How should I put model into distributed devices if I have multiple optimization loops?

[NoTody]

Hi, guys

I'm trying to incorporate average weight perturbation (https://proceedings.neurips.cc/paper/2020/file/1ef91c212e30e14bf125e9374262401f-Paper.pdf) to my lightning code. However, because it calculates adversary weight perturbation by another optimization loop, the model is not put into devices correctly in distributed setting. How should I do this correctly?

Specifically, how should I incorporate the perturb function in this link when I have another model need to be trained in train_loop?

https://github.com/csdongxian/AWP/blob/main/AT_AWP/awp.py

Right now, I did something like this, but it doesn't work out.

During initialization in def init(self, hparams, backbone):

self.model = backbone
if hparams.use_awp:
   self.proxy = copy.deepcopy(self.model)
   proxy_opt = torch.optim.SGD(self.proxy.parameters(), lr=0.01)
   self.awp_adversary = awp.AdvWeightPerturb(proxy=self.proxy,
                                      proxy_optim=proxy_opt,
                                      gamma=1e-2)

During training in train_loop

    if stage == "train":
        images_adv = self.adversary(images_pair, labels.repeat(2))
    else:
        with torch.enable_grad():
            images_adv = self.adversary(images_pair, labels.repeat(2))
    if stage == "train" and self.hparams.awp:
      self.awp = self.awp_adversary.calc_awp(model=self.model,
                                inputs_adv=images_adv,
                                targets=labels.repeat(2))
      self.awp_adversary.perturb(self.model, self.awp)
    # register hook to get intermediate output
    activation = {}
    def get_activation(name):
        def hook(model, input, output):
            activation[name] = output
        return hook
    self.model.relu.register_forward_hook(get_activation('penultimate'))
    # get original output
    outputs_adv = self.model(images_adv)
    # get latent
    outputs_latent = activation['penultimate']
    outputs_latent = F.avg_pool2d(outputs_latent, 8)
    outputs_latent = outputs_latent.view(outputs_latent.size(0), -1)
    loss_ce = self.criterion(outputs_adv, labels.repeat(2))

With the error shown as:

RuntimeError: Expected to have finished reduction in the prior iteration before starting a new one. This error indicates that your module has parameters that were not used in producing loss. You can enable unused parameter detection by passing the keyword argument `find_unused_parameters=True` to `torch.nn.parallel.DistributedDataParallel`, and by
making sure all `forward` function outputs participate in calculating loss.
If you already have done the above, then the distributed data parallel module wasn't able to locate the output tensors in the return value of your module's `forward` function. Please include the loss function and the structure of the return value of `forward` of your module when reporting this issue (e.g. list, dict, iterable).