louis-she/center-loss.pytorch

center loss for face recognition

你好，I'm Chenglu 👋

🦄 Status

• 🔭 I’m currently working on featurize.cn
• 🌱 I’m currently learning Computer Vision
• 👯 I’m looking to collaborate on Web Full Stack Development/Computer Vision
• 🤔 I’m looking for help with PyTorch Tao
• 📫 How to reach me: [email protected]

🏁 Kaggle

I solo at Kaggle competitions randomly.

center-loss.pytorch's Issues

Questions About the model

Hi, thanks for your pytorch implementation of the center face .
I played with your code for a while, and I may found some problem.

1. lfw dataset, It seems that you use the whole picture instead of using the face detection bounding boxes or just using center cropped images for training and testing, take a look at some pictures in the dataset,

by this way, too much background are envoled.

2. overlap between training set and test set.

    for klass, name in enumerate(names):
        def add_class(image):
            image_path = os.path.join(images_root, name, image)
            return (image_path, klass, name)

        images_of_person = os.listdir(os.path.join(images_root, name))
        total = len(images_of_person)

        training_set += map(
                add_class,
                images_of_person[:ceil(total * train_val_split)])
        validation_set += map(
                add_class,
                images_of_person[floor(total * train_val_split):])

I think you should use samples listed in the pairsDevTrain.txt to train and pairsDevTest.txt to test by the lfw paper

how to use the lfw dataset

thanks for your codes!

i have a question that what is the protocol do you use in your training,

there are
PAIRS_TRAIN = "http://vis-www.cs.umass.edu/lfw/pairsDevTrain.txt"
PAIRS_VAL = "http://vis-www.cs.umass.edu/lfw/pairsDevTest.txt"
in the codes,but doesnot use.

Evaluate

Hi, thanks for your work, i want to ask you a question, how to do the test? I tested the resnet50 and epoch_100.pth.tar on lfw, with crop face, but the accuracy is 0.66. How can I reproduce this result? thank you.

how can i test the l2 distance between one and many picture?

RuntimeError: CUDA out of memory. Tried to allocate 12.00 MiB (GPU 0; 7.77 GiB total capacity; 5.54 GiB already allocated; 7.56 MiB free; 35.91 MiB cached)

how to

can not tar -xvf the pretrained model.

can not tar -xvf the pretrained model.

Is it necessary to normalize features?

Is it necessary to normalize features?

Why separate parameters?

Hello, I am a beginner. So maybe my question is simple. Please forgive me.
In the main.py. The code:

    model = model_class(num_classes).to(device)

    trainables_wo_bn = [param for name, param in model.named_parameters() if
                        param.requires_grad and 'bn' not in name]
    trainables_only_bn = [param for name, param in model.named_parameters() if
                          param.requires_grad and 'bn' in name]

    optimizer = torch.optim.SGD([
        {'params': trainables_wo_bn, 'weight_decay': 0.0001},
        {'params': trainables_only_bn}
    ], lr=args.lr, momentum=0.9)

I want to know why you separate the parameters to two parts: trainables_wo_bn and trainables_only_bn,
In the optimizer, I always write:
optimizer = torch.optim.SGD(model.parameters(), lr = args.lr, momentum=0.9)
Did I do something wrong? Please help me, thank you so much!

pretrain.model can not open, an error

Does this implementation require landmarks of dataset?

Could you please let me know if you are generating landmarks for lfw? I am currently using a dataset with low resolution images for which I am unable to generate landmarks (using MTCNN), so I'm looking for an implementation that isn't using landmarks..

如何生成各个类别的特征分布图呢？