by timesler

Pretrained Pytorch face detection (MTCNN) and recognition (InceptionResnet) models

1.5K Stars 301 Forks Last release: 22 days ago (v2.4.1) MIT License 226 Commits 6 Releases

Available items

1 Support Hour

Developer services

Email/Chat Support

Ask me anything

Source changes not included

Bug Fix

Developer services

Bug Fix

Fast resolution

Contact me before purchase

Bug Fix - Minor

Developer services

Minor Bug Fix

Fast resolution

Contact me before purchase

5 Support Hours

Developer services

Email/Chat Support

Ask me anything

Source changes not included


Developer services

Integration help

Custom code optional

Contact me before purchase

Custom Feature

Developer services

Custom feature

Contact me before purchase


Face Recognition Using Pytorch


Code Coverage

| Python | 3.7 | 3.6 | 3.5 | | :---: | :---: | :---: | :---: | | Status | Build Status | Build Status | Build Status |


This is a repository for Inception Resnet (V1) models in pytorch, pretrained on VGGFace2 and CASIA-Webface.

Pytorch model weights were initialized using parameters ported from David Sandberg's tensorflow facenet repo.

Also included in this repo is an efficient pytorch implementation of MTCNN for face detection prior to inference. These models are also pretrained. To our knowledge, this is the fastest MTCNN implementation available.

Table of contents

Quick start

  1. Install:

    # With pip:
    pip install facenet-pytorch

    or clone this repo, removing the '-' to allow python imports:

    git clone https://github.com/timesler/facenet-pytorch.git facenet_pytorch

    or use a docker container (see https://github.com/timesler/docker-jupyter-dl-gpu):

    docker run -it --rm timesler/jupyter-dl-gpu pip install facenet-pytorch && ipython

  2. In python, import facenet-pytorch and instantiate models:

    from facenet_pytorch import MTCNN, InceptionResnetV1

    If required, create a face detection pipeline using MTCNN:

    mtcnn = MTCNN(image_size=, margin=)

    Create an inception resnet (in eval mode):

    resnet = InceptionResnetV1(pretrained='vggface2').eval()

  3. Process an image:

    from PIL import Image

    img = Image.open()

    Get cropped and prewhitened image tensor

    img_cropped = mtcnn(img, save_path=)

    Calculate embedding (unsqueeze to add batch dimension)

    img_embedding = resnet(img_cropped.unsqueeze(0))

    Or, if using for VGGFace2 classification

    resnet.classify = True img_probs = resnet(img_cropped.unsqueeze(0))


for usage and implementation details.

Pretrained models

See: models/inceptionresnetv1.py

The following models have been ported to pytorch (with links to download pytorch state_dict's):

|Model name|LFW accuracy (as listed here)|Training dataset| | :- | :-: | -: | |20180408-102900 (111MB)|0.9905|CASIA-Webface| |20180402-114759 (107MB)|0.9965|VGGFace2|

There is no need to manually download the pretrained state_dict's; they are downloaded automatically on model instantiation and cached for future use in the torch cache. To use an Inception Resnet (V1) model for facial recognition/identification in pytorch, use:

from facenet_pytorch import InceptionResnetV1

For a model pretrained on VGGFace2

model = InceptionResnetV1(pretrained='vggface2').eval()

For a model pretrained on CASIA-Webface

model = InceptionResnetV1(pretrained='casia-webface').eval()

For an untrained model with 100 classes

model = InceptionResnetV1(num_classes=100).eval()

For an untrained 1001-class classifier

model = InceptionResnetV1(classify=True, num_classes=1001).eval()

Both pretrained models were trained on 160x160 px images, so will perform best if applied to images resized to this shape. For best results, images should also be cropped to the face using MTCNN (see below).

By default, the above models will return 512-dimensional embeddings of images. To enable classification instead, either pass

to the model constructor, or you can set the object attribute afterwards with
model.classify = True
. For VGGFace2, the pretrained model will output logit vectors of length 8631, and for CASIA-Webface logit vectors of length 10575.

Example notebooks

Complete detection and recognition pipeline

Face recognition can be easily applied to raw images by first detecting faces using MTCNN before calculating embedding or probabilities using an Inception Resnet model. The example code at examples/infer.ipynb provides a complete example pipeline utilizing datasets, dataloaders, and optional GPU processing.

Face tracking in video streams

MTCNN can be used to build a face tracking system (using the

method). A full face tracking example can be found at examples/face_tracking.ipynb.

Finetuning pretrained models with new data

In most situations, the best way to implement face recognition is to use the pretrained models directly, with either a clustering algorithm or a simple distance metrics to determine the identity of a face. However, if finetuning is required (i.e., if you want to select identity based on the model's output logits), an example can be found at examples/finetune.ipynb.

Guide to MTCNN in facenet-pytorch

This guide demonstrates the functionality of the MTCNN module. Topics covered are:

  • Basic usage
  • Image normalization
  • Face margins
  • Multiple faces in a single image
  • Batched detection
  • Bounding boxes and facial landmarks
  • Saving face datasets

See the notebook on kaggle.

Performance comparison of face detection packages

This notebook demonstrates the use of three face detection packages:

  1. facenet-pytorch
  2. mtcnn
  3. dlib

Each package is tested for its speed in detecting the faces in a set of 300 images (all frames from one video), with GPU support enabled. Performance is based on Kaggle's P100 notebook kernel. Results are summarized below.

|Package|FPS (1080x1920)|FPS (720x1280)|FPS (540x960)| |---|---|---|---| |facenet-pytorch|12.97|20.32|25.50| |facenet-pytorch (non-batched)|9.75|14.81|19.68| |dlib|3.80|8.39|14.53| |mtcnn|3.04|5.70|8.23|

See the notebook on kaggle.

The FastMTCNN algorithm

This algorithm demonstrates how to achieve extremely efficient face detection specifically in videos, by taking advantage of similarities between adjacent frames.

See the notebook on kaggle.

Running with docker

The package and any of the example notebooks can be run with docker (or nvidia-docker) using:

docker run --rm -p 8888:8888
    -v ./facenet-pytorch:/home/jovyan timesler/jupyter-dl-gpu \
    -v :/home/jovyan/data
    pip install facenet-pytorch && jupyter lab 

Navigate to the examples/ directory and run any of the ipython notebooks.

See timesler/jupyter-dl-gpu for docker container details.

Use this repo in your own git project

To use this code in your own git repo, I recommend first adding this repo as a submodule. Note that the dash ('-') in the repo name should be removed when cloning as a submodule as it will break python when importing:

git submodule add https://github.com/timesler/facenet-pytorch.git facenet_pytorch

Alternatively, the code can be installed as a package using pip:

pip install facenet-pytorch

Conversion of parameters from Tensorflow to Pytorch

See: models/utils/tensorflow2pytorch.py

Note that this functionality is not needed to use the models in this repo, which depend only on the saved pytorch


Following instantiation of the pytorch model, each layer's weights were loaded from equivalent layers in the pretrained tensorflow models from davidsandberg/facenet.

The equivalence of the outputs from the original tensorflow models and the pytorch-ported models have been tested and are identical:

>>> compare_model_outputs(mdl, sess, torch.randn(5, 160, 160, 3).detach())
Passing test data through TF model

tensor([[-0.0142, 0.0615, 0.0057, ..., 0.0497, 0.0375, -0.0838], [-0.0139, 0.0611, 0.0054, ..., 0.0472, 0.0343, -0.0850], [-0.0238, 0.0619, 0.0124, ..., 0.0598, 0.0334, -0.0852], [-0.0089, 0.0548, 0.0032, ..., 0.0506, 0.0337, -0.0881], [-0.0173, 0.0630, -0.0042, ..., 0.0487, 0.0295, -0.0791]])

Passing test data through PT model

tensor([[-0.0142, 0.0615, 0.0057, ..., 0.0497, 0.0375, -0.0838], [-0.0139, 0.0611, 0.0054, ..., 0.0472, 0.0343, -0.0850], [-0.0238, 0.0619, 0.0124, ..., 0.0598, 0.0334, -0.0852], [-0.0089, 0.0548, 0.0032, ..., 0.0506, 0.0337, -0.0881], [-0.0173, 0.0630, -0.0042, ..., 0.0487, 0.0295, -0.0791]], grad_fn=)

Distance 1.2874517096861382e-06

In order to re-run the conversion of tensorflow parameters into the pytorch model, ensure you clone this repo with submodules, as the davidsandberg/facenet repo is included as a submodule and parts of it are required for the conversion.


  1. David Sandberg's facenet repo: https://github.com/davidsandberg/facenet

  2. F. Schroff, D. Kalenichenko, J. Philbin. FaceNet: A Unified Embedding for Face Recognition and Clustering, arXiv:1503.03832, 2015. PDF

  3. Q. Cao, L. Shen, W. Xie, O. M. Parkhi, A. Zisserman. VGGFace2: A dataset for recognising face across pose and age, International Conference on Automatic Face and Gesture Recognition, 2018. PDF

  4. D. Yi, Z. Lei, S. Liao and S. Z. Li. CASIAWebface: Learning Face Representation from Scratch, arXiv:1411.7923, 2014. PDF

  5. K. Zhang, Z. Zhang, Z. Li and Y. Qiao. Joint Face Detection and Alignment Using Multitask Cascaded Convolutional Networks, IEEE Signal Processing Letters, 2016. PDF

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.