def main(args):
env = gym.make('Walker2d-v1')
reward_history = []
agent = DDPG(env)
agent.construct_model(args.gpu)
saver = tf.train.Saver()
if args.model_path is not None:
# reuse saved model
saver.restore(agent.sess, args.model_path)
else:
# build a new model
agent.sess.run(tf.global_variables_initializer())
for episode in range(args.ep):
# env init
state = env.reset()
total_rewards = 0
for step in range(env.spec.timestep_limit):
env.render()
action = agent.sample_action(state[np.newaxis, :], explore=False)
# act
next_state, reward, done, _ = env.step(action[0])
total_rewards += reward
agent.store_experience(state, action, reward, next_state, done)
agent.update_model()
# shift
state = next_state
if done:
break
reward_history.append(total_rewards)
print('Ep%d reward:%d' % (episode+1, total_rewards))
print('Average rewards: ', np.mean(reward_history))
def main(args):
env = gym.make('Walker2d-v1')
agent = DDPG(env)
agent.construct_model(args.gpu)
saver = tf.train.Saver(max_to_keep=1)
if args.model_path is not None:
# reuse saved model
saver.restore(agent.sess, args.model_path)
ep_base = int(args.model_path.split('_')[-1])
else:
# build a new model
agent.sess.run(tf.global_variables_initializer())
ep_base = 0
MAX_EPISODES = 100000
TEST = 10
for episode in range(MAX_EPISODES):
# env init
state = env.reset()
total_rewards = 0
for step in range(env.spec.timestep_limit):
action = agent.sample_action(state[np.newaxis, :], explore=True)
# act
next_state, reward, done, _ = env.step(action[0])
total_rewards += reward
agent.store_experience(state, action, reward, next_state, done)
agent.update_model()
# shift
state = next_state
if done:
print('Ep %d global_steps: %d Reward: %.2f' %
(episode + 1, agent.global_steps, total_rewards))
# reset ou noise
agent.ou.reset()
break
# Evaluation per 100 ep
if episode % 100 == 0 and episode > 100:
total_rewards = 0
for ep_eval in range(TEST):
state = env.reset()
for step_eval in range(env.spec.timestep_limit):
action = agent.sample_action(state[np.newaxis, :],
explore=False)
next_state, reward, done, _ = env.step(action[0])
total_rewards += reward
state = next_state
if done:
break
mean_rewards = total_rewards / TEST
# logging
print('\n')
print('Episode: %d' % (episode + 1))
print('Gloabal steps: %d' % agent.global_steps)
print('Mean reward: %.2f' % mean_rewards)
print('\n')
if not os.path.isdir(args.save_path):
os.makedirs(args.save_path)
save_name = args.save_path \
+ str(episode) + '_' + str(round(mean_rewards, 2))
saver.save(agent.sess, save_name)
def main(args):
set_random_seed(args.seed)
env = gym.make('Walker2d-v1')
agent = DDPG(env, args)
agent.construct_model(args.gpu)
saver = tf.train.Saver(max_to_keep=1)
if args.model_path is not None:
# reuse saved model
saver.restore(agent.sess, args.model_path)
ep_base = int(args.model_path.split('_')[-1])
best_avg_rewards = float(args.model_path.split('/')[-1].split('_')[0])
else:
# build a new model
agent.sess.run(tf.global_variables_initializer())
ep_base = 0
best_avg_rewards = None
reward_history, step_history = [], []
train_steps = 0
for ep in range(args.max_ep):
# env init
state = env.reset()
ep_rewards = 0
for step in range(env.spec.timestep_limit):
action = agent.sample_action(state[np.newaxis, :], noise=True)
# act
next_state, reward, done, _ = env.step(action[0])
train_steps += 1
ep_rewards += reward
agent.store_experience(state, action, reward, next_state, done)
agent.update_model()
# shift
state = next_state
if done:
print('Ep %d global_steps: %d Reward: %.2f' %
(ep + 1, agent.global_steps, ep_rewards))
# reset ou noise
agent.ou.reset()
break
step_history.append(train_steps)
if not reward_history:
reward_history.append(ep_rewards)
else:
reward_history.append(reward_history[-1] * 0.99 + ep_rewards + 0.01)
# Evaluate during training
if ep % args.log_every == 0 and ep > 0:
ep_rewards = 0
for ep_eval in range(args.test_ep):
state = env.reset()
for step_eval in range(env.spec.timestep_limit):
action = agent.sample_action(
state[np.newaxis, :], noise=False)
next_state, reward, done, _ = env.step(action[0])
ep_rewards += reward
state = next_state
if done:
break
curr_avg_rewards = ep_rewards / args.test_ep
# logging
print('\n')
print('Episode: %d' % (ep + 1))
print('Gloabal steps: %d' % agent.global_steps)
print('Mean reward: %.2f' % curr_avg_rewards)
print('\n')
if not best_avg_rewards or (curr_avg_rewards >= best_avg_rewards):
best_avg_rewards = curr_avg_rewards
if not os.path.isdir(args.save_path):
os.makedirs(args.save_path)
save_name = args.save_path + str(round(best_avg_rewards, 2)) \
+ '_' + str(ep_base + ep + 1)
saver.save(agent.sess, save_name)
print('Model save %s' % save_name)
plt.plot(step_history, reward_history)
plt.xlabel('steps')
plt.ylabel('running reward')
plt.show()