PyTorch's grid_sample conversion to CoreML (via coremltools)

torch.nn.functional.grid_sample (source here, click on docs for documentation) is currently unsupported operation by CoreML (and their conversion utilities library: coremltools).

What I'm looking for is a way to export layer shown below from PyTorch's torchscript (docs here) to CoreML (either using custom op created via Swift or via efficient PyTorch rewrite of grid_sample).

For details and tips to get you started see Tips section

Minimal verifiable example

import coremltools as ct
import torch


class GridSample(torch.nn.Module):
    def forward(self, inputs, grid):
        # Rest could be the default behaviour, e.g. bilinear
        return torch.nn.functional.grid_sample(inputs, grid, align_corners=True)


# Image could also have more in_channels, different dimension etc.,
# for example (2, 32, 64, 64)
image = torch.randn(2, 3, 32, 32)  # (batch, in_channels, width, height)
grid = torch.randint(low=-1, high=2, size=(2, 64, 64, 2)).float()

layer = GridSample()
# You could use `torch.jit.script` if preferable
scripted = torch.jit.trace(layer, (image, grid))

# Sanity check
print(scripted(image, grid).shape)


# Error during conversion
coreml_layer = ct.converters.convert(
    scripted,
    source="pytorch",
    inputs=[
        ct.TensorType(name="image", shape=image.shape),
        ct.TensorType(name="grid", shape=grid.shape),
    ],
)

which raises the following error:

Traceback (most recent call last):
  File "/home/REDACTED/Downloads/sample.py", line 23, in <module>
    coreml_layer = ct.converters.convert(
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/_converters_entry.py", line 175, in convert
    mlmodel = mil_convert(
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/converter.py", line 128, in mil_convert
    proto = mil_convert_to_proto(, convert_from, convert_to,
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/converter.py", line 171, in mil_convert_to_proto
    prog = frontend_converter(, **kwargs)
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/converter.py", line 85, in __call__
    return load(*args, **kwargs)
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/frontend/torch/load.py", line 81, in load
    raise e
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/frontend/torch/load.py", line 73, in load
    prog = converter.convert()
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/frontend/torch/converter.py", line 227, in convert
    convert_nodes(self.context, self.graph)
  File "/home/REDACTED/.conda/envs/REDACTED/lib/python3.9/site-packages/coremltools/converters/mil/frontend/torch/ops.py", line 54, in convert_nodes
    raise RuntimeError(
RuntimeError: PyTorch convert function for op 'grid_sampler' not implemented.

Dependencies

Python (conda):

• coremltools==4.1
• torch==1.8.0

You could also use nightly/master builds (at least for the day of writing: 2021-03-20)

Tips

Those were split into two possible solutions I currently see:

PyTorch only

Rewrite torch.nn.functional.grid_sample from scratch.

• This would require sticking only to PyTorch operations on tensors as loops (e.g. triple nested) would hang the converter and be too inefficient
• You cannot use __getitem__ on list or related types - seems to work with torch.Tensor but had problems with that so you should have it in mind if you get RuntimeError: PyTorch convert function for op '__getitem__' not implemented

Pros:

• No need for two languages & sticking to single technology

Cons:

• Limited with loops and would require sticking to vectorized operations (most/all of the time)

Swift & CoreML

Register custom layer which is responsible for running grid_sample. CPU only implementation would be fine (although using Apple's Metal for GPU speedups would be great).

As I'm not into Swift, I've gathered a few resources which might help you:

• https://coremltools.readme.io/docs/custom-operators - starting point, Python only, quite easy, only registering layer for conversion
• https://developer.apple.com/documentation/coreml/mlcustomlayer - API of the layer which one would have to code in Swift
• https://developer.apple.com/documentation/coreml/core_ml_api/creating_a_custom_layer - more about aforementioned (but not much)
• https://machinethink.net/blog/coreml-custom-layers/ - blog post with example and dispatching layer to devices (GPU, CPU). Needs Swift (CPU version), Metal (GPU implementation). Eventual Metal implementation might be based off PyTorch's CUDA impl, CPU and Swift might also be related. 3 years old, so be aware of that, swish activation layer, seems to be a good starting point (and other posts from the same author cast some light on the CoreML itself).
• https://github.com/hollance/CoreML-Custom-Layers - repo for the above

Pros:

• Possibility to use loops and finer control over the algorithm
• Might be easier as we're not limited to operations which CoreML can currently read

Cons:

• Two languages
• Sparse documentation

from PyTorch's grid_sample conversion to CoreML (via coremltools)