[eeeficus]

If you’re using pretrained models then they should run ok. It’s the same reason why Tensorflow runs on phones: to run models, not necessarily to train models. DNNs take a lot of time to train.

[missosoup]

TF runs at seconds-per-frame speeds on phones without dedicated accelerators.

Even inference is going to be slow when working on a 20 megapixel image. That's as much work as a 1220 sample batch at 128x128 resolution.