Any tutorial for customizing the model architecture before re-training?

Hey, great thanks for the awesome pre-trained models. I have some customized datasets with different data structures that cannot be directly applied or converted to the designated data format of the provided pre-trained models. Namely, I need to change the input and output layers of the pre-trained models before conducting transfer learning. I failed to find references for how to retrieve the original model structure details or how to customize the model structures. Any help with this matter will be very much appreciated!

Which network will you focus on?

Hi @Morganh, thanks a lot for getting back to me! Iā€™m trying to use the UNET networks for semantic segmentation, I know that there are resnet10/resnet18/resnet34/resnet50/resnet101/vgg16/vgg19 available, and I will most likely use vgg19/16 and resnet101/50 for my project.

For Unet, currently, there are some parameters which can be specified. Refer to UNET ā€” TAO Toolkit 3.22.05 documentation

BTW, next TAO release will support converting UNet open-source ONNX model to a TAO-compatible model.

Hi @Morganh, thank you for the information! Do you happen to know a rough time when the next TAO release will be available?

It will release in June.

1 Like

Topic closed as new version released: Release Notes ā€” TAO Toolkit 3.22.05 documentation

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.