def step()

in timm/optim/rmsprop_tf.py [0:0]

• 59 lines of code
• 18 McCabe index (conditional complexity)

    def step(self, closure=None):
        """Performs a single optimization step.

        Arguments:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError('RMSprop does not support sparse gradients')
                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    state['square_avg'] = torch.ones_like(p)  # PyTorch inits to zero
                    if group['momentum'] > 0:
                        state['momentum_buffer'] = torch.zeros_like(p)
                    if group['centered']:
                        state['grad_avg'] = torch.zeros_like(p)

                square_avg = state['square_avg']
                one_minus_alpha = 1. - group['alpha']

                state['step'] += 1

                if group['weight_decay'] != 0:
                    if group['decoupled_decay']:
                        if group['corrected_weight_decay']:
                            wd_scale = group['lr'] ** 2 / self.defaults['lr']
                        else:
                            wd_scale = group['lr']
                        p.mul_(1. - wd_scale * group['weight_decay'])
                    else:
                        grad = grad.add(p, alpha=group['weight_decay'])

                # Tensorflow order of ops for updating squared avg
                square_avg.add_(grad.pow(2) - square_avg, alpha=one_minus_alpha)
                # square_avg.mul_(alpha).addcmul_(grad, grad, value=1 - alpha)  # PyTorch original

                if group['centered']:
                    grad_avg = state['grad_avg']
                    grad_avg.add_(grad - grad_avg, alpha=one_minus_alpha)
                    avg = square_avg.addcmul(grad_avg, grad_avg, value=-1).add(group['eps']).sqrt_()  # eps in sqrt
                    # grad_avg.mul_(alpha).add_(grad, alpha=1 - alpha)  # PyTorch original
                else:
                    avg = square_avg.add(group['eps']).sqrt_()  # eps moved in sqrt

                if group['momentum'] > 0:
                    buf = state['momentum_buffer']
                    buf.mul_(group['momentum'])

                    def _apply_caution(_m, _g):
                        # Apply caution as per 'Cautious Optimizers' - https://arxiv.org/abs/2411.16085
                        mask = (_m * _g > 0).to(_g.dtype)
                        mask.div_(mask.mean().clamp_(min=1e-3))
                        return _m * mask

                    if group['lr_in_momentum']:
                        # Tensorflow accumulates the LR scaling in the momentum buffer
                        buf.addcdiv_(grad, avg, value=group['lr'])
                        if group['caution']:
                            buf = _apply_caution(buf, grad)
                        p.add_(-buf)
                    else:
                        # PyTorch scales the param update by LR
                        buf.addcdiv_(grad, avg)
                        if group['caution']:
                            buf = _apply_caution(buf, grad)
                        p.add_(buf, alpha=-group['lr'])
                else:
                    p.addcdiv_(grad, avg, value=-group['lr'])

        return loss