Incompatible with latest version of tensor flow etc about starnet HOT 10 CLOSED

from starnet.

Working on it, but currently stuck at the step of transferring pre-trained weights to the new version

from starnet.

The problem is specifically with TF - old weights output complete garbage in the new version. Seems that Conv2D in TF 2.x is somehow different and I can not find out what is the problem.

I asked a friend of mine who has done quite a bit of TF work, and he said:

You might need to store the weights in an external format for transfer, if you can store the network from 1.4 in Keras or ONNX format, you'll be able to read that cleanly in tf 2.0

Don't know if that helps at all?

from starnet.

Amazing, thanks so much @nekitmm.

from starnet.

from starnet.

TF 2.0 is completely different, you need to install 1.4.0 or similar.

I am thinking of updating it actually, but never got time to do that so far... I need at some point though

from starnet.

I was looking into effort to do this since i'd like to try the DIrectML for Windows to see if DX cards can offload some startnet++ in PixInsight. Have you had any time to look into 2.x yet?

from starnet.

@nekitmm, first of all, many thanks for making startnet available. It's invaluable! I don't know much about tensorflow, but I write c++/python for a living, so please let me know if I can help with any of the other aspects of the update at all.

from starnet.

@nekitmm, first of all, many thanks for making startnet available. It's invaluable! I don't know much about tensorflow, but I write c++/python for a living, so please let me know if I can help with any of the other aspects of the update at all.

Thanks! I am ok on python code and everything - already wrote the whole thing. The problem is specifically with TF - old weights output complete garbage in the new version. Seems that Conv2D in TF 2.x is somehow different and I can not find out what is the problem.

from starnet.

I finally pushed a new version made in TF2. You can check it out!

from starnet.