face-property-detection-pytorch

1. Data structure

The structure of landmarks_jpg is like below:

|--celeba1
|----celeba_face
|------000001.jpg
|------000002.jpg
|------ .....
|------020000.jpg
|----celeba_raw_pic
|------000001.jpg
|------000002.jpg
|------ .....
|------020000.jpg

The celeba_raw_pic is the original picture that we do not make any processing.
The celeba_face is the face region of the raw pricture.

figure1: raw picture

figure2: face region of raw picture

You can run the below command to finish the data processing.

python3 create_data.py

This command will use MTCNN model to extract the face region. However, some pictures cannot be extracted by the model.
For my test, I can not cut out the face region of the below picture.

# file 000199.jpg cannot detect face
# file 001401.jpg cannot detect face
# file 002214.jpg cannot detect face
# file 002432.jpg cannot detect face
# file 002920.jpg cannot detect face
# file 003928.jpg cannot detect face
# file 003946.jpg cannot detect face
# file 004932.jpg cannot detect face
# file 005283.jpg cannot detect face
# file 006010.jpg cannot detect face
# file 006531.jpg cannot detect face
# file 007726.jpg cannot detect face
# file 008287.jpg cannot detect face
# file 011529.jpg cannot detect face
# file 011793.jpg cannot detect face
# file 013374.jpg cannot detect face
# file 013654.jpg cannot detect face
# file 014999.jpg cannot detect face
# file 016530.jpg cannot detect face
# file 016797.jpg cannot detect face
# file 017282.jpg cannot detect face
# file 017586.jpg cannot detect face
# file 018309.jpg cannot detect face
# file 018599.jpg cannot detect face
# file 018884.jpg cannot detect face
# file 019205.jpg cannot detect face
# file 019377.jpg cannot detect face

So I replace them with 000001.jpg.
Also, I revise the label file list_attr_celeba.txt. Replace the issue items with 000001.jpg and I get the
list_attr_celeba-face.txt
You can use BeyondCompare to diff the changes that I make

You can download the data from the cloud drive:

name	link
celeba_face.zip	https://pan.baidu.com/s/15nsbvla8eCy_n3EsUMH36Q code：5ipn
celeba_raw_pic.zip	https://pan.baidu.com/s/1WM3Zo3zLfKsAFvrDl03suQ code：3q70

2. how to train

First, install the third-party package:

pip install -r requirements.txt

Then just simply run the below command:

python3 train.py

if you want to use the pretrained models, you can revise the below code as you need:

load_pretrain_model = False
model_dir=r".\pretrain_models\model-resnet-50-justface-state.ptn"
if load_pretrain_model:
    checkpoint = torch.load(model_dir)
    model.load_state_dict(checkpoint)