khrlimam / res-facenet

Pretrained Facenet model

ResFacenet

This package is intended as a pytorch hub entry point for my trained facenet model on this repo khrlimam/facenet. This pretrained model can be used for anyone who want to use it for transfer learning or any other applications.

This model trained on VGGv2 (used the test set one) dataset and tested on LFW dataset and gained 92% accuracy.

This work is distributed under MIT license, so I hope you consider the license's policy. Cheers!

How to use

1. First thing first! Please install Pytorch framework by following the installation guide that suites your OS distribution. Refer to this link https://pytorch.org. After you successfully installed the pytorch framework then you're ready to continue to the next step.
2. Install the package pip install res-facenet
3. the pretrained model can be used with:

from res_facenet.models import model_920, model_921
model920 = model_920()
model921 = model_921()

There're two model you can use first model_920 that is the model with 92% accuracy and model_921 that is model with 92.135% accuracy. After instantiating the model you are free to use the pretrained weight for transfer learning or just to test the model. To use this model for inference you can follow the instructions below:

import torch.nn.functional as F
from torchvision import transforms
from PIL import Image

#prepare preprocess pipeline
preprocess_pipelines = [transforms.Resize(224), 
                       transforms.CenterCrop(224), 
                       transforms.ToTensor(), 
                       transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                                            std=[0.229, 0.224, 0.225])]
trfrm = transforms.Compose(preprocess_pipelines)
topil = transforms.ToPILImage()
totensor = transforms.Compose(preprocess_pipelines[:-1])

# read the image and transform it into tensor then normalize it with our trfrm function pipeline
img1 = trfrm(Image.open(img1_path)).unsqueeze(0)
img2 = trfrm(Image.open(img2_path)).unsqueeze(0)

# do forward pass
embed1, embed2 = model920(img1), model920(img2)

# compute the distance using euclidean distance of image embeddings
distance = F.pairwise_distance(embed1, embed2)

# we use 1.5 threshold to decide whether images are genuine or impostor

threshold = 1.5

genuine = distance <= threshold

The demo app can be found at this repo: http://github.com/khrlimam/demo-facenet

res-facenet's Issues

Proper summary of the project is Missing

You haven't mentioned about input size and other things required for using this pretrained model

clarification regarding the training set

Hi,
thanks for your work. i am a bit confused about the statement "This model trained on VGGv2 (used the test set one) dataset and tested on LFW dataset and gained 92% accuracy"
Does that mean the model was trained on VGG Face2 test set?