Transfer Learning Toolkit 2.0

HI,
I was looking at the release notes for Transfer Learning Toolkit 2.0(Developer Preview) and there are some questions which I wish to ask -
a) Is retraining from checkpoints supported or still only pretrained models can be used?
b) Is custom online augmentations supported ?
c) Is the previous pretrained DetectNetV2 resnet10 .hdf5 model has been upgraded/changed ?
d) Also, can the TLT trained models only be deployed with Deepstream5.0 or they support 4.0 also.
e) Lastly, where can I find resource to help me cleanly install the developer preview alongside the already existing TLT installation i.e I wish to test the Developer Preview but also wish to keep the previous TLT version.

Hi,
a) Retraining from checkpoints is supported in 2.0 docker.
b) Custom online augmentations is not supported.
c) Yes.
d) The etlt models are only compatible with DS5.0
e) Yes, you can. Because the dockers are different. You can just download the tlt 2.0 docker and generate a new folder to keep all the training results, etc.

Hi,
Thanks for the quick reply. Due to certain limitations, I can’t employ DS5.0 on my nano(since it requires an upgradation of jetpack). But I am very much interested in using the functionalities of TLT2.0. Is there a way to deploy the TLT2.0 retrained pruned model with Deepstream 4?

I am specifically concerned with DetectnetV2 resnet10 model.

I am currently also trying to do the same, are you not able to update the jetpack to 4.4?

Hi neophyte1,
Suggest you to use Jetpack 4.4 to install new version of TRT7, cudnn8, DS5 and CUDA10.2 remotely on your nano.

Hi Morganh,
Can you point out to some resources regarding this - i.e to cleanly achieve the same without any name or library clashes ? Should I also be uninstalling the previous versions of the libraries manually or will it be taken care of ?

Hi neophyte1,
I modify my previous comment. Suggest to flash the new OS/image when your install the new version of TRT7, cudnn8, DS5 and CUDA10.2.

For links of SDK manager and Jetpack 4.4 DP,
Please search it in https://developer.nvidia.com/embedded/downloads
How to install SDK manager: https://docs.nvidia.com/sdk-manager/download-run-sdkm/index.html
How to install Jetpack 4.4 DP with sdk-manager: https://docs.nvidia.com/sdk-manager/install-with-sdkm-jetson/index.html

@Morganh

You say that the .etlt model files from TLT 2.0 are only compatible with DS5.0. If we take the .etlt model from TLT2.0 and then create a trt engine file using a Nano with DS5.0, will that trt engine file work on a Nano using DS4.0?

It should be not working. Because when you build a trt engine with tlt-converter-7.1, the trt engine should be not working at a device whose trt version is not 7.1.