def test_create_dict(self):
env_dict = create_env_dict(self.env)
self.assertIn("obs", env_dict)
self.assertIn("act0", env_dict)
self.assertIn("act1", env_dict)
self.assertIn("act2", env_dict)
def __init__(self, max_length, seed_number, env):
env_dict = create_env_dict(env)
#override the observation length in the replay memory
env_dict['obs'] = {"dtype": numpy.float32, "shape": (17, )}
env_dict['next_obs'] = {"dtype": numpy.float32, "shape": (17, )}
print('!!!!', env_dict['obs'])
self.before_add = create_before_add_func(env)
self.storage = ReplayBuffer(max_length, env_dict)
def test_add(self):
env_dict = create_env_dict(self.env)
before_add_func = create_before_add_func(self.env)
rb = ReplayBuffer(256, env_dict)
obs = self.env.reset()
for i in range(100):
act = self.env.action_space.sample()
next_obs, rew, done, _ = self.env.step(act)
rb.add(**before_add_func(obs, act, next_obs, rew, done))
if done:
obs = self.env.reset()
else:
obs = next_obs
def __init__(self, max_length, seed_number, env):
env_dict = create_env_dict(env)
self.before_add = create_before_add_func(env)
self.storage = ReplayBuffer(max_length, env_dict)
# q.put(w)
# %% run_cell
if __name__ == "__main__":
ENV_ID = 'pong'
THREADS = 4
mp.set_start_method('spawn')
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# device = 'cpu'
params = data.params[ENV_ID]
test_env = createEnv(params.env)
shape = test_env.observation_space.shape
actions = test_env.action_space.n
env_dict = create_env_dict(test_env)
net = model.DDQN(shape, actions).to(device)
net.share_memory()
tgt_net = ptan.agent.TargetNet(net)
selector = ptan.actions.ArgmaxActionSelector()
agent = ptan.agent.DQNAgent(net,
selector,
device=device,
preprocessor=preprocess)
optimizer = torch.optim.Adam(net.parameters(), lr=params.lr)
done_training = mp.Event()
done_training.clear()
# create_buffer_with_helper_func.py
#
# Create `ReplayBuffer` for non simple space `gym.Env` with helper functions.
import gym
from cpprb import ReplayBuffer, create_env_dict, create_before_add_func
env = gym.make("Blackjack-v0")
# https://github.com/openai/gym/blob/master/gym/envs/toy_text/blackjack.py
# BlackjackEnv
# observation_space: Tuple(Discrete(32),Discrete(11),Discrete(2))
# action_space : Discrete(2)
env_dict = create_env_dict(env)
# >>> env_dict
#{'act': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'done': {'add_shape': array([-1, 1]), 'dtype': numpy.float32, 'shape': 1},
# 'next_obs0': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'next_obs1': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'next_obs2': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'obs0': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'obs1': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'obs2': {'add_shape': array([-1, 1]), 'dtype': numpy.int32, 'shape': 1},
# 'rew': {'add_shape': array([-1, 1]), 'dtype': numpy.float32, 'shape': 1}}
rb = ReplayBuffer(256, env_dict)