Any example about using microsoft/nni with NVIDIA modulus?

[Zhao-ZC]

Or is there any example of auto-tuning the hyperparameters of the net?

What makes me confused is how to report_intermediate_result.

I can not find a method to report the loss during training.

Is modulus capable of some tricks, like calling the callback function every 1000 steps?

Thank you very much.

[ngeneva]

Hi @Zhao-ZC

If you want to add a custom callback inside the training loop you can add it in the trainer.py file. For example, if you want to record the losses at line 546, the losses for that iteration are stored in a dictionary losses which can be used to save to file.

# Report losses every 100 iterations in trainer.py
if step % 100 == 0:
    for name, loss in losses.items():
         print(name, loss)

[Zhao-ZC]

Thank you very much!