Can A3C share model in multiple GPU?

[Jie_Han_Chen]

Accroding to original A3C, there exist 1 shared global model and multiple local models corresponded to each worker process. A3C is designed for multi-core CPU for parallel processing with poor GPU usage.

With little modification like model.cuda() we can run the model on GPU to speed up training.

[Question]
If we have 2 GPUs in computer. Can we run the A3C just without modification to share the global model between each worker?

[dgriff]

See post here:

Trying to modify this code to use cuda

Hey Checkout my repo here I implemented a3c with gpu use and hog wild training. I call it A3G and it’s works very well it’s actually even faster at training Atari games than mass parallelized evolutionary algorithms with also much higher overall performance. update Just realized you had already visited repo and we spoke on issues on github. Guess you can disregard then lol

It will be slower if you share gpu to gpu. This way is faster cause you get benefits from gpu and cpu

[Jie_Han_Chen]

Thank you

[Jie_Han_Chen]

Hi, @dgriff:

In your A3G, we put the shared model in CPU. Why don’t we put the shared model in GPU?

Thank you.

[dgriff]

cause cpu is faster for sharing model cause we can do “hogwild training”, sharing without locks asynchronous updates. The act of acquiring locks and releasing locks is actually quite time consuming. This is a benefit of CPU has over GPU. GPU requires locks for sharing in almost all cases except for atomic operations