How do you convert tf weights to pytorch?

[bamps53]

How do you normally convert tensorflow model's weights to pytorch equivalent model? Sometimes I want to convert some tensorflow model to pytorch or timm model back to tensorflow implement, but it's so tough and there seems to be no silver bullet. But you always convert released tensorflow pretrained weights so quickly, do you have any secret or magical script??
If you could share some insight or give practical advice, it would be very helpful!
Thanks,

[rwightman]

@bamps53 you can follow some discussion about doing it in this PR (
#554) where there is some useful pointers and questions that you'll likely run into...

I have no magic script, just a collection of really ugly scripts. It's a fiddly, detail oriented and time consuming process. I am only fast because I've done it before. When you have a template for a given 'style' of TF model it's easier to hack and slash it into a working state. First time through for a given arch / style is a slow process to note the patterns. Details like padding equivalence can be hard to track dow (PR above didn't ahve that issue). TF models are a mish mash of model level name scoping (where layers increment across whole model, irrespective of hierarchy) and local scoping (where modeller specifies their own scopes).

[bamps53]

Thanks for the answer and now I even more thank you for your tremendous effort after knowing there is no magic:)
And the PR is very useful and has many practical tips, I'll try based on this.
Thanks!