Please provide the following information when requesting support.
• Network Type: Mobilenetv2
• TLT Version: TAO v5.5.0
We are carrying out training, pruning and retraining Mobilenetv2 model (converted to .tltb format) in TAO BYOM. We are able to get checkpoints after training in .hdf5 format rather than .tlt format. Is there any changes with current version of TAO(v5.5.0) for BYOM.
Also, we see in few posts about “BYOM might be deprecated in future” - Does this mean we cannot use any BYOM commands of TAO v5.5.0 or v5.3.0 ?
What should users could do who are dependent on TAO BYOM approach? Please suggest if there are any alternatives ?
Yes, please refer to Bring Your Own Model (BYOM) - NVIDIA Docs.
No, there is not any change.
BYOM is aim to convert 3rd-party open-source ONNX model to a TAO-compatible model. BYOM will be deprecated because since TAO5.0, the TAO source code is open-sourced, and onnx file is also possible to generate during TAO deployment.
Thanks for your time.
Please confirm on the below statement:
We are able to get checkpoints after training in .hdf5 format rather than .tlt format. Is it expected in BYOM to get model checkpoint after training in .tlt or .hdf5 format ?
BYOM is aim to convert 3rd-party open-source ONNX model to a TAO-compatible model. So, the only important step is to convert 3rd-party open-source ONNX model to .tltb model. Then trigger TAO training. After training, the coming pipeline is the same as non-BYOM pipeline.
Since TAO5.0, it is expected for you to get .hdf5 format after training. The .tlt is actually an encryption of .hdf5 file.
Please note, .tltb is not related to .tlt.
Thanks for your response. You may close this ticket