aab3902d4a7d55f5a86058854adc36b8a12c873f,catalyst/dl/callbacks/base.py,OptimizerCallback,on_batch_end,#OptimizerCallback#Any#,202

Before Change

        else:
            model = state.model
            model.zero_grad()
            optimizer = state.get_key(
                key="optimizer", inner_key=self.optimizer_key
            )
            loss = state.get_key(key="loss", inner_key=self.optimizer_key)
            scaled_loss = self.fp16_grad_scale * loss.float()
            scaled_loss.backward()

            master_params = list(optimizer.param_groups[0]["params"])
            model_params = list(
                filter(lambda p: p.requires_grad, model.parameters())
            )
            copy_grads(source=model_params, target=master_params)
            for param in master_params:
                param.grad.data.mul_(1. / self.fp16_grad_scale)
            self.grad_step(
                optimizer=optimizer,
                optimizer_wd=self._optimizer_wd,
                grad_clip_fn=self.grad_clip_fn

After Change

    def on_batch_end(self, state):
        loss = state.get_key(key="loss", inner_key=self.loss_key)
        if isinstance(loss, dict):
            loss = list(loss.values())
        if isinstance(loss, list):
            loss = torch.mean(torch.stack(loss))

        if self.prefix is not None:

Instances