In DifferentiableAdam, sqrt() is non-differentiable at zero

When using higher with Adam as the inner optimizer, calling the outer loss.backward() sometimes raises the following torch error:

RuntimeError: Function 'SqrtBackward' returned nan values in its 0th output.

The problem occurs when the exp_avg_sq tensor in DifferentiableAdam contains a zero, in which case exp_avg_sq.sqrt() is non-differentiable.

The problem disappears when I add a tiny value to exp_avg_sq before applying the sqrt():

exp_avg_sq = exp_avg_sq + 1e-16

But I don’t know whether this would cause other problems.

Is the _maybe_mask function designed to deal with zeros in exp_avg_sq?

I’m using the latest pip-installed version (higher=0.2.1).

Forgot to mention, anomaly detection needs to be enabled to see the runtime error:

torch.autograd.set_detect_anomaly(True)

@rickyloynd-microsoft

@rickyloynd-microsoft I'm facing the same problem. Did you solve it without adding a small value?

Yes, the solution above has been working for me.

Although I'm having another issue now with Adam's internal state not getting persisted (and detached) between rollouts (#114).

Yes, the solution above has been working for me.

Can you tell me how to solve the problem?

@rickyloynd-microsoft

@rickyloynd-microsoft I'm facing the same problem. Did you solve it without adding a small value?
@zhiqihuang 老哥，你解决这个问题了么？我也遇到了这个问题

Navigation Menu