This repo is for the adavanced training on deeping learning with the purpose of building a face attributes analysis application.
- CelebA: about 162770 train images and about 39829 test images. （with precropped images, but it's better to try wild images.）
- LFWA: (todo...because the original dataset link is invalid...)
- v1:resnet101 (self-defined fc + BCE loss with logits + average accuracy)
- v2:resnet101 (self-defined fc + BCE loss with logits + threshold rescaling (
from page 67,Machine Learning, Zhihua Zhou) + average accracy + more metrics (recall, precision, TPR, FPR, F1from page 30-33,Machine Learning, Zhihua Zhou)) - v3+v5: try GC_resnet101 after modifying label format and loss.
- v4: resnet101 + focal loss.
- v6: se_resnet101
- v7: resnet152
- v8: densenet121
- v9: SGE_resnet101.
- V10: SK_resnet101.
Some model's codes are from PytorchInsight and pytorch.torchvision.Thx!
| Solution | average accuracy(%) | macro-precision(%) | macro-recall(%) | macro-F1(%) |
|---|---|---|---|---|
| Resnet101 | 91.53 | 79.81 | 63.67 | 68.52 |
| GCnet101 | 91.94 | 79.45 | 65.64 | 69.94 |
| SEnet101 | 91.95 | 79.81 | 65.64 | 69.99 |
| Densenet121 | 91.64 | 79.25 | 65.42 | 69.81 |
| SGEnet101 | 91.60 | 79.23 | 65.40 | 69.77 |
| SKnet101 | 91.93 | 79.69 | 65.54 | 69.95 |
| CBAMnet101 | 91.42 | 78.12 | 65.32 | 69.24 |
| Resnet152 | 91.95 | 79.46 | 65.98 | 70.14 |
More detailed data can be seen in folder \result.
Some explanable results for the model by using Grad-CAM.
Bald.
Eyeglasses
Female and Male
Wear hat
The image is from this paper--FAN,Youtu Search, IJCAI 2018
Done:
- Dataset attribute analysis.
- Data process and load.
- Built baseline model(Resnet18 and 101).
- Train and evaluate.
- Save and load model.
- More detailed analysis about the experiment results.
- Parse the input script command.
- Try more famous models, such as
ResNet101,DenseNet,SKNet,SGENet,SENet,GCNet. - Customize the network structure for better performance.
- upload images in html and return the analysis result.
More study notes on the DailyNote.md
The pretrianed model Resnet18 is availabel. Click Google Drive .
Open the run.sh file, read the main.py and you can see the running arguments of model.
bash run.sh
If you want to train and test model by yourself, you can run this command:
nohup python main.py --model_type Resnet18 --batch_size 32 --epochs 128 --learning_rate 1e-3 --exp_version v1 &If you want to visulize the model, you can download pretrianed model and see comments in main part in file featuremap_visulize.py for the usage, and just run:
python featuremap_visulize.py- Sample attributes imbalance.
- Need more effective learning strategies on multi-label task. (loss function? network structure?...)
- The wild images differ from training and test image (aligned)
pip install -r requirements.txt # requirements.txt created by cmd: pipreqs ./
linux 16.0.1
MIT.