Data parallelism is a way to process multiple data batches across multiple devices simultaneously to achieve better performance. In PyTorch, the DistributedSampler ensures each device gets a non-overlapping input batch. The model is replicated on all the devices; each replica calculates gradients and simultaneously synchronizes with the others using the ring all-reduce algorithm(梯度在設備間環狀傳遞、求和、更新).

Gym 是 OpenAI 做的 RL tasks 模拟库 (虽然 Play Atari w/ Deep RL 等都是 Deepmind 发的)，比如 CartPole, Robot Locomotion, Atari 2600 都有模拟。除了 Python 外在 C++ 也实现了此 API。
其中 ALE (Arcade-Learning-Environment) 是 Atari 2600 part，stable-baselines3 在这个模拟库的逻辑之上实现 RL 经典算法。

两年没看论文，实在难以专注，写笔记更有劲。

文章里的大部分术语都能在cs231n筆記里找到解释。

在文獻裡 Molecular Generation 和 Molecule Generation 這倆說法各占一半。