Warmup is a training technique often used in training deep neural networks. In this post, I will try to explain what is warmup, and how does it work.

Warmup was originally proposed in this paper: Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour. It gives a good explanation on why warmup is needed, and explains different strategies of warmup.

Why do we need warmup#

Suppose that we use learning rate $\eta$ on a single GPU with batch size $n$, when we train the network on 8 GPUs, now the batch size becomes $8n$. The learning rate also needs to change to suit the distributed training scenario. The author find that in practice, the linear scaling of learning rate works pretty well. For example, when we use initial learning rate 0.01 for one GPU, we may use an initial learning rate of 0.08 for distributed training, i.e., 0.01*8.

However, to use linear scaling of learning rate, certain condition have to be met¹. On the initial training stage, due to the rapid change of network parameters, the condition that makes linear scaling work does not hold any more. So, in the initial training stage, the authors propose warmup to tackle this issue.

The basic idea is that we should use a small learning rate than the value calculated by linear scaling policy. There are two strategies for warmup:

• constant: Use a low learning rate than 0.08 for the initial few epochs.
• gradual: In the first few epochs, the learning rate is set to be lower than 0.08 and increased gradually to approach 0.08 as epoch number increases. In maskrcnn, a linear warmup strategy is used for control warmup factor in the initial learning stage.

After the warmup epochs, the learning rate strategy would return to normal. You can change the learning rate based on the task at hand.

Warmup applications#

Warmup in ResNet#

In Deep residual learning, when training a 110-layer ResNet on CIFAR-10 (section 4.2), the authors used constant warmup to ease the initial training iterations:

In this case, we find that the initial learning rate of 0.1 is slightly too large to start converging. So we use 0.01 to warm up the training until the training error is below 80% (about 400 iterations), and then go back to 0.1 and continue training.

How does linear warmup work in maskrcnn#

In maskrcnn-benchmark, there is some config parameters about warmup in solver (WARMUP_FACTOR, WARMUP_ITERS, WARMUP_METHOD). The warmup method used by maskrcnn-benchmark can be found here:

def get_lr(self):
    warmup_factor = 1
    if self.last_epoch < self.warmup_iters:
        if self.warmup_method == "constant":
            warmup_factor = self.warmup_factor
        elif self.warmup_method == "linear":
            alpha = float(self.last_epoch) / self.warmup_iters
            warmup_factor = self.warmup_factor * (1 - alpha) + alpha
    return [
        base_lr
        * warmup_factor
        * self.gamma ** bisect_right(self.milestones, self.last_epoch)
        for base_lr in self.base_lrs
    ]

In the above code, self.last_epoch is the current training iteration (because maskrcnn-benchmark use iteration instead of the usual epoch to measure the training process). self.warmup_iters is the number of iterations for warmup in the initial training stage. self.warmup_factors are a constant (0.333 in this case).

Only when current iteration number is below self.warmup_iters, will the warmup_factor be used. Otherwise, it will be 1 and not affect the learning rate.

When current iteration is below warmup_iters and warmup method is linear. The warmup factor used is calculated as follows:

warmp_factor = 0.667 * (current_iter/warmup_iters) + 0.333

So as current iteration approaches warmup_iters, warmup_factor will gradually approach 1. As a result, the learning rate used will approach base learning rate.

References#

• How the learning rate change?
• Discussions about warmup policy