support tensorflow version 2.0-2.4

[llpin1992]

when i install tfimm, it need tensorflow version <3.0,>=2.5 , please support tf version 2.0-2.4

[martinsbruveris]

It could be that I am more restrictive with the TF version than I have to be. At the same time several of the models are using grouped convolutions and those were only introduced fairly recently in TF. I will see what happens if I run the unit tests with TF 2.4 or lower.

[llpin1992]

ok, thanks, expect your results!

[martinsbruveris]

Unfortunately, there are a bunch of things that go wrong when I try TF 2.4.

• TF 2.4. does not support grouped convolutions on the CPU. A few versions before that there was also no grouped convolution support on GPUs.
• The cosine decay LR schedule is not available in TF 2.4. Well, it is, but it is part of the experimental API.
• There is a numpy error coming from K.batch_set_value(weight_value_tuples). It has something to do with a change in a numpy API. It can probably be fixed if the numpy version is set correctly. Probably...

In short, I won't be able to support TF <2.5 for tfimm officially. It should be possible to build a custom version of tfimm by changing the version requirements in pyproject.toml, and finding the right version of numpy. Then poetry build will generate a wheel that can be installed with earlier versions of TF. Models that don't use grouped convolutions will probably work. But I can't make any guarantees.