dde.gradients support 3D outputs #1928
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
Support Python 3.12
|
This PR should only modify dde.gradients, not pde_operator.py. |
update to latest version
|
You need to test all backends. |
The only case where the output is 3D is shown to be supported only by tensorflow v2 and pytorch |
|
Not necessary testing PI-DeepONet. You can manually construct a function, and jus test if dde.gradients works correctly. |
I have tested pytorch, tf v2, paddle, jax. I don't have access to the environment for tf v1 |
deepxde/gradients/gradients.py
Outdated
| When `ys` has shape (batch_size, dim_y), the output shape is (batch_size, 1). | ||
| When `ys` has shape (batch_size_out, batch_size, dim_y), the output shape is | ||
| (batch_size_out, batch_size, 1) if forward-mode autodiff is used or | ||
| (batch_size, 1) if reverse-mode autodiff is used. |
There was a problem hiding this comment.
For reverse-mode, if it is (batch_size, 1), then does it mean that reverse mode doesn't support batch_size_out?
There was a problem hiding this comment.
reverse mode does support batch_size_out. In reverse mode, the gradient has the same shape as the input since it uses vector jacobian product. so when the ys has the shape (batch_size, dim_y), the shape of gradient is (batch_size, 1) no matter reverse_mode or forward-mode is used.
There was a problem hiding this comment.
I don't understand. If revise-mode supports batch_size_out, then why the output is (batch_size, 1)? Where is batch_size_out?
There was a problem hiding this comment.
reverse-mode support ys shape has the shape (batch_size_out, batch_size, dim_y) and xs has the shape (batch_size, dim_x). That's this PR care about, 3D output.
No description provided.