Reproduction of memory leak for CoreML inference on iOS device

This repository contains a simple example to reproduce the memory leak observed when running inference on a CoreML model on an iOS device (specifically tested on the iPhone12). To reproduce the memory leak, you must first generate a model by converting a PyTorch model to a CoreML MLModel, and then build and launch the model in a mobile application using XCode.

The bug report submitted to coremltools is available here.

This example assumes you have installed and loaded the following environment listed.

System environment:

• coremltools version: 5.0b5:
• OS: build on MacOS targetting iOS for mobile application:
• macOS version: Big Sur (version 11.4)
• iOS version: 14.7.1 (run on iPhone 12)
• XCode version: Version 12.5.1 (12E507)
• How you install python: Install from source
• python version: 3.8.10
• How you install Pytorch: Install from source
• PyTorch version: 1.8.1

Generate CoreML MLModel

A simple model involving three convolution and one pixel-shuffle layers is used to reproduce the memory leak. This PyTorch model is converted to a CoreML MLModel so that it can be used in the iOS application.

The CoreML model is built by converting a simple PyTorch model coreml-memory-leak/convert_torch_to_coreml.py contains a simple PyTorch model and a function to convert the Torch model to a CoreML MLModel. - Simply run this script to produce a CoreML model expecting input tensor shape [1,48,480,270] and outputs a tensor of shape [1,3,3840,2160]

```python
python convert_torch_to_coreml.py
```
- This produces the model `toy.mlmodel`.

Run mobile application

The iOS application TestApp can be used to launch the generated model on an iOS device. The application was taken and adapted from PyTorch's source TestApp example. This application should be built and run using XCode. You may need to modify the input arguments and for the application, which can be set as follows:

• input_type : float
• coreml_model : true
• input_dims : 1,48,480,270
• iter : 50
• warmup : 1

You will also need to move the MLModel to coreml-memory-leak/TestApp/TestApp/model.pt.