def test_single():
env = agnes.make_env('CartPole-v0')
runner = agnes.Single(env, agnes.PPO, agnes.MLP, config=test_config())
runner.log(agnes.log)
runner.run(100)
runner.trainer.save("Test.pth")
runner.worker.load("Test.pth")
def test_single():
env = agnes.make_env('CartPole-v0')
runner = agnes.Single(env, agnes.PPO, agnes.RNN, config=test_config())
runner.log(agnes.log)
runner.run()
import agnes
env_name = "Ant-v2"
if __name__ == '__main__':
env = agnes.make_env(env_name)
runner = agnes.Single(env, agnes.PPO, agnes.MLP)
runner.load("results/MuJoCo/Ant-v2_MLP/PPO/weights.pth")
agnes.common.Visualize(runner.worker, env).prerun(1000).run()
import time
import agnes
env_name = "InvertedDoublePendulum-v2" # "Swimmer-v2" # "BreakoutNoFrameskip-v4" #
if __name__ == '__main__':
envs = agnes.make_vec_env(env_name)
runner = agnes.Single(envs, agnes.PPORND, agnes.MLP)
runner.log(agnes.TensorboardLogger('.logs/'), agnes.log)
runner.run()
env = agnes.make_env(env_name)
agnes.common.Visualize(runner.worker, env).run()
import agnes
from agnes.algos.base import _BaseAlgo
from gym.spaces import Space
class RandomAlgo(_BaseAlgo):
get_config = agnes.PPO.get_config
def __init__(self, nn,
observation_space: Space,
action_space: Space,
*args, **kwargs):
super().__init__()
self.action_space = action_space
def __call__(self, state, done):
return self.action_space.sample(), None, None
env_name = "Ant-v2" # "InvertedDoublePendulum-v2" # "Swimmer-v2" #
if __name__ == '__main__':
env = agnes.make_env(env_name)
runner = agnes.Single(env, RandomAlgo, agnes.MLP)
agnes.common.Visualize(runner.worker, env).run()
)
video.write(prep)
self.state, _, done, _ = self.env.step(dist.sample().cpu().numpy())
if done.item():
self.hidden = None
video.release()
def _save_output(self, module, input, output):
self.outputs.append(output[0])
def _save_gradient(self, module, grad_input, grad_output):
self.gradients.append(grad_output[0])
env_name = "BreakoutNoFrameskip-v4"
env = agnes.make_env(env_name, config={"frame_stack": True})
config, _ = agnes.PPO.get_config(env["env_type"])
runner = agnes.Single(env, agnes.PPO, agnes.LSTMCNN, config=config)
runner.trainer.load("results/Atari-BreakoutNoFrameskip-v4-PPO-10M/Breakout.pth")
VisualizeAttention(env, runner, seconds=60, layer_num=1).run()
print("Done!")
def test_vec():
envs = agnes.make_vec_env('Pendulum-v0')
runner = agnes.Single(envs, agnes.PPO, agnes.MLP, config=test_config())
runner.log(agnes.log)
runner.run()
def test_single():
env = agnes.make_env('Pendulum-v0')
runner = agnes.Single(env, agnes.PPO, agnes.MLP, config=test_config())
runner.log(agnes.CsvLogger())
runner.run()
import agnes
import time
env_name = "BreakoutNoFrameskip-v4"
if __name__ == '__main__':
envs = agnes.make_vec_env(env_name, envs_num=4)
runner = agnes.Single(envs, agnes.PPO, agnes.CNN)
runner.log(agnes.TensorboardLogger(), agnes.log)
runner.run()
import agnes
def test_config():
return dict(timesteps=30000,
nsteps=128,
nminibatches=4,
gamma=1.0,
lam=1.0,
noptepochs=4,
max_grad_norm=40.0,
learning_rate=1e-3,
cliprange=lambda x: 0.3 * x,
vf_coef=1.0,
ent_coef=.005)
env_name = "CartPole-v0"
if __name__ == '__main__':
envs = agnes.make_vec_env(env_name, envs_num=32)
runner = agnes.Single(envs, agnes.PPO, agnes.RNN, config=test_config())
runner.log(agnes.log)
runner.run()
import time
import agnes
env_name = "BreakoutNoFrameskip-v4"
if __name__ == '__main__':
env = agnes.make_vec_env(env_name,
envs_num=4,
config={"frame_stack": True})
runner = agnes.Single(env, agnes.PPO, agnes.LSTMCNN)
# runner.worker.load("examples/distributed_rnn/Breakout.pth")
# runner.trainer.load("examples/distributed_rnn/Breakout.pth")
runner.log(agnes.TensorboardLogger(".logs/"), agnes.CsvLogger(".logs/"))
runner.save_every("temp.pth", int(1e6))
runner.run()
runner.save("final.pth")
