Quantization aware training resume with checkpoint

[aab60526]

Hello, when I training a resuming model with checkpoint,
It raise an error

And this is my command line

Where is the error come from?

[guyjacob]

It's not 100% clear from the command line, but I assume you're trying to resume training from a previous quantization-aware training session. Unfortunately, that's not supported at the moment (see comment here).
(In case you want to load a quantization-aware training checkpoint just for evaluation, you can use the workaround detailed here)

We're currently working on fixing some other issues related to loading checkpoints (see #182), once that's finalized I'll try to get quantization-aware training resume soon after.

[parap1uie-s]

Maybe, you could try is:

1. Follow the steps here to hack the distiller framework.

2. Create two model

resume_model = create_model(<your params>)
new_model = create_model(<your params same as resume model>)

3. Resume quantized weights to resume model

    if args.resume:
        resume_model, _, _ = apputils.load_checkpoint(resume_model, chkpt_file=args.resume)
        resume_model.to(args.device)

4. Quantize the new model

    if args.compress:
        compression_scheduler = distiller.file_config(new_model, optimizer, args.compress, compression_scheduler)

Now, you have a resume model which contains quantized trained weights, and a new model contains random initial quantized weights.

5. Load weights

new_model.load_state_dict(copy.deepcopy(resume_model.state_dict()))
del(resume_model)

[thnguyen996]

Hi, does anyone figure out any quick way to fix this bug? I really need to run evaluation on quantized model