Somshubra Majumdar

NASNet-A is extremely picky about it's hyperparameter combinations. Try matching the parameters from the prebuilt models in the NASNet function. The prebut models only wrap around the actual NASNet function...

Nasnet has been used multiple times without issues, and can load weights and predict and train. I don't think there is a bug. Please post the stack trace.

I have no issues with the pr wrr adapters, please have @vadam5 review the rest of it

Either you must provide max steps to the trainer, or you have to initialize the train_ds + set the trainer so that we can resolve the max_steps. Scheduler will not...

Check the area in your logs that prints out the optimizer being loaded and near it scheduler info should also be posted or it will mention that the scheduler could...

Use exp manager to resume your jobs. If your scheduled has already hit the min lr then the previous run is over, it will not update the scheduler anymore.

I know that these things were required to make Conformer train properly in fp16, but @bmwshop can provide details of why this specific case was necessary

@bmwshop could you take a look ?

What if pycuda isnt installed at all?

Can you expand the entire stack frame and paste the entire error here. It's a Numba error when calculating the loss, it has probably little to do with spec augment