy = Policy_net(
'policy', S_Class) #buiding the actor critic graph / object , Passing
Old_Policy = Policy_net('old_policy',
S_Class) #same thing as the other PPO
PPO = PPOTrain(Policy, Old_Policy,
gamma=args.gamma) #gradiet updatror object or the graph
D = Discriminator(S_Class) #discriminator of the Gan Kind of thing
'''
batch_n=tf.get_collection(tf.GraphKeys.UPDATE_OPS)
var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Siamese')
'''
#Loading Expert Data State/Tragets etc
expert_observations = np.genfromtxt(
'trajectory/observations.csv') #load expert demnetrations
expert_targets = np.genfromtxt('trajectory/targets.csv')
expert_actions = np.genfromtxt('trajectory/actions.csv', dtype=np.int32)
expert_observations = np.reshape(expert_observations,
newshape=[-1, 2048, 4])
expert_targets = np.reshape(expert_targets, newshape=[-1, 2048, 4])
saver = tf.train.Saver(
) #Assign another save if you want to use BC weights
if args.restore: #We need a seperate saver only for assigning paramters from BC trained thing
saver2 = tf.tran.Saver([
tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='policy'),
tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='Siamese')
])
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(
tf.global_variables_initializer()
) #here already variables get intialized both old policy and new policy net
if args.restore:
if args.model == '':
saver2.restore(
sess,
args.modeldir + '/' + args.alg + '/' + 'shamane.ckpt')
print("Model Reastored")
else:
saver.restore(
sess, args.modeldir + '/' + args.alg + '/' +
'model.ckpt-' + args.model)
success_num = 0 #This is use to check whether my agent went to the terminal point
#var_list=tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
for iteration in range(
100000): #args.iteration):#Here start the adversial training
print(
"Starting ........ The Iteration---------------------------------------------------- :",
iteration)
observations = []
actions = []
#rewards = []
targets = [] #for the gail
v_preds = []
run_policy_steps = 0
while (
True
): #Here what is happenning is , this again samples trajectories from untrain agent
run_policy_steps += 1
obs = np.stack([env.s_t]).astype(
dtype=np.float32
) # prepare to feed placeholder Policy.obs #Initial observation
target = np.stack([env.s_target]).astype(
dtype=np.float32
) #This is to make sure that input is [batch_size,2048,4]
act, v_pred, prob = Policy.act(
state=obs, target=target,
stochastic=True) # Agents action and values
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs) #save the set of observations
targets.append(target)
actions.append(act) #save the set of actions
v_preds.append(v_pred)
#next_obs, reward, done, info = env.step(act) #get the next observation and reward acording to the observation
next_obs, is_terminal, is_collided = env.step(act)
if is_terminal:
success_num = success_num + 1
print(
"Congratz yoy just reach the terminal state which is:",
env.terminal_state_id)
if is_collided:
print(
"Bad Luck your agent just collided couldn't made it to the terminal state which is :",
env.terminal_state_id)
if (is_terminal or is_collided
or (run_policy_steps
== 100)): #run one episode till the termination
print("Number Of Exploration by the AGENT:",
run_policy_steps)
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value #this list use to update the parameters of the calue net
print(
"Environment is resetting after the collition/Terminal"
)
obs = env.reset()
#reward = -1
break #with tihs vreak all obsercation ,action and other lists get empty
#print(sum(rewards))
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_length',
simple_value=run_policy_steps)
]), iteration)
#writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_reward', simple_value=sum(rewards))])
#, iteration)
if success_num >= 5000:
saver.save(sess, args.savedir + '/model.ckpt')
print('Clear!! Model saved.')
break
#else:
#success_num = 0
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations,
newshape=[-1, 2048,
4]) #collect observations
targets = np.reshape(targets, newshape=[-1, 2048, 4])
actions = np.array(actions).astype(
dtype=np.int32) #collect the actions
# train discriminator #Here comes the Discriminator !!
Dis_input = [
expert_observations, expert_targets, expert_actions,
observations, targets, actions
]
observations.shape[0]
expert_observations.shape[0]
if observations.shape[0] < expert_observations.shape[0]:
High = observations.shape[0]
else:
High = expert_observations.shape[0]
for i in range(100):
sample_indices = np.random.randint(low=0, high=High, size=32)
sampled_inp_D = [
np.take(a=a, indices=sample_indices, axis=0)
for a in Dis_input
]
D.train(expert_s=sampled_inp_D[0],
expert_t=sampled_inp_D[1],
expert_a=sampled_inp_D[2],
agent_s=sampled_inp_D[3],
agent_t=sampled_inp_D[4],
agent_a=sampled_inp_D[5])
'''
D.train(expert_s=expert_observations,
expert_t=expert_targets,
expert_a=expert_actions,
agent_s=observations,
agent_t=targets,
agent_a=actions)
'''
#To get rewards we can use a RNN , then we can get the each time unit output to collect the reward function
d_rewards = D.get_rewards(
agent_s=observations, agent_t=targets, agent_a=actions
) #how well our agent performed with respect to the expert
d_rewards = np.reshape(d_rewards, newshape=[-1]).astype(
dtype=np.float32) #rewards for each action pair
gaes = PPO.get_gaes(
rewards=d_rewards, v_preds=v_preds, v_preds_next=v_preds_next
) #this to calcuate the advantage function in PPO
gaes = np.array(gaes).astype(dtype=np.float32)
# gaes = (gaes - gaes.mean()) / gaes.std()
v_preds_next = np.array(v_preds_next).astype(
dtype=np.float32) #This is the next value function
#train policy
inp = [
observations, targets, actions, gaes, d_rewards, v_preds_next
]
PPO.assign_policy_parameters(
) #Assigning policy params means assigning the weights to the default policy nets
for epoch in range(
100
): #This is to train the Agent (Actor Critic ) from the obtaiend agent performances and already trained discriminator
sample_indices = np.random.randint(
low=0, high=observations.shape[0],
size=32) # indices are in [low, high)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # Here trainign the policy network
PPO.train(state=sampled_inp[0],
targets=sampled_inp[1],
actions=sampled_inp[2],
gaes=sampled_inp[3],
rewards=sampled_inp[4],
v_preds_next=sampled_inp[5])
summary = PPO.get_summary(obs=inp[0],
target=inp[1],
actions=inp[2],
gaes=inp[3],
rewards=inp[4],
v_preds_next=inp[5])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
# prepare log dir
if not os.path.exists(args.logdir):
os.makedirs(args.logdir)
if not os.path.exists(args.savedir):
os.makedirs(args.savedir)
# gym環境作成
env = gym.make("CartPole-v0")
env.seed(0)
ob_space = env.observation_space
# policy net
Policy = Policy_net("policy", env)
Old_Policy = Policy_net("old_policy", env)
# ppo学習インスタンス
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
# tensorflow saver
saver = tf.train.Saver()
# session config
config = tf.ConfigProto(
gpu_options=tf.GPUOptions(visible_device_list=args.gpu_num, allow_growth=True)
)
# start session
with tf.Session(config=config) as sess:
# summary writer
writer = tf.summary.FileWriter(args.logdir, sess.graph)
# Sessionの初期化
sess.run(tf.global_variables_initializer())
# 状態の初期化
obs = env.reset()
# episodeの成功回数
success_num = 0
# episode loop
for iteration in tqdm(range(args.iteration)):
# episodeのtrajectory配列
# buffer
observations = []
actions = []
v_preds = []
rewards = []
# episodeのstep回数
episode_length = 0
# run episode
while True:
episode_length += 1
# プレースホルダー用に変換
obs = np.stack([obs]).astype(dtype=np.float32)
# 行動と状態価値を推定
act, v_pred = Policy.act(obs=obs, stochastic=True)
# 要素数が1の配列をスカラーに変換
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
# policyによる行動で状態を更新
next_obs, reward, done, info = env.step(act)
# episodeの各変数を追加
# (s_t, a_t, v_t, r_t)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
# episode終了判定
# episodeが終了していたら次のepisodeを開始
if done:
# v_t+1の配列
v_preds_next = v_preds[1:] + [0]
obs = env.reset()
reward = -1
break
else:
obs = next_obs
# summary追加
writer.add_summary(
tf.Summary(
value=[
tf.Summary.Value(
tag="episode_length", simple_value=episode_length
)
]
),
iteration,
)
writer.add_summary(
tf.Summary(
value=[
tf.Summary.Value(
tag="episode_reward", simple_value=sum(rewards)
)
]
),
iteration,
)
# episode成功判定
if sum(rewards) >= 195:
success_num += 1
# 連続で100回成功していればepisode loopを終了
if success_num >= 100:
saver.save(sess, args.savedir + "/model.ckpt")
print("Clear!! Model saved.")
break
else:
success_num = 0
# policy netによるtrajectryをプレースホルダー用に変換
observations = np.reshape(
observations, newshape=[-1] + list(ob_space.shape)
)
actions = np.array(actions).astype(dtype=np.int32)
# rewardsをプレースホルダー用に変換
rewards = np.array(rewards).astype(dtype=np.float32)
# gaesの取得
gaes = PPO.get_gaes(
rewards=rewards, v_preds=v_preds, v_preds_next=v_preds_next
)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
# エージェントのexperience
inp = [observations, actions, gaes, rewards, v_preds_next]
# Old_Policyにパラメータを代入
PPO.assign_policy_parameters()
# PPOの学習
for epoch in range(6):
# 学習データサンプル用のインデックスを取得
sample_indices = np.random.randint(
low=0, high=observations.shape[0], size=32
)
# PPO学習データをサンプル
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
]
PPO.train(
obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4],
)
# summaryの取得
summary = PPO.get_summary(
obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4],
)
writer.add_summary(summary, iteration)
writer.close()
def main(args):
writer = SummaryWriter(args.logdir)
logger = ResultLogger(writer)
env = Environment() # 自定义环境
ob_space = env.observation_space
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy,
Old_Policy,
gamma=args.gamma,
args=args,
logger=logger)
saver = tf.train.Saver()
if args.continue_train:
tf.reset_default_graph()
tf.train.import_meta_graph(args.continue_meta)
with tf.Session() as sess:
if args.continue_train:
saver.restore(sess, args.continue_modeldir)
sess.run(tf.global_variables_initializer())
reward = 0
winnum = 0
drawnum = 0
for episode in range(args.episode):
observations = []
actions = []
v_preds = []
rewards = []
run_policy_steps = 0
total_reward = 0
obs = env.reset()
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
run_policy_steps += 1
obs = np.stack([obs]).astype(
dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
next_obs, reward, sparse_rew, done, info = env.step(act)
if reward < -1000:
reward = -10
reward = utils.get_curriculum_reward(reward, sparse_rew, 1.0,
run_policy_steps)
# if episode==1:
# print(reward)
obs = next_obs
if done:
total_reward = sum(rewards)
total_reward /= run_policy_steps
total_reward += reward
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value
reward = -1
if info == 3:
winnum += 1
if info == 2:
drawnum += 1
break
if episode % 100 == 0:
winnum = 0
drawnum = 0
logger.log_result(total_reward, winnum, drawnum, episode)
print(episode, total_reward)
if episode % 1000 == 0:
saver.save(sess, args.savedir + '/model.ckpt')
####
## GAE
####
gaes = PPO.get_gaes(rewards=rewards,
v_preds=v_preds,
v_preds_next=v_preds_next)
# 把list 转成 适应于tf.placeholder 的numpy array
observations = np.reshape(observations, newshape=(-1, ob_space))
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
logger.log_gaes(gaes.mean(), episode)
PPO.log_parameter(observations, actions, gaes, rewards,
v_preds_next)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, rewards, v_preds_next]
# train
for epoch in range(2):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0,
high=observations.shape[0],
size=32)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
def main(args):
#env.seed(0)
env = gym.make('MineRLNavigateDense-v0')
ob_space = env.observation_space
action_space = env.action_space
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
reward = 0
success_num = 0
render = False
for iteration in range(args.iteration):
observations = []
actions = []
v_preds = []
rewards = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.stack([obs['pov']]).astype(dtype=np.float32)
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs['pov'])
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
next_obs, reward, done, info = env.step(
[int(act / 3) + 1, act - int(act / 3) * 3])
if (episode_length % 2500 == 0):
print(sum(rewards))
if render:
env.render()
if done:
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value
obs = env.reset()
print('done')
break
else:
obs = next_obs
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_length',
simple_value=episode_length)
]), iteration)
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_reward',
simple_value=sum(rewards))
]), iteration)
if sum(rewards) >= 1:
success_num += 1
render = True
if success_num >= 10:
saver.save(sess, args.savedir + '/model.ckpt')
print('Clear!! Model saved.')
break
gaes = PPO.get_gaes(rewards=rewards,
v_preds=v_preds,
v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations, newshape=[-1, 64, 64, 3])
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, rewards, v_preds_next]
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0,
high=observations.shape[0],
size=32)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
env = gym.make('CartPole-v0')
BCPolicy = Policy_net('bcpolicy', env)
BC = BehavioralCloning(BCPolicy)
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
saver = tf.train.Saver(max_to_keep=args.max_to_keep) #实例化一个Saver对象，在训练过程中，定期调用saver.save方法，像文件夹中写入包含当前模型中所有可训练变量的checkpoint文件 saver.save(sess,FLAGG.train_dir,global_step=step)
exp_obs = np.genfromtxt('trajectory/observations.csv')[0:exp_len] #exp_len=200
exp_acts = np.genfromtxt('trajectory/actions.csv', dtype=np.int32)[0:exp_len]
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph) #指定一个文件用来保存图。格式：tf.summary.FileWritter(path,sess.graph)，可以调用其add_summary（）方法将训练过程数据保存在filewriter指定的文件中
sess.run(tf.global_variables_initializer())
inp = [exp_obs, exp_acts] #inp[0]就是observations， inp[1]就是actoins
for iteration in range(args.iteration): # episode
# train
for epoch in range(args.epoch_num):
# select sample indices in [low, high)
sample_indices = np.random.randint(low=0, high=exp_obs.shape[0], size=args.minibatch_size) #函数的作用是，返回一个随机整型数，范围从低（包括）到高（不包括），即[low, high)
sampled_inp = [np.take(a=a, indices=sample_indices, axis=0) for a in inp] # sample training data
BC.train(obs=sampled_inp[0], actions=sampled_inp[1])
bc_summary = BC.get_summary(obs=inp[0], actions=inp[1])
if (iteration+1) % args.interval == 0:
saver.save(sess, args.savedir + '/model.ckpt', global_step=iteration+1)
writer.add_summary(bc_summary, iteration)
print("Done with BC. Start RL")
# Start RL
obs = env.reset()
ob_space = env.observation_space
reward = 0
alter_reward = 0
success_num = 0
render = False
ep_reward=[]
for iteration in range(5*args.iteration):
print("iter:{}".format(iteration))
observations = []
actions = []
v_preds = []
rewards = []
alter_rewards = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.stack([obs]).astype(dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
alter_rewards.append(alter_reward)
rewards.append(reward)
next_obs, reward, done, info = env.step(act)
alter_reward = np.log(1/(kl_divergence(obs, BCPolicy, Policy)+0.00001))
#alter_reward = -kl_divergence(obs, BCPolicy, Policy)
#alter_reward = kl_divergence(obs, BCPolicy, Policy)
#print(alter_reward)
if render:
#env.render()
pass
if done:
v_preds_next = v_preds[1:] + [0] # next state of terminate state has 0 state value
obs = env.reset()
reward = -1
alter_reward = -1
print("episode_len: ",episode_length)
break
else:
obs = next_obs
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_length', simple_value=episode_length)])
, iteration)
writer.add_summary(tf.Summary(value=[tf.Summary.Value(tag='episode_reward', simple_value=sum(rewards))])
, iteration)
if sum(rewards) >= 195:
success_num += 1
render = True
if success_num >= 100:
saver.save(sess, args.savedir+'/model.ckpt')
print('Clear!! Model saved.')
break
else:
success_num = 0
ep_reward.append(sum(rewards))
print("rewards: ",sum(rewards))
print("alter_rewards: ",sum(alter_rewards))
print("Sample done in one traj.")
gaes = PPO.get_gaes(rewards=alter_rewards, v_preds=v_preds, v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations, newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
alter_rewards = np.array(alter_rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, alter_rewards, v_preds_next]
print("Begin Training")
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0, high=observations.shape[0], size=32)
sampled_inp = [np.take(a=a, indices=sample_indices, axis=0) for a in inp]
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
"""
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
"""
#writer.add_summary(summary, iteration)
writer.close()
plt.plot(ep_reward)
def main(args):
env = gym.make('CartPole-v0')
env.seed(0)
ob_space = env.observation_space #This is the environment for the gym to observe
Policy = Policy_net(
'policy', env) #take the environments #this is normal policy class
Old_Policy = Policy_net('old_policy', env) #this is for the old policy
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma) #this is for training
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter(args.logdir, sess.graph)
sess.run(tf.global_variables_initializer()
) #Here all the variabls get initialized
obs = env.reset(
) # [position of cart, velocity of cart, angle of pole, rotation rate of pole] Initial observation
reward = 0
success_num = 0
for iteration in range(args.iteration):
observations = [] #to store observations
actions = []
v_preds = []
rewards = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length #Starting to run the
episode_length += 1 #episode length is something dynamic
obs = np.stack([obs]).astype(
dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(
obs=obs, stochastic=True
) #get the action and value prediction (actor and critic network output)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
next_obs, reward, done, info = env.step(
act) #get the observation from the environments
#The episode ends when the pole is more than 15 degrees from vertical, or the cart moves more than 2.4 units from the center. That is done
if done: #This is a termination stage
#this has all the next state eliements of the episode inputs
v_preds_next = v_preds[1:] + [
0
] # next state of terminate state has 0 state value #after the terminal stage there shouldn;t be a value function
obs = env.reset()
reward = -1
break
else: #here your break the episode
obs = next_obs #if the system do not get terminated it will run for ever
#After a one episode get terminated
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_length',
simple_value=episode_length)
]) #From this we can learn how long the episode went
,
iteration)
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_reward',
simple_value=sum(rewards))
]) #
,
iteration)
if sum(rewards) >= 195:
success_num += 1
if success_num >= 100:
saver.save(sess, args.savedir + '/model.ckpt')
print(
'Clear!! Model saved.'
) #this is like after this much sucessfull attempts we are confident about the model
break
else:
success_num = 0
gaes = PPO.get_gaes(
rewards=rewards, v_preds=v_preds,
v_preds_next=v_preds_next) #this is the advantage function
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(
observations, newshape=[-1] +
list(ob_space.shape)) #observations from the current policy
actions = np.array(actions).astype(
dtype=np.int32) #actions taken from current policy
gaes = np.array(gaes).astype(
dtype=np.float32) #generalized advantage enstimation
gaes = (gaes - gaes.mean()) / gaes.std() #Normalize it
rewards = np.array(rewards).astype(
dtype=np.float32) #Extracted rewrds
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters(
) #before updating the new policy we assign current policy parameters to old policy
inp = [observations, actions, gaes, rewards, v_preds_next]
# train
for epoch in range(6): #starting the optimization
# sample indices from [low, high)
sample_indices = np.random.randint(low=0,
high=observations.shape[0],
size=32)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data Randomly take one sample from the training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()
def main(args):
# prepare log dir
if not os.path.exists(args.logdir):
os.makedirs(args.logdir)
if not os.path.exists(args.savedir):
os.makedirs(args.savedir)
# gym環境作成
env = gym.make('CartPole-v0')
env.seed(0)
ob_space = env.observation_space
# policy net
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
# ppo学習インスタンス
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
# discriminator
D = Discriminator(env)
# エキスパートtrajectory読み込み
expert_observations = np.genfromtxt('trajectory/observations.csv')
expert_actions = np.genfromtxt('trajectory/actions.csv', dtype=np.int32)
# tensorflow saver
saver = tf.train.Saver()
# session config
config = tf.ConfigProto(
gpu_options=tf.GPUOptions(
visible_device_list=args.gpu_num,
allow_growth=True
))
# start session
with tf.Session(config=config) as sess:
# summary writer
writer = tf.summary.FileWriter(args.logdir, sess.graph)
# Sessionの初期化
sess.run(tf.global_variables_initializer())
# 状態の初期化
obs = env.reset()
success_num = 0
# episode loop
for iteration in tqdm(range(args.iteration)):
# buffer
observations = []
actions = []
rewards = []
v_preds = []
run_policy_steps = 0
# run episode
while True:
run_policy_steps += 1
# ネットワーク入力用にobsを変換
obs = np.stack([obs]).astype(dtype=np.float32)
# 行動と価値を推定
act, v_pred = Policy.act(obs=obs, stochastic=True)
# 要素数が1の配列をスカラーに変換
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
# policy netの推定行動で状態の更新
next_obs, reward, done, info = env.step(act)
# episodeの各変数を追加
observations.append(obs)
actions.append(act)
v_preds.append(v_pred)
rewards.append(reward)
# episode終了判定
if done:
v_preds_next = v_preds[1:] + [0]
obs = env.reset()
reward = -1
break
else:
obs = next_obs
# summary追加
writer.add_summary(
tf.Summary(value=[tf.Summary.Value(
tag='episode_length',
simple_value=run_policy_steps)]),
iteration)
writer.add_summary(
tf.Summary(value=[tf.Summary.Value(
tag='episode_reward',
simple_value=sum(rewards))]),
iteration)
# episode成功判定
if sum(rewards) >= 195:
success_num += 1
# 連続で100回成功していればepisode loopを終了
if success_num >= 100:
saver.save(sess, args.savedir+'/model.ckpt')
print('Clear!! Model saved.')
break
else:
success_num = 0
# policy netによるtrajectryをプレースホルダー用に変換
observations = np.reshape(observations, newshape=[-1] + list(ob_space.shape))
actions = np.array(actions).astype(dtype=np.int32)
###########################
# GAILの変更点はここだけ
# discriminatorでエキスパートの報酬に近づける
# discriminator学習 2回
for i in range(2):
D.train(expert_s=expert_observations,
expert_a=expert_actions,
agent_s=observations,
agent_a=actions)
# get d_rewards from discrminator
d_rewards = D.get_rewards(agent_s=observations, agent_a=actions)
# transform d_rewards to numpy for placeholder
d_rewards = np.reshape(d_rewards, newshape=[-1]).astype(dtype=np.float32)
###########################
# get generalized advantage estimator
gaes = PPO.get_gaes(rewards=d_rewards, v_preds=v_preds, v_preds_next=v_preds_next)
gaes = np.array(gaes).astype(dtype=np.float32)
# gaes = (gaes - gaes.mean()) / gaes.std()
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
# ppo input data whose rewards is discriminator rewards
inp = [observations, actions, gaes, d_rewards, v_preds_next]
# assign parameters to old policy
PPO.assign_policy_parameters()
# train PPO
for epoch in range(6):
# sample index
sample_indices = np.random.randint(
low=0,
high=observations.shape[0],
size=32)
# sampling from input data
sampled_inp = [np.take(a=a, indices=sample_indices, axis=0) for a in inp]
# run ppo
PPO.train(
obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
# get summary
summary = PPO.get_summary(
obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
# add summary
writer.add_summary(summary, iteration)
writer.close()
writer.add_summary(
tf.Summary(value=[tf.Summary.Value(
tag='episode_reward',
simple_value=sum(episode_d_rewards))]),
iteration)
# Get generalized advantage estimator
gaes = Agent.get_gaes(rewards=d_rewards, v_preds=v_preds,
v_preds_next=v_preds_next)
#gaes = (gaes - gaes.mean()) / (gaes.std() + 1e-5)
# Train Agent
Agent_step = args.G_step
# Assign parameters to old policy
Agent.assign_policy_parameters()
total_loss = 0
clip_loss = 0
vf_loss = 0
entropy_loss = 0
l1_loss = 0
for i, _ in enumerate(observations):
# Run ppo train operation
losses = Agent.train(
obs=observations[i],
gaes=gaes[i],
rewards=d_rewards[i],
v_preds_next=v_preds_next[i],
expert_act=expert_actions[i],
lr=lr)
total_loss += losses[1]
def main(args):
env = gym.make('CartPole-v0')
env.seed(0)
ob_space = env.observation_space
Policy = Policy_net('policy', env)
Old_Policy = Policy_net('old_policy', env)
PPO = PPOTrain(Policy, Old_Policy, gamma=args.gamma)
saver = tf.train.Saver()
tl = args.train_level
if tl == 'expert':
threshold = 195
savedir = 'trained_models/ppo/expert'
logdir = 'log/train/ppo/expert/'
elif tl == 'med':
threshold = 100
savedir = 'trained_models/ppo/med'
logdir = 'log/train/ppo/med/'
else:
print("[run_ppo.py] Error: Unrecognized train level: {}".format(tl))
exit(1)
with tf.Session() as sess:
writer = tf.summary.FileWriter(logdir, sess.graph)
sess.run(tf.global_variables_initializer())
obs = env.reset()
success_num = 0
for iteration in range(args.iteration):
observations = []
actions = []
rewards = []
v_preds = []
episode_length = 0
while True: # run policy RUN_POLICY_STEPS which is much less than episode length
episode_length += 1
obs = np.stack([obs]).astype(
dtype=np.float32) # prepare to feed placeholder Policy.obs
act, v_pred = Policy.act(obs=obs, stochastic=True)
act = np.asscalar(act)
v_pred = np.asscalar(v_pred)
next_obs, reward, done, info = env.step(act)
observations.append(obs)
actions.append(act)
rewards.append(reward)
v_preds.append(v_pred)
if done:
next_obs = np.stack([next_obs]).astype(
dtype=np.float32
) # prepare to feed placeholder Policy.obs
_, v_pred = Policy.act(obs=next_obs, stochastic=True)
v_preds_next = v_preds[1:] + [np.asscalar(v_pred)]
obs = env.reset()
break
else:
obs = next_obs
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_length',
simple_value=episode_length)
]), iteration)
writer.add_summary(
tf.Summary(value=[
tf.Summary.Value(tag='episode_reward',
simple_value=sum(rewards))
]), iteration)
if sum(rewards) >= threshold:
success_num += 1
if success_num >= 100:
saver.save(sess, savedir + '/model.ckpt')
print('Clear!! Model saved.')
break
else:
success_num = 0
print("Iteration: {}, Rewards: {}".format(iteration, sum(rewards)),
end='\r')
gaes = PPO.get_gaes(rewards=rewards,
v_preds=v_preds,
v_preds_next=v_preds_next)
# convert list to numpy array for feeding tf.placeholder
observations = np.reshape(observations,
newshape=(-1, ) + ob_space.shape)
actions = np.array(actions).astype(dtype=np.int32)
gaes = np.array(gaes).astype(dtype=np.float32)
gaes = (gaes - gaes.mean()) / gaes.std()
rewards = np.array(rewards).astype(dtype=np.float32)
v_preds_next = np.array(v_preds_next).astype(dtype=np.float32)
PPO.assign_policy_parameters()
inp = [observations, actions, gaes, rewards, v_preds_next]
# train
for epoch in range(6):
# sample indices from [low, high)
sample_indices = np.random.randint(low=0,
high=observations.shape[0],
size=32)
sampled_inp = [
np.take(a=a, indices=sample_indices, axis=0) for a in inp
] # sample training data
PPO.train(obs=sampled_inp[0],
actions=sampled_inp[1],
gaes=sampled_inp[2],
rewards=sampled_inp[3],
v_preds_next=sampled_inp[4])
summary = PPO.get_summary(obs=inp[0],
actions=inp[1],
gaes=inp[2],
rewards=inp[3],
v_preds_next=inp[4])
writer.add_summary(summary, iteration)
writer.close()