Steps for data-parallel DeepXDE

[pescap]

• DeepXDE wak scaling for tensorflow.compat.v1: (Weak) data-parallel Horovod acceleration for tensorflow.compat.v1 #1205
• Update Docker image with mpi4py: updated Dockerfile and docker image #1278
• Add parameter for scaling mode (weak or strong): Added scaling parameter #1279
• Separate the weak and strong scaling implementations: prepare strong scaling #1282
• Define custom parameters for hvd.DsitributedOptimizer: Added options for hvd.DistributedOptimzer #1285
• Implement strong scaling for tensorflow.compat.v1: Strong scaling data parallel acceleration for tensorflow.compat.v1 #1284
• Add documentation for Data Parallel DeepXDE
• Study Sobol distribution for weak scaling
• Study resampling strategies for weak scaling

[pescap]

Concerning documentation for data parallel DeepXDE, I would create a section after Demos of Operator Learning and before FAQ, do you agree @lululxvi?

The name would be: Data parallel acceleration.

[lululxvi]

Sounds good.

[pescap]

Hi! Now, pde.PDE is with data parallel sampling.

According to you @lululxvi, are there other class(Data) that shall be accelerated?

Also, it would be interesting to study the data parallel acceleration for DeepONets. Which data class or example do you suggest me to start with?

Thank you!

[lululxvi]

Maybe double check pde.TimePDE.

Yes, DeepONet parallel training is a good next step. You can simply use the demo examples.

[pescap]

Hi, how can I obtain the training points in:

deepxde/examples/operator/antiderivative_aligned.py

Line 7 in 4363618

d = np.load("antiderivative_aligned_train.npz", allow_pickle=True)

Was antiderivative_aligned_train.npz too large to be added to the repository?

[pescap]

Sorry I did not read the documentation.

As stated in the tutorial, the datasets are available here.

A few comments:

• Why not adding directly the datasets to DeepXDE?
• It would be interesting to release the code to generate the datasets.

[lululxvi]

• I want to keep the repo size small.
• It is a very simple ODE. The code can be found at https://github.com/lululxvi/deeponet

[pescap]

Hi @lululxvi, it would be very useful to add the capability for:

deepxde/deepxde/optimizers/tensorflow_compat_v1/optimizers.py

Line 59 in 47127ac

optim = hvd.DistributedOptimizer(optim)

to have other parameters. For example, Horovod has Adasum algorithm, or fp16 compression. See:

https://github.com/pescap/horovodPINNs/blob/46aa5faee24f9512d2a9e7ae136286c558d9f94b/Laplace/hvd_hval.py#LL155C1-L161C1

Are you ok if I follow a similar structure to what was used for L-BFGS (defining the parameters as an external dictionary)?

[lululxvi]

Good idea.

[pescap]

Hi @lululxvi, which is the easiest way to obtain the training domain points (to split them appropriately)?

Wouldn't it be easier to define here:

deepxde/deepxde/data/pde.py

Line 263 in 742ef4d

)

self.train_domain = X

When I apply strong scaling, I want to split the domain training points over ranks. Here:

deepxde/deepxde/data/pde.py

Line 180 in 742ef4d

self.train_x_all = self.train_points()

it sounds like they correspond to the num_domain last rows. Finally, I also have to take the anchor points into account.

[lululxvi]

Yes, but we need to do the splitting after computing the BC points.

[pescap]

Ok, I understand.

From what I observe, self.train_x_all has num_domain + num_boundary points. I am not sure to understand why. I assumed that the PDE points were defined in the interior of the domain.

[lululxvi]

Boundary points are also used for PDE loss.

[pescap]

According to me, the PDE is not valid on the boundary (only in the interior of the domain), as the operators are defined over open domains. Could you please give more details about this choice?

[lululxvi]

In general, boundary points should also satisfy the PDE. Could you give a counterexample?