def main():
torch.set_num_threads(1)
torch.manual_seed(0)
env = gym.make(env_name)
env.seed(seed)
print('New model')
policy = Policy('actor_critic', env.observation_space.shape[0], env.action_space.n)
policy.to(device)
optimizer = PPO(policy, clip_param, ppo_epoch, mini_batch_size,
value_loss_coef, entropy_coef, learning_rate,
max_grad_norm)
episode_rewards = deque(maxlen=50)
for eps in range(0, n_eps + 1):
state = env.reset()
storage = Storage(device=device)
policy.eval()
episode_rewards.append(test_env(policy, gym.make(env_name)))
if eps % 5 == 0:
print('Avg reward', np.mean(episode_rewards))
for step in range(n_steps):
state = torch.FloatTensor(state).to(device)
with torch.no_grad():
value, action, log_prob = policy.act(state)
next_state, reward, done, _ = env.step(action.item())
storage.push(state, action, log_prob, value, reward, done)
state = next_state
if done:
state = env.reset()
next_state = torch.FloatTensor(next_state).to(device)
with torch.no_grad():
next_value = policy.get_value(next_state).detach()
storage.compute(next_value)
policy.train()
value_loss, action_loss, dist_entropy = optimizer.update(storage)
with open('metrics.csv', 'a') as metrics:
metrics.write('{},{},{}\n'.format(value_loss, action_loss, dist_entropy))
def main():
torch.set_num_threads(1)
torch.manual_seed(0)
env = gym.make(env_name)
env.seed(42)
print('New model')
policy = Policy('dqn', env.observation_space.shape[0], env.action_space.n)
target_policy = Policy('dqn', env.observation_space.shape[0],
env.action_space.n)
policy.to(device)
target_policy.to(device)
target_policy.load_state_dict(policy.state_dict())
optimizer = DQNOptimizer(policy, target_policy, mini_batch_size, discount,
learning_rate, update_epochs)
episode_rewards = deque(maxlen=50)
get_epsilon = lambda episode: np.exp(-episode * e_decay)
for eps in range(0, n_eps + 1):
state = env.reset()
storage = Storage(device=device)
episode_rewards.append(test_env(target_policy, gym.make(env_name)))
if eps % 5 == 0:
print('Avg reward', np.mean(episode_rewards))
for step in range(n_steps):
state = torch.FloatTensor(state).to(device)
with torch.no_grad():
action = policy.act(state, get_epsilon(eps))
next_state, reward, done, _ = env.step(action.item())
storage.push(state, action, reward, next_state, done)
state = next_state
if done:
state = env.reset()
storage.compute()
loss = optimizer.update(storage)
if eps % target_policy_update:
target_policy.load_state_dict(policy.state_dict())
with open('metrics.csv', 'a') as metrics:
metrics.write('{}\n'.format(loss))
def main():
torch.set_num_threads(1)
torch.manual_seed(0)
env = MountainCarEnvInherit()
env.seed(42)
meta_policy = Policy('dqn', env.observation_space.shape[0], goal_object.get_size())
target_meta_policy = Policy('dqn', env.observation_space.shape[0], goal_object.get_size())
policy = Policy('dqn', env.observation_space.shape[0] + goal_object.get_size(), env.action_space.n)
target_policy = Policy('dqn', env.observation_space.shape[0] + goal_object.get_size(), env.action_space.n)
meta_policy.to(device)
target_meta_policy.to(device)
target_meta_policy.load_state_dict(meta_policy.state_dict())
policy.to(device)
target_policy.to(device)
target_policy.load_state_dict(policy.state_dict())
optimizer_meta_policy = DQNOptimizer(meta_policy, target_meta_policy, mini_batch_size, discount, learning_rate, update_epochs)
optimizer_policy = DQNOptimizer(policy, target_policy, mini_batch_size, discount, learning_rate, update_epochs)
episode_rewards = deque(maxlen=50)
get_meta_epsilon = lambda episode: np.exp(-episode * e_meta_decay)
get_epsilon = lambda episode: np.exp(-episode * e_decay)
frame = 0
meta_frame = 0
for eps in range(0, n_eps + 1):
if eps % 1 == 0:
episode_rewards.append(test_env(meta_policy, policy, MountainCarEnvInherit()))
print('Avg reward', np.mean(episode_rewards))
storage = Storage(device=device)
storage_meta = Storage(device=device)
print('Game', eps)
state0 = env.reset()
state = state0.copy()
state = torch.FloatTensor(state).to(device)
done = False
for step in range(100):
extrinsic_reward = 0
goal = meta_policy.act(state, get_meta_epsilon(step))
onehot_goal = to_onehot(goal, goal_object.get_size())
print('Goal', goal)
goal_reached = False
for i in range(100):
joint_state = torch.FloatTensor(np.concatenate([state.cpu().numpy(), onehot_goal], axis=0)).to(device)
with torch.no_grad():
action = policy.act(joint_state, get_epsilon(frame))
next_state, reward, done, _ = env.step(action.item())
intrinsic_reward = get_intrinsic_reward(goal, next_state)
goal_reached = True if intrinsic_reward else False
joint_next_state = np.concatenate([next_state, onehot_goal], axis=0)
storage.push(joint_state, action, intrinsic_reward, joint_next_state, done)
extrinsic_reward += reward
state = next_state
state = torch.FloatTensor(state).to(device)
frame += 1
if done or goal_reached:
break
goal = torch.LongTensor([goal]).to(device)
storage_meta.push(torch.FloatTensor(state0).to(device), goal, extrinsic_reward, next_state, done)
meta_frame += 1
if done:
break
storage.compute()
storage_meta.compute()
loss_meta = optimizer_meta_policy.update(storage_meta)
loss = optimizer_policy.update(storage)
if eps % target_policy_update:
target_meta_policy.load_state_dict(meta_policy.state_dict())
target_policy.load_state_dict(policy.state_dict())
with open('metrics.csv', 'a') as metrics:
metrics.write('{},{}\n'.format(loss_meta, loss))
