Fix a bug of forward-mode AD when multi-output is needed #1925
Conversation
Update example of heat equation (lululxvi#706)
Add document for Lorenz inverse with exogenous input (lululxvi#709)
OperatorPredictor supports backends tensorflow.compat.v1, tensorflow,…
… method of Lr decay in Pytorch
update to latest version
|
Here, we compute the dim deepxde/deepxde/gradients/jacobian.py Line 31 in b0d239b Would this a problem? |
|
What you mentioned is indeed a problem, the dim computed is wrong. This error has nothing to do with this pull request. It is about the last Forward-mode AD pull request. Since we changed the forward-mode AD to loop-free, then the dim computed is wrong. I have used another way to compute the dim of |
| if bkd.ndim(ys[0]) == 2: | ||
| self.dim_y = 1 | ||
| elif bkd.ndim(ys[0]) == 3: | ||
| self.dim_y = ys[0].shape[2] |
There was a problem hiding this comment.
How about using
self.dim_y = ys[0].shape[-1]
There was a problem hiding this comment.
you mean using self.dim_y = ys[0].shape[-1]when ndim of ys[0] is 3?
deepxde/data/pde_operator.py
Outdated
| losses_bc = [bkd.reduce_mean(bkd.stack(loss, 0)) for loss in losses_bc] | ||
| losses.append(losses_bc) | ||
| for loss in losses_bc: | ||
| losses.append(bkd.reduce_mean(bkd.stack(loss, 0))) |
There was a problem hiding this comment.
How about
losses_bc = [bkd.reduce_mean(bkd.stack(loss, 0)) for loss in losses_bc]
losses.extend(losses_bc)
I find the bug from this case: https://github.com/lululxvi/deepxde/blob/master/examples/operator/stokes_aligned_pideeponet.py
when setting the
num_output=3andmulti_output_strategy="independent", the output shape will be (batch size, # of coordinates, 3).Then, the slice here is wrong, what we really want is slicing in the last dim.