Faceie: A Pure-Numpy Face Recognition Library

Faceie is an inference-only, CPU-only, pure Numpy implementation of MTCNN face detection and an Inception-ResNet-v1 based, FaceNet-style face encoder. Both components are intended to be weight-compatible with FaceNet-PyTorch.

Some more detailed notes (and references) on the two models can be found here:

• MTCNN (Face Detection)
• Inception-ResNet-v1/FaceNet (Face Embedding)

Usage

>>> from faceie import detect_faces, encode_faces

>>> from PIL import Image
>>> image = image.load("path/to/image.jpg")

>>> # Use MTCNN to find faces in an image (the returned named tuple
>>> # includes probabilities, bounding boxes, facial landmark coordinates
>>> # and thumbnails of each face by default).
>>> faces = detect_faces(image)

>>> # Produce a 512-dimensional embedding for each detected face
>>> embeddings = encode_faces(faces.images)

>>> # Compare detected faces against a target face using squared-distance
>>> # metric. Smaller distances (around 1 or less) imply the same face.
>>> import numpy as np
>>> target = image.load("path/to/target.jpg")
>>> target_embedding = encode_faces(detect_faces(target).images[0])
>>> square_dists = np.sum(np.power(embeddings - target_embedding, 2), axis=1)

Weights

This implementation is weight-compatible with FaceNet-PyTorch but uses its own file formats (primarily to avoid the PyTorch dependency).

The MTCNN model weights are very small (a couple of megabytes) and are stored included with Faceie's source.

The Inception-ResNet-v2 based FaceNet-style network model weights are larger (around one hundred megabytes) and are distributed separately. By default, Faceie will automatically download and use a copy of the VGGFace-trained weights (preconverted into Faceie's format) from the corresponding Faceie GitHub release. (See weightie for the weight storage format and auto-download mechanism.).

You can also use the faceie-convert-facenet-weights script to convert a PyTorch weights file from the reference FaceNet-PyTorch implementation's format into the native weights format:

$ pip install path/to/faceie[convert]  # Extra packages needed for weights file conversion
$ faceie-convert-facenet-weights /path/to/weights.pt output.weights

The conversion script requires extra packages to be installed in order to unpack the PyTorch weights file format (including PyTorch). After conversion, these dependencies are no longer required.

If using a custom weights file, you can provide the path to the converted weights file (as a Path object) to faceie.facenet.load to load the needed weights and pass them to the encoder functions:

>>> # Load the weights...
>>> from pathlib import Path
>>> from faceie.facenet import load
>>> weights = load(Path("/path/to/converted.weights"))

>>> # Use them...
>>> embeddings = encode_faces(..., weights)

Preemptive FAQ

Why does Faceie exist?

As with Clippie, I am gradually building my own search facilities for my own photo collection which doesn't depend on 3rd party services (e.g. Google Photos). The MTCNN and Inception-ResNet-v1 FaceNet networks hit a sweet spot:

• Both are fairly well documented in the literature
• Both produce fairly robust detections/embeddings respectively. Compared with much faster popular models and algorithms, this combination does a good job at handling differing facial poses, orientations, lighting and so on. Compared with the current state of the art, they're not that far behind. In any case, in my casual experimentation, they seem to be generally as good as or better than the functions built into popular photo management services and software.
• Both run "well enough" on a CPU to be practical for a personal photo collection.
• Both have readily available pre-trained models available under permissive licenses. Specifically FaceNet-PyTorch includes both!

As for the obvious follow-up question...

Why not use FaceNet-PyTorch?

By contrast with FaceNet-PyTorch, Faceie has only a few comparatively light-weight and stable dependencies (chiefly Numpy and Pillow). As such, the largest download needed is a copy of the weights, not gigabytes of software. Furthermore, unlike most deep learning libraries -- which cater to a fast moving field -- all of the dependencies used have been stable and well supported for many years and are likely to remain so for many more years.

With that said, FaceNet-PyTorch is a really cool project with a track record of active development and ongoing improvements. It is around twice as fast as Faceie run on my CPU and dramatically faster running on a GPU -- which Faceie doesn't support. If you're not me, you should probably use FaceNet-PyTorch!

Last but not least, re-implementing things like this is a good way for me to get to know how they work since I'm entirely new to the field!

Why CPU only?

Faceie can detect a set of faces in a photograph and produce 512 embeddings for them in a second or two on my laptop. This is adequately fast for my use. Even at this (quite pitiful) speed, processing my lifetimes' back-catalogue is just a one-off few days of CPU time. In any case, if speed was a real concern, I'd definitely be looking at other solutions.

Faceie's Numpy based implementation runs approximately twice as slowly as the PyTorch-based implementation on a CPU. This is obviously a bit disappointing (I see no reason why this should be the case) but I'm insufficiently motivated to dig deeper.

Why inference only?

Since I'm only interested in using these models, and published weights work well, I had no need. I'm also especially keen to avoid falling down the rabbit hole of model training...

Does Faceie include any clustering algorithms?

No, this is out of scope for now.

Does this reuse any code from FaceNet-PyTorch?

No -- though it does use its model weights.

This software is a from-scratch implementation of MTCNN and Inception-ResNet-v1/FaceNet primarily based on their respective papers. However, for the sake of weight compatibility, some parts necessarily mimic the behaviour of FaceNet-PyTorch.

Whilst no code is reused, Faceie directly uses the pretrained weights distributed by FaceNet-PyTorch under the MIT License:

• The (converted) MTCNN model weights are included in this repository in faceie/mtcnn/data

• The (converted) Inception-ResNet-v1/FaceNet weights are included alongside GitHub releases of Faceie.