Thomas Capelle
Thomas Capelle
I also wrote a report on my findings: https://wandb.ai/tcapelle/apple_m1_pro/reports/Deep-Learning-on-the-M1-Pro-with-Apple-Silicon---VmlldzoxMjQ0NjY3 - I would absolutely recommend you re-run your benchmarks on NGC containers, as the performance for Tensorflow can be substantially higher.
Nice! Yeah, I ran on 128x128.
sorry, but I had many issues trying to run this, that's why I ported my own version. https://github.com/tcapelle/action_recognition
This is a pure code implementation, no experiments or training code or test. I am currently using this and `TimeSformer` for regression, you don't need to modify anything, just set...
No, you ahve to skip frames or use less. The paper deals with this issue skipping frames. (one every 15)
Lol, he already provided us with a neat and clean implementation of the arch! you can check a train code [here](https://github.com/tcapelle/action_recognition/blob/master/03_train_transformer.ipynb) using this repo's implementation of the Transformer
can you try the timesformer also, I am also having issues with STAM.
I am not sure it is possible, but ViT weights are available on `timm` and others.
You can even download a pretrained ViT from huggingface: https://huggingface.co/google/vit-large-patch16-224
yeah, variables are silent inside this type of expressions. You could call `h` or `z` and it doe snot matter.