error report when implementing GEM on Tox21 datasets

QueuQ / CGLB

Other

45 stars 13 forks source link

error report when implementing GEM on Tox21 datasets #21

Closed ghost closed 1 year ago

ghost commented 1 year ago

I encountered the error while I implementing GEM on Tox21-tIL in TaskIncremental setting.

ValueError: Input contains NaN, infinity or a value too large for dtype('float32').

The parameters of the model were nan so consequently logits were also nan. This error does not occur in other settings and datasets, but only Tox21 one.

ghost commented 1 year ago

The problem is solved. I missed the code " store_grad(self.net.parameters, self.grads, self.grad_dims, old_task_i)" in observe function.