Switching out Optimizer after a certain amount of epochs

Is it possible to use multiple optimizers to train a model?
For example, I would like to start training with the Adam optimizer and then at a certain point continue training with SGD. How would I correctly do this in pytorch?

Would something like this work?

optimizer = torch.optim.Adam(model.parameters(), lr=0.0004)

for e in range(epochs):
    for i, data in enumerate(dataloader()):
        output = model(data)
        loss = criterion(output, label)

        if e == 100:
             optimizer = torch.optim.SGD(model.parameters(), lr=0.002, momentum=0.9)

There is an optimizer called AdaBound, which uses adam and SGD together.

optimizer = AdaBound(model.parameters(), lr=1e-3, final_lr=0.1)

class AdaBound(Optimizer):
    """Implements AdaBound algorithm.

    It has been proposed in `Adaptive Gradient Methods with Dynamic Bound of Learning Rate`_.


        params (iterable): iterable of parameters to optimize or dicts defining

            parameter groups

        lr (float, optional): Adam learning rate (default: 1e-3)

        betas (Tuple[float, float], optional): coefficients used for computing

            running averages of gradient and its square (default: (0.9, 0.999))

        final_lr (float, optional): final (SGD) learning rate (default: 0.1)

        gamma (float, optional): convergence speed of the bound functions (default: 1e-3)

        eps (float, optional): term added to the denominator to improve

            numerical stability (default: 1e-8)

        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)

        amsbound (boolean, optional): whether to use the AMSBound variant of this algorithm

    .. Adaptive Gradient Methods with Dynamic Bound of Learning Rate:



    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), final_lr=0.1, gamma=1e-3,

                 eps=1e-8, weight_decay=0, amsbound=False):

        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))

        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))

        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))

        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))

        if not 0.0 <= final_lr:
            raise ValueError("Invalid final learning rate: {}".format(final_lr))

        if not 0.0 <= gamma < 1.0:
            raise ValueError("Invalid gamma parameter: {}".format(gamma))

        defaults = dict(lr=lr, betas=betas, final_lr=final_lr, gamma=gamma, eps=eps,

                        weight_decay=weight_decay, amsbound=amsbound)

        super(AdaBound, self).__init__(params, defaults)

        self.base_lrs = list(map(lambda group: group['lr'], self.param_groups))

    def __setstate__(self, state):

        super(AdaBound, self).__setstate__(state)

        for group in self.param_groups:
            group.setdefault('amsbound', False)

    def step(self, closure=None):

        """Performs a single optimization step.


            closure (callable, optional): A closure that reevaluates the model

                and returns the loss.


        loss = None

        if closure is not None:
            loss = closure()

        for group, base_lr in zip(self.param_groups, self.base_lrs):

            for p in group['params']:

                if p.grad is None:

                grad = p.grad.data

                if grad.is_sparse:
                    raise RuntimeError(

                        'Adam does not support sparse gradients, please consider SparseAdam instead')

                amsbound = group['amsbound']

                state = self.state[p]

                # State initialization

                if len(state) == 0:

                    state['step'] = 0

                    # Exponential moving average of gradient values

                    state['exp_avg'] = torch.zeros_like(p.data)

                    # Exponential moving average of squared gradient values

                    state['exp_avg_sq'] = torch.zeros_like(p.data)

                    if amsbound:
                        # Maintains max of all exp. moving avg. of sq. grad. values

                        state['max_exp_avg_sq'] = torch.zeros_like(p.data)

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']

                if amsbound:
                    max_exp_avg_sq = state['max_exp_avg_sq']

                beta1, beta2 = group['betas']

                state['step'] += 1

                if group['weight_decay'] != 0:
                    grad = grad.add(group['weight_decay'], p.data)

                # Decay the first and second moment running average coefficient

                exp_avg.mul_(beta1).add_(1 - beta1, grad)

                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)

                if amsbound:

                    # Maintains the maximum of all 2nd moment running avg. till now

                    torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)

                    # Use the max. for normalizing running avg. of gradient

                    denom = max_exp_avg_sq.sqrt().add_(group['eps'])


                    denom = exp_avg_sq.sqrt().add_(group['eps'])

                bias_correction1 = 1 - beta1 ** state['step']

                bias_correction2 = 1 - beta2 ** state['step']

                step_size = group['lr'] * math.sqrt(bias_correction2) / bias_correction1

                # Applies bounds on actual learning rate

                # lr_scheduler cannot affect final_lr, this is a workaround to apply lr decay

                final_lr = group['final_lr'] * group['lr'] / base_lr

                lower_bound = final_lr * (1 - 1 / (group['gamma'] * state['step'] + 1))

                upper_bound = final_lr * (1 + 1 / (group['gamma'] * state['step']))

                step_size = torch.full_like(denom, step_size)

                step_size.div_(denom).clamp_(lower_bound, upper_bound).mul_(exp_avg)


        return loss

1 Like