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.
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)
3. how to test
Revise the test file name in predict.py and then run the below command:
python3 predict.py
48 Dec 20, 2022
3.1k Dec 29, 2022
32 Dec 13, 2022
183 Dec 01, 2022
65 Dec 19, 2022
152 Dec 28, 2022
383 Dec 16, 2022
70 Dec 22, 2022
41 Sep 30, 2022
30 Jan 01, 2023
108 Dec 23, 2022
30 Jan 01, 2023
2 Feb 11, 2022
145 Dec 30, 2022
1 Dec 19, 2021
35 Jan 03, 2023
1 Jan 11, 2022
5 Sep 07, 2021
4 Feb 05, 2022
0 Jul 21, 2022