def run():
parser = argparse.ArgumentParser()
parser.add_argument('--env_id', type=str, default='AntBulletEnv-v0')
parser.add_argument('--log_name', type=str, default='')
parser.add_argument('--cuda', action='store_true')
parser.add_argument('--seed', type=int, default=0)
args = parser.parse_args()
if args.log_name:
log_dir = os.path.join('logs', args.env_id, args.log_name)
else:
env_dir = os.path.join('logs', args.env_id, '*')
dirs = glob.glob(env_dir)
log_dir = max(dirs, key=os.path.getctime)
print(f'using {log_dir}')
env = gym.make(args.env_id)
device = torch.device(
"cuda" if args.cuda and torch.cuda.is_available() else "cpu")
policy = GaussianPolicy(
env.observation_space.shape[0],
env.action_space.shape[0],
hidden_units=[256, 256]).to(device)
policy.load(os.path.join(log_dir, 'model', 'policy.pth'))
grad_false(policy)
def exploit(state):
state = torch.FloatTensor(state).unsqueeze(0).to(device)
with torch.no_grad():
_, _, action = policy.sample(state)
return action.cpu().numpy().reshape(-1)
env.render()
while True:
state = env.reset()
episode_reward = 0.
done = False
while not done:
env.render()
action = exploit(state)
next_state, reward, done, _ = env.step(action)
episode_reward += reward
state = next_state
print(f'total reward: {episode_reward}')
time.sleep(1)
def testing():
parser = argparse.ArgumentParser()
parser.add_argument('--env_name', type=str, default='HalfCheetah-v2')
parser.add_argument('--num_episode', type=int, default=10)
args = parser.parse_args()
num_episode = args.num_episode
env = gym.make(args.env_name)
device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
policy = GaussianPolicy(
env.observation_space.shape[0],
env.action_space.shape[0],
hidden_units=[256, 256]).to(device)
policy.load(os.path.join('models', args.env_name, 'policy.pth'))
grad_false(policy)
def exploit(state):
state = torch.FloatTensor(state).unsqueeze(0).to(device)
with torch.no_grad():
_, _, action = policy.sample(state)
return action.cpu().numpy().reshape(-1)
e_rewrads = []
for _ in range(num_episode):
state = env.reset()
episode_reward = 0.
done = False
while not done:
if num_episode <= 1:
env.render()
action = exploit(state)
next_state, reward, done, _ = env.step(action)
episode_reward += reward
state = next_state
e_rewrads.append(episode_reward)
print("Average reward of " + args.env_name + " is %.1f"%(np.mean(e_rewrads)))
print("Average std of " + args.env_name + " is %.1f"%(np.std(e_rewrads)))
def run():
parser = argparse.ArgumentParser()
parser.add_argument('--env_id', type=str, default='HalfCheetah-v2')
parser.add_argument('--log_name', type=str, default='sac-seed0-datetime')
parser.add_argument('--cuda', action='store_true')
parser.add_argument('--seed', type=int, default=0)
args = parser.parse_args()
log_dir = os.path.join('logs', args.env_id, args.log_name)
env = gym.make(args.env_id)
device = torch.device(
"cuda" if args.cuda and torch.cuda.is_available() else "cpu")
policy = GaussianPolicy(env.observation_space.shape[0],
env.action_space.shape[0],
hidden_units=[256, 256]).to(device)
policy.load(os.path.join(log_dir, 'model', 'policy.pth'))
grad_false(policy)
def exploit(state):
state = torch.FloatTensor(state).unsqueeze(0).to(device)
with torch.no_grad():
_, _, action = policy.sample(state)
return action.cpu().numpy().reshape(-1)
state = env.reset()
episode_reward = 0.
done = False
while not done:
env.render()
action = exploit(state)
next_state, reward, done, _ = env.step(action)
episode_reward += reward
state = next_state
import torch
from torch.optim import Adam
from torch.utils.tensorboard import SummaryWriter
from rltorch.memory import MultiStepMemory, PrioritizedMemory
from model import TwinnedQNetwork, GaussianPolicy
import random
import gym
from dst_d import DeepSeaTreasure
date = 'sac-seed0-20210512-2219'
critic = TwinnedQNetwork(2, 2, 2, [256, 256])
critic.load('./logs/dst_d-v0/' + date + '/model/critic.pth')
policy = GaussianPolicy(4, 2, [256, 256])
policy.load('./logs/dst_d-v0/' + date + '/model/policy.pth')
device = 'cuda'
vis = visdom.Visdom()
env = gym.make('dst_d-v0')
def q_heatmap(action, prefer):
prefer = torch.tensor(prefer, dtype=torch.float32)
action = torch.tensor(action, dtype=torch.float32)
value = np.empty([11, 11])
time = np.empty([11, 11])
for i in range(11):