Hacker News Reader

Comment by abetusk | Hacker News Reader

Sorry, what's not open source about them? I've only checked a few models but they look to be under a BSD-3-clause license. The first few I looked at all have the same BSD-3 license [0] [1] [2].

Are you saying they've just repacked other existing models under their own banner but haven't opened sourced some other component?

[0] https://huggingface.co/qualcomm/FFNet-40S

[1] https://huggingface.co/qualcomm/ResNet50

[2] https://huggingface.co/qualcomm/VIT

View on HN

From the link the the person you're reply to posted:

> The program must include source code, and must allow distribution in source code as well as compiled form.

The compute graph with trained weight is very much a compiled form of the model. The source code would include everything needed to train that model and reproduce it.

EDIT: wasn't fast enough

> The source code would include everything needed to train that model and reproduce it.

You know these models are trained on internet scrape which contains copyrighted content, so the dataset can't be open sourced. It's either this or bad models.

In theory, you must have written some code to train the models + download the data ... just openning this code + adding logging to store the sources trained on, you could achieve trully "open source" (anybody can now go and scrape + train the same way you did and achieve the same outcome/model)

I'm not saying "opening models is bad", it's good. However imo it would be nice to have a semantic way to differentiate between those two

And there's a lot of third party closed source code in windows that means it would be difficult for them to open source it.

They don't get to claim that it's open source just because it would be too hard to actually open source.

> The program must include source code, and must allow distribution in source code as well as compiled form.

Can you reproduce these models? if not then it's probably not open source. With a model the closest analog seems to be the training data. Is that all published?

ML training runs are not reproducible, GPUs are non-deterministic when doing large sums, the order operands are added changes the result, thread execution times also depend on caching, which is hard to predict. If you want to force deterministic mode be prepared for a huge slowdown.

A lot of work at larger orgs is put into reproducible training runs. It's about the only way to debug 'did the small parameter tweak tank performance because of a hardware failure or is there something special about that parameter?'

Reply on news.ycombinator.com