Method _log_grad_norm

improved_diffusion/train_util.py:251–255

(self)

            update_ema(params, self.master_params, rate=rate)

    def _log_grad_norm(self):
        sqsum = 0.0
        for p in self.master_params:
            sqsum += (p.grad ** 2).sum().item()
        logger.logkv_mean("grad_norm", np.sqrt(sqsum))

    def _anneal_lr(self):
        if not self.lr_anneal_steps: