Phiflow grid to Pytorch tensor/model

[muacq]

Hi, I have encountered several problems while combining PhiFlow with Pytorch:

When combining PhiFlow with the self-defined PyTorch model, how to gracefully input PhiFlow grid data, and output grid data (with/without the same layout)?

• I tried using field.native_call(model, grid) and it worked. But how can I make sure the native call processes the grid data in the right dimension I trained (e.g. 'vector, y, x' or 'vector, x, y')?

• What if the model's output doesn't have the same layout as the input (e.g. input 'vector, y, x' output 'y, x')?

• Also, I tried converting the PhiFlow tensor to Pytorch tensor (e.g. CenterGrid velocity.values.native("vector,y,x"), then input the Pytorch tensor into the Pytorch model. I wondered if there is a graceful way to input and output grid data for the Pytorch model.

Looking forward to your reply!

[holl-]

Sorry for the late response.

• field.native_call(model, grid) is the right approach. For PyTorch, it will always use the order batch, channel, spatial.
• If your output does not have the vector/channel dimension, it should insert a singleton dimension (size 1). You can also use the channel_dim and spatial_dim arguments to describe the output if its shape varies from the input.

[muacq]

Thanks for your reply! I have another question...Where can I see all solver's names supported by PhiFlow? I just found some names in the demo code like: 'auto', 'CG', 'CG-adaptive', 'CG-native', 'direct', 'GMres', 'lGMres', 'biCG', 'CGS', 'QMR', 'GCrotMK' Is there a Jacobian or Gauss-Seidel solver?

[holl-]

Only 'CG' and 'CG-adaptive' (and soon biCG) are currently supported by all backends. If you use NumPy, you can also use the ones you mentioned but they won't work with TensorFlow or PyTorch.

[muacq]

Got it, Thanks for your reply!