def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump(
(self.policy_net, self.value_net, self.running_state),
open("{}/{}_trpo.p".format(save_path, self.env_id), "wb"),
)
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump(
(self.policy_net, self.q_net_1, self.q_net_2, self.running_state),
open('{}/{}_sac_alpha.p'.format(save_path, self.env_id), 'wb'))
def save_model(self, save_path):
check_path(save_path)
# torch.save((self.discriminator, self.policy, self.value), f"{save_path}/{self.exp_name}.pt")
torch.save(self.discriminator,
f"{save_path}/{self.env_id}_Discriminator.pt")
torch.save(self.policy, f"{save_path}/{self.env_id}_Policy.pt")
torch.save(self.value, f"{save_path}/{self.env_id}_Value.pt")
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump(self.running_state,
open('{}/{}_dueling_dqn_tf2.p'.format(save_path, self.env_id), 'wb'))
self.value_net.save_weights(
"{}/{}_dueling_dqn_tf2".format(save_path, self.env_id))
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump(
self.running_state,
open('{}/{}_reinforce_tf2.p'.format(save_path, self.env_id), 'wb'))
self.policy_net.save_weights("{}/{}_reinforce_tf2".format(
save_path, self.env_id))
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump(
self.running_state,
open("{}/{}_trpo_tf2.p".format(save_path, self.env_id), "wb"),
)
self.policy_net.save_weights(
"{}/{}_trpo_tf2_p".format(save_path, self.env_id)
)
self.value_net.save_weights(
"{}/{}_trpo_tf2_v".format(save_path, self.env_id)
)
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.policy_net, self.value_net, self.running_state),
open('{}/{}_ppo.p'.format(save_path, self.env_id), 'wb'))
def save(self, save_path):
"""save model"""
check_path(save_path)
pickle.dump((self.ac_net, self.running_state),
open('{}/{}_a2c.p'.format(save_path, self.env_id), 'wb'))
def main(env_id, n_trajs, model_path, data_path, render, seed, obs_type):
"""
Collect trajectories from pre-trained models by PPO
"""
if data_path is not None:
check_path(data_path)
env, _, num_states, num_actions = get_env_info(env_id)
# seed
env.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
model = pickle.load(open(model_path, 'rb'))
model.running_state.fix = True
states, actions, rewards, dones, next_states = [], [], [], [], []
for i_iter in range(1, n_trajs + 1):
state = env.reset()
ep_reward = 0
n_step = 0
ep_states, ep_actions, ep_rewards, ep_dones, ep_next_states = [], [], [], [], []
while True:
if render:
env.render()
normalized_state = model.running_state(state)
action = model.choose_action(normalized_state)
next_state, reward, done, _ = env.step(action)
normalized_next_state = model.running_state(next_state)
ep_reward += reward
n_step += 1
ep_states.append(state if obs_type == 0 else normalized_state)
ep_actions.append(action)
ep_rewards.append(reward)
ep_dones.append(done)
ep_next_states.append(next_state if obs_type == 0
else
normalized_next_state)
if done:
states.extend(ep_states)
actions.extend(ep_actions)
rewards.extend(ep_rewards)
dones.extend(ep_dones)
next_states.extend(ep_next_states)
print(
f"Iter: {i_iter}, step: {n_step}, episode Reward: {ep_reward}")
break
state = next_state
env.close()
states = np.r_[states].reshape((-1, num_states))
next_states = np.r_[next_states].reshape((-1, num_states))
actions = np.r_[actions].reshape((-1, 1))
rewards = np.r_[rewards].reshape((-1, 1))
dones = np.r_[dones].reshape((-1, 1))
numpy_dict = {
'obs': states,
'action': actions,
'reward': rewards,
'done': dones,
'next_obs': next_states
} # type: Dict[str, np.ndarray]
save_path = f"{data_path}/{env_id}" if data_path is not None else env_id
np.savez(f"{save_path}.npz", **numpy_dict)