This repository contains a minimum code to reproduce coreML issue on Mac M1. Below are the instructions to run it:

1. Clone this repository to your local machine.
2. Navigate to the directory where you cloned the repository.

Install the requirements using the following command, (if needed, create a new conda environment and activate it first):

conda create -n test_coreML python=3.11
conda activate test_coreML
pip install -r requirements.txt

1. Open a terminal or command prompt.
2. Navigate to the directory where the Python file is located.
3. Run the Python script using the following command:

4. python coreML_grid_sample.py
   

The code will convert two grid_sample layer to coreML models and save them as GridSampleModel.mlpackage and CustomGridSampleModel.mlpackage respectively in the same directory.

The GridSampleModel.mlpackage is created using the torch.nn.functional.grid_sample function and the CustomGridSampleModel.mlpackage is created using the custom grid_sample function.

Performance of GridSampleModel.mlpackage when measured on M1 it could be noticed that it does not run on Neural Engine. However, it runs on Neural Engine on M2.

We created our own grid_sample function and converted it to coreML model: CustomGridSampleModel.mlpackage it runs on GPU.

We would like to know if there is a way to convert torch.nn.functional.grid_sample to coreML model that runs on Neural Engine on M1. Or, if there is a way to convert CustomGridSampleModel.mlpackage to run on Neural Engine .