def check_agent():
from elegantrl2.agent import AgentPPO as Agent
agent = Agent()
net_dim = 2**7
state_dim = 8
action_dim = 2
agent.init(net_dim,
state_dim,
action_dim,
learning_rate=1e-4,
if_use_gae=False,
gpu_id=0)
# print(agent.act.state_dict())
for key, value in agent.act.state_dict().items():
print(key, value.shape)
def demo_get_video_to_watch_gym_render():
import cv2 # pip3 install opencv-python
import gym # pip3 install gym==0.17 pyglet==1.5.0 # env.render() bug in gym==0.18, pyglet==1.6
import torch
"""parameters"""
env_name = 'LunarLanderContinuous-v2'
env = PreprocessEnv(env=gym.make(env_name))
'''initialize agent'''
agent = None # means use random action
if agent is None: # use random action
device = None
else:
from elegantrl2.agent import AgentPPO
agent = AgentPPO() # means use the policy network which saved in cwd
cwd = f'./{env_name}_{agent.__class__.__name__}/' # current working directory path
net_dim = 2 ** 9 # 2 ** 7
state_dim = env.state_dim
action_dim = env.action_dim
agent.init(net_dim, state_dim, action_dim)
agent.save_load_model(cwd=cwd, if_save=False)
device = agent.device
'''initialize evaluete and env.render()'''
save_frame_dir = '' # means don't save video, just open the env.render()
# save_frame_dir = 'frames' # means save video in this directory
if save_frame_dir:
os.makedirs(save_frame_dir, exist_ok=True)
state = env.reset()
episode_return = 0
step = 0
for i in range(2 ** 10):
print(i) if i % 128 == 0 else None
for j in range(1):
if agent is None:
action = env.action_space.sample()
else:
s_tensor = torch.as_tensor((state,), dtype=torch.float32, device=device)
a_tensor = agent.act(s_tensor)
action = a_tensor.detach().cpu().numpy()[0] # if use 'with torch.no_grad()', then '.detach()' not need.
next_state, reward, done, _ = env.step(action)
episode_return += reward
step += 1
if done:
print(f'{i:>6}, {step:6.0f}, {episode_return:8.3f}, {reward:8.3f}')
state = env.reset()
episode_return = 0
step = 0
else:
state = next_state
if save_frame_dir:
frame = env.render('rgb_array')
cv2.imwrite(f'{save_frame_dir}/{i:06}.png', frame)
cv2.imshow('OpenCV Window', frame)
cv2.waitKey(1)
else:
env.render()
env.close()
'''convert frames png/jpg to video mp4/avi using ffmpeg'''
if save_frame_dir:
frame_shape = cv2.imread(f'{save_frame_dir}/{3:06}.png').shape
print(f"frame_shape: {frame_shape}")
save_video = 'gym_render.mp4'
os.system(f"| Convert frames to video using ffmpeg. Save in {save_video}")
os.system(f'ffmpeg -r 60 -f image2 -s {frame_shape[0]}x{frame_shape[1]} '
f'-i ./{save_frame_dir}/%06d.png '
f'-crf 25 -vb 20M -pix_fmt yuv420p {save_video}')
def get_video_to_watch_gym_render():
import cv2 # pip3 install opencv-python
import gym # pip3 install gym==0.17 pyglet==1.5.0 # env.render() bug in gym==0.18, pyglet==1.6
import torch
'''choose env'''
import pybullet_envs # for python-bullet-gym
dir(pybullet_envs)
# from elegantrl2.env import PreprocessEnv
env_name = [
'BipedalWalker-v3', 'AntBulletEnv-v0', 'KukaBulletEnv-v0',
'ReacherBulletEnv-v0', 'PusherBulletEnv-v0', "ThrowerBulletEnv-v0",
"StrikerBulletEnv-v0"
][1]
env = PreprocessEnv(env=gym.make(env_name))
'''initialize agent'''
agent = None
from elegantrl2.agent import AgentPPO
agent = AgentPPO()
agent.if_use_dn = True
net_dim = 2**8
cwd = f'./{env_name}_4/'
# from elegantrl2.agent import AgentModSAC
# agent = AgentModSAC()
# agent.if_use_dn = True
# net_dim = 2 ** 8
# cwd = f'./{env_name}_2/'
device = None
if agent is not None:
state_dim = env.state_dim
action_dim = env.action_dim
agent.init(net_dim, state_dim, action_dim)
agent.save_load_model(cwd=cwd, if_save=False)
device = agent.device
rd.seed(194686)
torch.manual_seed(1942876)
'''initialize evaluete and env.render()'''
save_frame_dir = 'frames'
if save_frame_dir:
os.makedirs(save_frame_dir, exist_ok=True)
state = env.reset()
episode_return = 0
step = 0
for i in range(2**9):
print(i) if i % 128 == 0 else None
for j in range(1):
if agent is not None:
s_tensor = torch.as_tensor((state, ),
dtype=torch.float32,
device=device)
a_tensor = agent.act(s_tensor)
action = a_tensor.detach().cpu().numpy(
)[0] # if use 'with torch.no_grad()', then '.detach()' not need.
else:
action = env.action_space.sample()
next_state, reward, done, _ = env.step(action)
episode_return += reward
step += 1
if done:
print(
f'{i:>6}, {step:6.0f}, {episode_return:8.3f}, {reward:8.3f}'
)
state = env.reset()
episode_return = 0
step = 0
else:
state = next_state
frame = env.render('rgb_array')
frame = frame[50:210, 50:270] # (240, 320) AntPyBulletEnv-v0
# frame = cv2.resize(frame[:, :500], (500//2, 720//2))
cv2.imwrite(f'{save_frame_dir}/{i:06}.png', frame)
cv2.imshow('', frame)
cv2.waitKey(1)
env.close()
# exit()
'''convert frames png/jpg to video mp4/avi using ffmpeg'''
if save_frame_dir:
frame_shape = cv2.imread(f'{save_frame_dir}/{3:06}.png').shape
print(f"frame_shape: {frame_shape}")
save_video = 'gym_render.mp4'
os.system(
f"| Convert frames to video using ffmpeg. Save in {save_video}")
os.system(
f'ffmpeg -r 60 -f image2 -s {frame_shape[0]}x{frame_shape[1]} '
f'-i ./{save_frame_dir}/%06d.png '
f'-crf 25 -vb 20M -pix_fmt yuv420p {save_video}')
