edgetpu-yolo VS transformers

I'm on the fence. It's a very nice device if you can get your models working on it - basically untouched at the price/power point. Drivers for me have been OK. I have an M.2 card connected to a Jetson devkit (makes for a nice embedded test bench) and it runs fine, no worse than the NCS for setup anyway. There were a couple of PCI settings to tweak but I documented the setup here [0]. For common use cases it's a decent option, I think. For custom models you really need to know what you're doing.

The main issue I've had is that the compiler behaviour differs between versions (and it's very difficult to find older releases), so where previously you could run a big model and delegate things to the CPU, now it sometimes won't compile at all. There were also problems where we trained a model in AutoML - using free credits but the real cost would have been over $100 - but edgetpu compiled model lost a lot of performance. The developers have been very helpful when I've contacted them, and generally you can get through to real devs (not generic support) who can look at your model for you. Mostly I think you need to take care when training models for these devices, but quantisation-aware training is not trivial to use in Tensorflow and there are only a few off-the-shelf models which are supported in the various toolkits. Model maker looks promising, but it's also finnicky in my experience [1].

I'm not super worried about hardware availability. They're suffering from the chip shortage like everyone else, so it's not surprising that lead times are long. I was able to buy my device in late 2020 without any trouble.

[0] https://github.com/jveitchmichaelis/edgetpu-yolo/blob/main/h...

Is t5x an encoder/decoder architecture?

Some more general options.

The Flax ecosystem

https://github.com/google/flax?tab=readme-ov-file

or dm-haiku

https://github.com/google-deepmind/dm-haiku

were some of the best developed communities in the Jax AI field

Perhaps the “trax” repo? https://github.com/google/trax

Some HF examples https://github.com/huggingface/transformers/tree/main/exampl...

Sadly it seems much of the work is proprietary these days, but one example could be Grok-1, if you customize the details. https://github.com/xai-org/grok-1/blob/main/run.py

The HuggingFace transformers library already has support for a similar method called prompt lookup decoding that uses the existing context to generate an ngram model: https://github.com/huggingface/transformers/issues/27722

I don't think it would be that hard to switch it out for a pretrained ngram model.

Most of this tutorial is based on Hugging Face course about Transformers and on Niels Rogge's Transformers tutorials: make sure to check their work and give them a star on GitHub, if you please ❤️

* Superconvergence + LR range finder + Fast AI's Ranger21 optimizer was the goto optimizer for CNNs, and worked fabulously well, but on transformers, the learning rate range finder sadi 1e-3 was the best, whilst 1e-5 was better. However, the 1 cycle learning rate stuck. https://github.com/huggingface/transformers/issues/16013

Thanks! :) I'm pushing them into transformers, pytorch-gemma and collabing with the Gemma team to resolve all the issues :)

The RoPE fix should already be in transformers 4.38.2: https://github.com/huggingface/transformers/pull/29285

My main PR for transformers which fixes most of the issues (some still left): https://github.com/huggingface/transformers/pull/29402

If you want to tinker with the architecture Hugging Face has a FOSS implementation in transformers: https://github.com/huggingface/transformers/blob/main/src/tr...

If you want to reproduce the training pipeline, you couldn't do that even if you wanted to because you don't have access to thousands of A100s.

I am aware that using mixed precision reduces the stability of weight and there will be little consistency but don't expect it to be this much. I have attached the graph of evaluation metrics. If someone can give me some insight into this issue, that would be great.