If plan to launch a pretrained model? is it possible to fine-tune or transfer-learning for tf-gnn?

[Neslihans]

Hi There!

Can you please describe in more detail what you want to accomplish? Is there a specific model you've trained that you'd want to later fine-tune or are you looking for architectures to help you transfer learn from one domain to another?

[SidneyLann]

Hi. I want a gnn pretrained model with architecture and weights like pretrained bert in pln, pretrained yolov8 model in object detect, and then use custom data to fine-tune or train some layers in the pretrained gnn model. the pretrained yolov8 model is trained for 80 classes, but I can fine-tune it with my small dataset for n classes(ie. 96 or 100 etc).

[phanein]

Yes, TF-GNN works just like normal Keras layers in this way (so you should be able to achieve fine-tuning / transfer learning by basically following this tutorial: https://keras.io/guides/transfer_learning/).

[SidneyLann]

I mean is there a gnn pretrained model provided by offical to use?

[phanein]

Sorry about the delay - no, we don't provide pretrained GNN models. (However you can use pretrained models provided by others with our framework, for example if you needed an encoder for images or text.)

[SidneyLann]

BERT, Yolov8, etc are powerfull for its weights, not only nn.