unity-tech-cn
/
ml-agents
45
14
派生 39
代码 合并请求 版本发布
浏览代码

porting the changes on ppo.py and removing AgentMonitor.cs

/tag-0.2.0
vincentpierre 7 年前
当前提交
2b8353b2
共有 2 个文件被更改，包括 12 次插入 和 195 次删除
  1. 24
      python/ppo.py
  2. 183
      unity-environment/Assets/ML-Agents/Scripts/AgentMonitor.cs

24
python/ppo.py
查看文件


hidden_units = int(options['--hidden-units'])
batch_size = int(options['--batch-size'])
env = UnityEnvironment(file_name=env_name, worker_id=worker_id, curriculum=curriculum_file)
env = UnityEnvironment(file_name=env_name, worker_id=worker_id, curriculum=curriculum_path)
print(str(env))
brain_name = env.brain_names[0]

def get_progress():
if curriculum_file is not None:
if use_curriculum:
if env._curriculum.measure_type == "progress":
return steps / max_steps
elif env._curriculum.measure_type == "reward":

else:
return None
with tf.Session() as sess:
# Instantiate model parameters

steps, last_reward = sess.run([ppo_model.global_step, ppo_model.last_reward])
summary_writer = tf.summary.FileWriter(summary_path)
info = env.reset(train_mode=train_model, progress=get_progress())[brain_name]
trainer = Trainer(ppo_model, sess, info, is_continuous, use_observations, use_states)
trainer = Trainer(ppo_model, sess, info, is_continuous, use_observations, use_states, train_model)
new_info = trainer.take_action(info, env, brain_name)
new_info = trainer.take_action(info, env, brain_name, steps)
info = new_info
trainer.process_experiences(info, time_horizon, gamma, lambd)
if len(trainer.training_buffer['actions']) > buffer_size and train_model:

# Write training statistics to tensorboard.
trainer.write_summary(summary_writer, steps, env._curriculum.lesson_number)
trainer.write_summary(summary_writer, steps)
steps += 1
sess.run(ppo_model.increment_step)
if len(trainer.stats['cumulative_reward']) > 0:
mean_reward = np.mean(trainer.stats['cumulative_reward'])
sess.run(ppo_model.update_reward, feed_dict={ppo_model.new_reward: mean_reward})
last_reward = sess.run(ppo_model.last_reward)
if train_model:
steps += 1
sess.run(ppo_model.increment_step)
if len(trainer.stats['cumulative_reward']) > 0:
mean_reward = np.mean(trainer.stats['cumulative_reward'])
sess.run(ppo_model.update_reward, feed_dict={ppo_model.new_reward: mean_reward})
last_reward = sess.run(ppo_model.last_reward)
# Final save Tensorflow model
if steps != 0 and train_model:
save_model(sess, model_path=model_path, steps=steps, saver=saver)

183
unity-environment/Assets/ML-Agents/Scripts/AgentMonitor.cs
查看文件


using System.Collections;
using System.Collections.Generic;
using UnityEngine;
using UnityEngine.UI;
/**
* Attach this component to a game object with an Agent component in order to
* visulaize relevant information about your agent.
*/
public class AgentMonitor : MonoBehaviour
{
public bool fixedPosition;
/**< \brief If true, the monitor will be fixed in the top left corner of
* the screen. If false, it will move with the position of the agent*/
public float verticalOffset = 10f;
/**< \brief If fixedPosition is false, defines how high above the agent
* the monitor is */
[Header("Displayed Values")]
public bool DisplayBrainName = true;
/**< \brief If true, the name of the brain will be displayed in the monitor */
public bool DisplayBrainType = true;
/**< \brief If true, the type of the brain will be displayed in the monitor */
public bool DisplayFrameCount = true;
/**< \brief If true, the number of steps since the agent was reset will be displayed in the monitor */
public bool DisplayCurrentReward = true;
/**< \brief If true, the current reward of the agent will be displayed in the monitor */
public bool DisplayMaxReward = true;
/**< \brief If true, the maximum reward accros episodes will be displayed in the monitor */
public bool DisplayState = false;
/**< \brief If true, the current state of the agetn will be displayed in the monitor */
public bool DisplayAction = false;
/**< \brief If true, the current action of the agetn will be displayed in the monitor */
private Agent agent;
private Text texts;
private GameObject monitor;
private Slider slider;
private Image fill;
private float maxReward;
void Start()
{
agent = gameObject.GetComponent<Agent>();
GameObject canvas = GameObject.Find("AgentMonitorCanvas");
if (canvas == null)
{
canvas = new GameObject();
canvas.name = "AgentMonitorCanvas";
}
if (canvas.GetComponent<Canvas>() == null)
{
Canvas c = canvas.AddComponent<Canvas>();
c.renderMode = RenderMode.ScreenSpaceOverlay;
c.pixelPerfect = true;
}
if (canvas.GetComponent<CanvasScaler>() == null)
{
CanvasScaler cs = canvas.AddComponent<CanvasScaler>() as CanvasScaler;
cs.uiScaleMode = CanvasScaler.ScaleMode.ScaleWithScreenSize;
}
if (canvas.GetComponent<GraphicRaycaster>() == null)
{
canvas.AddComponent<GraphicRaycaster>();
}
monitor = Instantiate(Resources.Load("AgentMonitor") as GameObject);
monitor.transform.SetParent(canvas.transform, false);
slider = monitor.GetComponentInChildren<Slider>();
fill = slider.GetComponentInChildren<Image>();
texts = monitor.GetComponent<Text>();
if (fixedPosition)
{
monitor.GetComponent<RectTransform>().anchorMin = new Vector2(0, 1);
monitor.GetComponent<RectTransform>().anchorMax = new Vector2(0, 1);
monitor.GetComponent<RectTransform>().pivot = new Vector2(0, 1);
monitor.GetComponent<RectTransform>().position = new Vector3(0, monitor.GetComponent<RectTransform>().position.y, 0);
}
}
void Update()
{
if (maxReward < agent.CummulativeReward)
{
maxReward = agent.CummulativeReward;
}
texts.text = "";
if (DisplayBrainName)
{
texts.text += "Brain name : " + agent.brain.gameObject.name + "\n";
}
if (DisplayBrainType)
{
texts.text += "Mode : " + agent.brain.brainType + "\n";
}
if (DisplayFrameCount)
{
texts.text += "Frame Count :" + agent.stepCounter + "\n";
}
if (DisplayCurrentReward)
{
texts.text += "Reward : " + agent.CummulativeReward.ToString("F2") + "\n";
}
if (DisplayMaxReward)
{
texts.text += "Maximum Reward : " + maxReward.ToString("F2") + "\n";
}
if (DisplayState)
{
if (agent.brain.brainParameters.stateSpaceType == StateType.continuous)
{
texts.text += "State : ";
foreach (float f in agent.CollectState())
{
texts.text += f.ToString("F2") + " ";
}
texts.text += "\n";
}
else
{
texts.text += "State : " + ((int)agent.CollectState()[0]).ToString() + "\n";
}
}
if (DisplayAction)
{
if (agent.brain.brainParameters.actionSpaceType == StateType.continuous)
{
texts.text += "Action : ";
foreach (float f in agent.agentStoredAction)
{
texts.text += f.ToString("F2") + " ";
}
texts.text += "\n";
}
else
{
texts.text += "Action : " + ((int)agent.agentStoredAction[0]).ToString() + "\n";
}
}
if (Mathf.Abs(agent.value) > slider.maxValue)
{
slider.maxValue = Mathf.Abs(agent.value);
}
slider.value = Mathf.Abs(agent.value);
if (agent.value > 0)
{
fill.color = Color.green;
}
else
{
fill.color = Color.red;
}
if (!fixedPosition)
{
Vector3 cam2obj = gameObject.transform.position - Camera.main.transform.position;
if (Vector3.Dot(cam2obj, Camera.main.transform.forward) < 0)
{
monitor.SetActive(false);
}
else
{
monitor.SetActive(true);
monitor.transform.position = Camera.main.WorldToScreenPoint(agent.transform.position + new Vector3(0, verticalOffset, 0));
monitor.transform.localScale = 20f * Mathf.Min(1, 1f / (Vector3.Dot(cam2obj, Camera.main.transform.forward))) * new Vector3(1, 1, 1);
}
}
}
void OnDestroy()
{
Destroy(monitor);
}
}
撰写 预览
正在加载...
取消
保存