medipixel / rl_algorithms Goto Github PK

TD3 network 통합

.

gym wrapper로 환경에서 state를 torch.Tensor로 만들어주는 코드 작성

I'd like to play pong with rainbow IQN algorithms, but the repository provides only dqn. How can I play RainbowIQN?

e.g. Callable[[torch.Tensor], torch.Tensor]

https://black.readthedocs.io/en/stable/the_black_code_style.html

mypy 형식으로 수정

e.g. https://github.com/openai/spinningup/blob/master/spinup/algos/sac/sac.py#L10

@all-contributors please add Curt-Park for code

Now it adjusts the action value ranges by using the first actions values' bound.
We need to consider the case that each action has a different range.

Hi. I'm just wondering why alpha get updated in 'log' space rather than in its own value.
Was doing that stable for entire training? Or any other reason?

# In https://github.com/medipixel/rl_algorithms/blob/master/algorithms/sac/agent.py

 # train alpha
        if self.hyper_params["AUTO_ENTROPY_TUNING"]:
            alpha_loss = (
                -self.log_alpha * (log_prob + self.target_entropy).detach()
            ).mean()

            self.alpha_optimizer.zero_grad()
            alpha_loss.backward()
            self.alpha_optimizer.step()

            alpha = self.log_alpha.exp()
        else:
            alpha_loss = torch.zeros(1)
            alpha = self.hyper_params["W_ENTROPY"]

algorithms/bc 디렉토리의 pkl 제거
data/ 디렉토리로 옮겼음

README.md와 각 알고리즘에 대한 문서 작성

50 --> 100 over

추상 클래스를 만들고 상속받는 방식으로 수정

Hello,
I'd like to use pretrained weight for the models (especially dqn.py for Pong), but I don't see one. If it exists, could you provide the link?

Thanks,

Hi, I run example a2c agent with lunarlander_continuous_v2 env with default hyper parameters for about 80K episode, but the total score is still about -500, Could any one help, please. By the way, I suspect it is because the reward is not normalize to mean 0.

[INFO] episode 86207    episode step: 58        total score: -549
total loss: -43826796714054.6172        policy loss: -46156932647123.8594       value loss: 2330135933069.2412

We need check performace TD3 algorithm because not as good as before.

TD3: @MrSyee
SAC: @Curt-Park

TD3: @MrSyee
SAC: @Curt-Park

Does this program use Python 3.6? If so, why does it use NumPy for Python 2.7 (CPython 2.7 - cp27)?

15:27:33|peter@laptop:[rl_algorithms]> make dep
pip install -r requirements.txt
Collecting numpy==1.16.2 (from -r requirements.txt (line 1))
Using cached https://files.pythonhosted.org/packages/c4/33/8ec8dcdb4ede5d453047bbdbd01916dbaccdb63e98bba60989718f5f0876/numpy-1.16.2-cp27-cp27mu-manylinux1_x86_64.whl

lgtm is recommended.

pylint 추가