No assignment for “self.grads”

[onlyyao]

Thanks for your good job!
I don’t find any assignment for “self.grads” in RotateOnly so the “self.grads” remains None when running _rep_grad, is there any solution？

[adrianjav]

Hi @onlyyao

self.grads is assigned in a backward_hook in each task-specific RotateModule:

rotograd/rotograd/rotograd.py

Lines 123 to 124 in 7259503

	def hook(self, g):
	self.p.grads[self.item] = g.clone()

These RotateModules are added to your architecture here:

rotograd/rotograd/rotograd.py

Lines 197 to 198 in 7259503

	for i in range(num_tasks):
	heads[i] = nn.Sequential(RotateModule(self, i), heads[i])

Are you using the model as described in the readme file? That is, using preds = model(x) to get your predictions?