Mixture of model prediction and update about pytorch-a3c HOT 4 CLOSED

The first approach is faster on CPU in practice (no need to do a second pass, no performance gain from using batches).

However, the second one can be faster for GPU training. I implemented the second approach here:
https://github.com/ikostrikov/pytorch-a2c-ppo-acktr

from pytorch-a3c.

Thanks, that seems to be reasonable. Currently, I am using Macbook Pro with i7 processor with 4 cores and I don't have any GPU available so that's why I am trying experimenting with A3C. I will try to implement first approach. Also, is it possible to use PPO algorithm with this asynchronous settings? I didn't read about PPO much and just wonder if it is possible. Also do you think PyTorch is good fit for implementing CPU reinforcement algorithms or are there any better? I like the way how PyTorch works and is easy to use.

from pytorch-a3c.

4 cores is fine for mujoco but not enough for atari (unless you are ready to wait for several days to get results).

Synchronous methods are much easier to debug and they usually perform just as well as asynchronous methods. See the OpenAI blogpost about a2c/acktr where they mention that.

In terms of the frameworks, the difference between all the frameworks nowadays is marginal (in terms of time you need to spend in different frameworks to implement exactly the same thing). It's usually +-1-2 hours between TensorFlow and PyTorch depending on a specific task. But PyTorch tends to work slightly faster than TensorFlow, especially on CPU. So if performance is the major concern then I would recommend PyTorch.

from pytorch-a3c.

Thanks for you time, it helps a lot. I will not use Atari nor Mujoco. I am trying to implement my own environment based on descrete grid and experiment with it, so hopefully my processor is enough.

from pytorch-a3c.