mathLab
DeepONet tutorial for Advection Problem
In this PR I uploaded a new tutorial for the DeepONet class since it was missing.
When writing the tutorial I noticed some bugs which I corrected. Here I report a list of the bugs:
- The aggregation (
MIONet._aggregator) had the problem that the input was already stacked, but sometimes it is not possible to stack (e.g.net1output shape[T, N, d],net2output shape[T, d]and aggregation performseinsum('TNd,Td->TNd', ...)). So to avoid this problem now theMIONet._aggregatorfunction takes as input a tuple of tensors (also more simple for the user to understand). - I replaced the
.reshapewith.unsqueezeinoutput_ = self._reduction(aggregated).reshape(-1, 1)since the final aim of the operation is to add an extra dimension, and not reshape the tensor.
Finally, I remove the check consistency of the network outputs before aggregation and put a warning message that the consistency is up to the user to check. This is not because I believe it should be up to the user the check, but it is a permanent fix since the way we were checking the consistency is wrong. Indeed consider what we were doing:
# check trunk branch nets consistency
shapes = []
for key, value in networks.items():
check_consistency(value, (str, int))
check_consistency(key, torch.nn.Module)
input_ = torch.rand(10, len(value))
shapes.append(key(input_).shape[-1])
if not all(map(lambda x: x == shapes[0], shapes)):
raise ValueError('The passed networks have not the same '
'output dimension.')
This is wrong if the user wants to specify a custom network whose inputs is not in the form [N, len(value)]. For example, consider the following net:
class Net(torch.nn.Module):
def __init__(self):
super().__init__()
self.layer = torch.nn.Linear(2, 100)
def forward(self, x):
x_ = x[:, 0, :]
return self.layer(x_)
Due to the input reshape the line shapes.append(key(input_).shape[-1]) will raise an error since we are trying to slice a two dimensional tensor (input_ = torch.rand(10, len(value))) like a three dimensional one (line x[:, 0, :]).
The problem to check network consistency is more general than just MIONet or DeepONet, as also in FNO we have the same problem. Maybe we can think to make a PR just for this problem.
The comments are made because the checks are just correct for 2D inputs. I commented because it is not still clear what is the data format for all models (2d/3d?).