face-attribute-prediction

Face Attribute Prediction on CelebA benchmark with PyTorch Implemantation.

Dependencies

• Anaconda3 (Python 3.6+, with Numpy etc.)
• PyTorch 0.4+
• tensorboard, tensorboardX

Dataset

• CelebA dataset is a large-scale face dataset with attribute-based annotations. Cropped and aligned face regions are utilized as the training source.
• Pre-processed data and specific split list has been uploaded to list directory.
• lfwA+ dataset is the private test dataset.

Features

• BCE loss for attributes recognition.
• Good capacity as well as generalization ability.Achieve 92%+ average accuracy on CelebA Val as well as >73% on LFWA+.
• ResNet-50 as backbone
• Focal Loss
• Class balanced sampler
• fast convergence: 91% acc on CelebA Val after 1 epoch.

Result

Method	CelebA Acc	LFWA ACC
BCE Baseline	91.3	72.2
Cos Decay + US + focal	92.14	73.31
Cos Decay + BS + focal	91.7	73.9

• link to pretrained model(ResNet50)
• see train.sh script for more details.

Attention:

• I manually changed CE loss to BCE to lint code as well as try some tricks. Basically they are the same.
• Normalization value are calculated on CelebA train_split set. You can refer to the code here
• Class Balance Sampler may hurt general representation learning, but lead to better generalization on tail classes. Better used in fiine-tuning process. For more details, kindly refer to the paper
  • BBN: Bilateral-Branch Network with Cumulative Learning for Long-Tailed Visual Recognition