davidsandberg / facenet

Face recognition using Tensorflow
MIT License
13.71k stars 4.8k forks source link

A new model trained on VGGFace2 #624

Open Shahnawazgrewal opened 6 years ago

Shahnawazgrewal commented 6 years ago

I trained a model on VGGFace2 using center loss. The embedding is powerful than the subset of MS-Celeb. I can make the model public with the two available mode. @davidsandberg

phoebushe commented 6 years ago

@syy6 hello,i also have the problem(OutOfRangeError). How can i do to solve it ?

syy6 commented 6 years ago

@phoebushe , I just checked if there is any broken images for the input and I find one image is broken.

phoebushe commented 6 years ago

@syy6 , VGGFACE2 dataset have over 300 million images, could u tell me how to check the images?Thx!!!!!

syy6 commented 6 years ago

@phoebushe , just use Python, code snip as below. `from PIL import Image, ExifTags

image=Image.open(file_path)

image.verify()`

phoebushe commented 6 years ago

@syy6 Thanks a lot!!

Honghe commented 6 years ago

@syy6 image.verify() is the best way to check up to now.

hungnv21292 commented 6 years ago

hi all, In this repo, @davidsandberg use LWF with LWF dataset and pairs.txt file. How to evaluate with your dataset (have use pairs.txt file, ...) when you fine-tune Asian-face dataset? I donn't know how to evaluate for new dataset (I fine tune for Asian-face dataset)? Thank you

hwpengTristin commented 5 years ago

Q1: What is your learning rate, learning rate decay epochs and learning rate decay epochs in the experiment using center loss? I find that the default settings of these hyperparameters are 0.1, 100 and 1.0, respectively. Did you modify them? If yes, can you share your modify value?

Q2: Have you tried training a initial inception-resnet-v1 with triplet loss using CASIA WebFace, MS-Celeb-1M or VGGFace2 ? I wonder if the triplet loss can well pre-train the inception-resnet-v1 and obtain better performance in LFW dataset in both rank-1 accuracy and the verification rates?

fisakhan commented 5 years ago

How to train on images of size 100,100 rather than 160,160 ?

zebi51 commented 5 years ago

Hi @Shahnawazgrewal , we are very much impressed with the performance of you model. Thank you.

btomtom5 commented 4 years ago

@Shahnawazgrewal Hey thanks for releasing your model! However, it looks like the link is broken. Did you change the location of the link by any chance? Thanks in advance!

adarsh0906 commented 4 years ago

can some one tell me what computer spec your using to train the model using this dataset ,or the cloud spec if trained in cloud thanks in advance .I am beginner in computer vision and trying to train a model of my own from scratch using this dataset , the model similar to inception model used in facenet

kiad4631 commented 4 years ago

Hi @Shahnawazgrewal Do you know how can I fine-tune this classifier for my dataset?

sonfire186 commented 4 years ago

@Shahnawazgrewal link doesn`t work

peererror commented 3 years ago

Link broken