dodox/metastable-baselines
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Trying to converge on simple columbus envs

Browse Source
This commit is contained in:
Dominik Moritz Roth 2022-06-20 23:12:42 +02:00
parent b9303416ac
commit e71735bf79
1 changed files with 25 additions and 55 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

80
test.py
View File

@ -2,91 +2,61 @@ import gym
from gym.envs.registration import register from gym.envs.registration import register
import numpy as np import numpy as np
import time import time
import datetime
from stable_baselines3 import SAC, PPO, A2C from stable_baselines3 import SAC, PPO, A2C
from stable_baselines3.common.evaluation import evaluate_policy from stable_baselines3.common.evaluation import evaluate_policy
from stable_baselines3.common.policies import ActorCriticCnnPolicy, ActorCriticPolicy, MultiInputActorCriticPolicy
from sb3_trl.trl_pg import TRL_PG from sb3_trl.trl_pg import TRL_PG
from columbus import env import columbus
register(
id='ColumbusTestRay-v0',
entry_point=env.ColumbusTestRay,
max_episode_steps=30*60*5,
)
def main(): def main(env_name='ColumbusEasyObstacles-v0'):
#env = gym.make("LunarLander-v2") env = gym.make(env_name)
env = gym.make("ColumbusTestRay-v0")
ppo = PPO(
"MlpPolicy",
env,
verbose=1,
tensorboard_log="./logs_tb/test/ppo",
use_sde=False,
ent_coef=0.0001,
learning_rate=0.0004
)
ppo_base_sde = PPO(
"MlpPolicy",
env,
verbose=1,
tensorboard_log="./logs_tb/test/ppo_base_sde/",
use_sde=True,
sde_sample_freq=30*20,
sde_net_arch=[],
ent_coef=0.000001,
learning_rate=0.0003
)
ppo_latent_sde = PPO( ppo_latent_sde = PPO(
"MlpPolicy", "MlpPolicy",
env, env,
verbose=1, verbose=1,
tensorboard_log="./logs_tb/test/ppo_latent_sde/", tensorboard_log="./logs_tb/"+env_name+"/ppo_latent_sde/",
use_sde=True, use_sde=True,
sde_sample_freq=30*20, sde_sample_freq=30*15,
ent_coef=0.000001, ent_coef=0.0032,
learning_rate=0.0003 vf_coef=0.0005,
) gamma=0.95,
a2c = A2C( learning_rate=0.02
"MlpPolicy",
env,
verbose=1,
tensorboard_log="./logs_tb/test/a2c/",
)
trl = TRL_PG(
"MlpPolicy",
env,
verbose=0,
tensorboard_log="./logs_tb/test/trl_pg/",
) )
#trl = TRL_PG(
# "MlpPolicy",
# env,
# verbose=0,
# tensorboard_log="./logs_tb/"+env_name+"/trl_pg/",
#)
#print('PPO:') print('PPO_LATENT_SDE:')
#testModel(ppo, 500000, showRes = True, saveModel=True, n_eval_episodes=4) testModel(ppo_latent_sde, 100000, showRes = True, saveModel=True, n_eval_episodes=0)
print('PPO_BASE_SDE:')
testModel(ppo_base_sde, 200000, showRes = True, saveModel=True, n_eval_episodes=0)
#print('A2C:')
#testModel(a2c, showRes = True)
#print('TRL_PG:') #print('TRL_PG:')
#testModel(trl) #testModel(trl)
def testModel(model, timesteps=100000, showRes=False, saveModel=False, n_eval_episodes=16): def testModel(model, timesteps=150000, showRes=False, saveModel=False, n_eval_episodes=16):
env = model.get_env() env = model.get_env()
model.learn(timesteps) model.learn(timesteps)
if saveModel:
now = datetime.datetime.now().strftime('%d.%m.%Y-%H:%M')
model.save(model.tensorboard_log.replace('./logs_tb/','').replace('/','_')+now+'.zip')
if n_eval_episodes: if n_eval_episodes:
mean_reward, std_reward = evaluate_policy(model, env, n_eval_episodes=n_eval_episodes, deterministic=False) mean_reward, std_reward = evaluate_policy(model, env, n_eval_episodes=n_eval_episodes, deterministic=False)
print('Reward: '+str(round(mean_reward,3))+'±'+str(round(std_reward,2))) print('Reward: '+str(round(mean_reward,3))+'±'+str(round(std_reward,2)))
if showRes: if showRes:
model.save("model")
input('<ready?>') input('<ready?>')
obs = env.reset() obs = env.reset()
# Evaluate the agent # Evaluate the agent
episode_reward = 0 episode_reward = 0
for _ in range(1000): for _ in range(30*60*5):
time.sleep(1/30) time.sleep(1/30)
action, _ = model.predict(obs, deterministic=False) action, _ = model.predict(obs, deterministic=False)
obs, reward, done, info = env.step(action) obs, reward, done, info = env.step(action)