Google seemed to have been released a tpu which have been working on their server for picture searching. Is it possible to use their Dev board(edge-TPU) to train the faceswap modle?
It's it possible to use edge-TPU to train a modle?
- jhsy1209021
- Posts: 4
- Joined: Thu Feb 27, 2020 2:26 pm
- bryanlyon
- Site Admin
- Posts: 793
- Joined: Fri Jul 12, 2019 12:49 am
- Location: San Francisco
- Has thanked: 4 times
- Been thanked: 218 times
- Contact:
Re: It's it possible to use edge-TPU to train a modle?
No, a TPU isn't usable, especially the edge TPU which is only made for inference and can't train.
- jhsy1209021
- Posts: 4
- Joined: Thu Feb 27, 2020 2:26 pm
- PooOnBirbs
- Posts: 4
- Joined: Thu Oct 07, 2021 6:41 pm
Re: It's it possible to use edge-TPU to train a modle?
I was about to drop ~$50 to slap an M.2 to PCI card + install one of those Coral (Google) Edge TPU's into my rig specifically for FaceSwap... until I read this. So, outside of GPUs, are there ANY plug-and-play TPU's available that would actually be able to train in FaceSwap? Obviously, I would need to find one that supports actual TensorFlow (not TensorFlow LITE). But could you point me in the right direction? Any suggestions?
- bryanlyon
- Site Admin
- Posts: 793
- Joined: Fri Jul 12, 2019 12:49 am
- Location: San Francisco
- Has thanked: 4 times
- Been thanked: 218 times
- Contact:
Re: It's it possible to use edge-TPU to train a modle?
No, we only support GPUs. TPUs, NPUs, the Apple M1, and any other "neural accelerators" are not supported. Most of those devices support inference ONLY and wont work for training at all. The only one we'd consider supporting is the M1 but that's hardware we can't justify purchasing and can't support without owning one.