def test_vecnormalize(self):
"""
Tests working of the VecNormalize wrapper
"""
env = VectorEnv("CartPole-v1", 2)
env = VecNormalize(env, norm_obs=True, norm_reward=True, clip_reward=1.0,)
env.reset()
_, rewards, _, _ = env.step(env.sample())
env.close()
assert np.all(-1.0 <= rewards)
assert np.all(1.0 >= rewards)
def test_vecenv_parallel(self):
"""
Tests working of parallel VecEnvs
"""
env = VectorEnv("CartPole-v1", 2, parallel=True)
env.seed(0)
observation_space, action_space = env.get_spaces()
env.reset()
env.step(env.sample())
env.close()
def test_vecmonitor(self):
"""
Tests working of the VecMonitor wrapper
"""
env = VectorEnv("CartPole-v1", 2)
env = VecMonitor(env, history_length=1)
env.reset()
_, _, _, info = env.step(env.sample())
env.close()
dones = [0, 0]
while not dones[0]:
_, _, dones, infos = env.step(env.sample())
info = infos[0]["episode"]
assert info["Episode Rewards"]
assert info["Episode Length"]
assert info["Time taken"]
import numpy as np
import optuna
import torch
from genrl.agents.a2c.a2c import A2C
from genrl.environments.suite import VectorEnv
from genrl.trainers.onpolicy import OnPolicyTrainer
env = VectorEnv("CartPole-v0")
def tune_A2C(trial):
# Define hyperparameters that are relevant for training
# Choose a suggestion type and range (float/int and log/uniform)
lr_value = trial.suggest_float("lr_value", 1e-5, 1e-2, log=True)
lr_policy = trial.suggest_float("lr_policy", 1e-5, 1e-2, log=True)
rollout_size = trial.suggest_int("rollout_size", 100, 10000, log=True)
entropy_coeff = trial.suggest_float("entropy_coeff", 5e-4, 2e-1, log=True)
agent = A2C(
"mlp",
env,
lr_value=lr_value,
lr_policy=lr_policy,
rollout_size=rollout_size,
entropy_coeff=entropy_coeff,
)
trainer = OnPolicyTrainer(
agent,
env,
log_interval=10,
import numpy as np
import optuna
import torch
from genrl.agents.td3.td3 import TD3
from genrl.environments.suite import VectorEnv
from genrl.trainers.offpolicy import OffPolicyTrainer
env = VectorEnv("Pendulum-v0")
def objective(trial):
lr_value = trial.suggest_float("lr_value", 1e-6, 1e-1, log=True)
lr_policy = trial.suggest_float("lr_policy", 1e-6, 1e-1, log=True)
replay_size = trial.suggest_int("replay_size", 1e2, 1e5, log=True)
max_ep_len = trial.suggest_int("max_ep_len", 1e3, 50000, log=True)
agent = TD3("mlp",
env,
lr_value=lr_value,
lr_policy=lr_policy,
replay_size=replay_size)
trainer = OffPolicyTrainer(
agent,
env,
log_interval=5,
epochs=100,
max_timesteps=16500,
evaluate_episodes=10,
max_ep_len=max_ep_len,
)
def test_vecenv_serial(self):
"""
Tests working of serial VecEnvs
"""
env = VectorEnv("CartPole-v1", 2, parallel=False)
env.seed(0)
ob, ac = env.observation_spaces, env.action_spaces
env.reset()
env.step(env.sample())
env.close()