AgeDB-DIR

Installation

Prerequisites

1. Download AgeDB dataset from here and extract the zip file (you may need to contact the authors of AgeDB dataset for the zip password) to folder ./data

2. (Optional) We have provided required AgeDB-DIR meta file agedb.csv to set up balanced val/test set in folder ./data. To reproduce the results in the paper, please directly use this file. If you want to try different balanced splits, you can generate it using

python data/create_agedb.py
python data/preprocess_agedb.py

Dependencies

• PyTorch (>= 1.2, tested on 1.6)
• tensorboard_logger
• numpy, pandas, scipy, tqdm, matplotlib, PIL

Code Overview

Main Files

• train.py: main training and evaluation script
• create_agedb.py: create AgeDB raw meta data
• preprocess_agedb.py: create AgeDB-DIR meta file agedb.csv with balanced val/test set

Main Arguments

• --data_dir: data directory to place data and meta file
• --lds: LDS switch (whether to enable LDS)
• --fds: FDS switch (whether to enable FDS)
• --reweight: cost-sensitive re-weighting scheme to use
• --retrain_fc: whether to retrain regressor
• --loss: training loss type
• --resume: path to resume checkpoint (for both training and evaluation)
• --evaluate: evaluate only flag
• --pretrained: path to load backbone weights for regressor re-training (RRT)

Getting Started

Train a vanilla model

CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --data_dir <path_to_data_dir> --reweight none

Always specify CUDA_VISIBLE_DEVICES for GPU IDs to be used (by default, 4 GPUs) and --data_dir when training a model or directly fix your default data directory path in the code. We will omit these arguments in the following for simplicity.

Train a model using re-weighting

To perform inverse re-weighting

python train.py --reweight inverse

To perform square-root inverse re-weighting

python train.py --reweight sqrt_inv

Train a model with different losses

To use Focal-R loss

python train.py --loss focal_l1

To use huber loss

python train.py --loss huber

Train a model using RRT

python train.py [...retrained model arguments...] --retrain_fc --pretrained <path_to_pretrained_ckpt>

Train a model using LDS

To use Gaussian kernel (kernel size: 5, sigma: 2)

python train.py --reweight sqrt_inv --lds --lds_kernel gaussian --lds_ks 5 --lds_sigma 2

Train a model using FDS

To use Gaussian kernel (kernel size: 5, sigma: 2)

python train.py --fds --fds_kernel gaussian --fds_ks 5 --fds_sigma 2

Train a model using LDS + FDS

python train.py --reweight sqrt_inv --lds --lds_kernel gaussian --lds_ks 5 --lds_sigma 2 --fds --fds_kernel gaussian --fds_ks 5 --fds_sigma 2

Evaluate a trained checkpoint

python train.py [...evaluation model arguments...] --evaluate --resume <path_to_evaluation_ckpt>

Reproduced Benchmarks and Model Zoo

We provide below reproduced results on AgeDB-DIR (base method SQINV, metric MAE). Note that some models could give better results than the reported numbers in the paper.

Model	Overall	Many-Shot	Medium-Shot	Few-Shot	Download
LDS	7.67	6.98	8.86	10.89	model
FDS	7.69	7.11	8.86	9.98	model
LDS + FDS	7.47	6.91	8.26	10.55	model