Skip to content

Latest commit

 

History

History
146 lines (94 loc) · 3.27 KB

README.md

File metadata and controls

146 lines (94 loc) · 3.27 KB

anecc

anecc, short for "ANE CoreML converter/compiler", cuts the middleman to let you directly run a CoreML MLModel on the Neural Engine with my reverse-engineered Linux driver.

Installation

From pip:

pip install anecc

From source:

git clone https://github.com/eiln/anecc.git
cd anecc
make install

Verify installation:

Usage: anecc [OPTIONS] PATH

Options:
  -n, --name TEXT  Model name
  -o, --out TEXT   Output file name (default $name.anec)
  -w, --write      Write anec.
  -p, --print      Print struct.
  -f, --force      Bypass warnings.
  --help           Show this message and exit.

What It Does

On macOS, an MLModel file would be loaded with:

import coremltools as ct
model = ct.models.MLModel("yolov5.mlmodel")

anecc converts the MLModel into a custom anec format that's even easier to load on Linux:

import ane
model = ane.model("yolov5.anec")

You'll notice load times are a lot faster than CoreML. And the Python "library" being a mere 50 LOC wrapper around the lightweight gnu99 C userspace driver library. That's because you actually own your own model.

#include "ane.h"

int main(void) {
	struct ane_nn *nn = ane_init("yolov5.anec");
	if (nn == NULL) {
		return -1;
	}
	// ...
	ane_free(nn);
	return 0;
}

Compile with gcc or g++:

g++ -I/usr/include/libane main.c -o main -lane

CoreML Conversion

To create or convert your own MLModel, you need to start in macOS, where a CoreML runtime is available. You do not need macOS to run a pre-converted ".anec" model.

TLDR at the bottom.

Step 0. obtain MLModel

Starting in macOS-land. You should have a "*.mlmodel" file ready. Do not use "mlpackages".

See coreml101.md.

Step 1. tohwx: Compile mlmodel -> hwx

Still in macOS-land.

I kinda lied about the MLModel part. When an MLModel is loaded,

mlmodel = ct.models.MLModel("yolov5.mlmodel")

the CoreML backend is called to deserialize, convert, and compile the model into a microcode executable on the ANE. Every. Time. I'm unsure whether the concept of "caching" is absent out of apathy, inability, or secrecy. The compiled model, which we must extract, is embedded in a macho suffixed "*.hwx". My guess is "hardware executable".

Extract the hwx under macOS using tohwx, a slimmed down version of freedomtan's obj-c conversion script. Install tohwx with:

git clone https://github.com/eiln/anecc.git
cd anecc
make -C tohwx
make -C tohwx install

Use tohwx to convert mlmodel -> hwx:

$ tohwx yolov5.mlmodel
tohwx: input mlmodel: /Users/eileen/yolov5.mlmodel
tohwx: output hwx: /Users/eileen/yolov5.hwx

If it fails, sorry, I don't know. Make a pull, attach the mlmodel, and I'll take a look at it.

Step 2. anecc: convert hwx -> anec

anecc is platform independent. You can also use it to check conversion ability or debug the model.

anecc yolov5.hwx -o yolov5.anec

TLDR

In macOS:

tohwx yolov5.mlmodel

In macOS or Linux or Windows or whatever:

anecc yolov5.hwx -o yolov5.anec

Use the yolov5.anec with libane.