This repository contains PyTorch implementation for the paper "MetaAge: Meta-Learning Personalized Age Estimators" (TIP 2022).
- torch 1.2.0 or later versions
- python 3.5.2
- Cuda compilation tools, release 10.1, V10.1.105
- cudnn 6
Codes in this file are used for training the model and validate the results, including data extraction, training loop, etc.
Model establishing are major parts of this file.
Loss function is inside this file.
This file include some less important functions, including model saving, pre-trained model loading, etc.
The file includes functions of establishing resnet model, which is used for extracting identity features of samples.
CUDA_VISIBLE_DEVICES=4,5 python train.py --pretrained_vgg_path xxx --pretrained_resnet_path xxxx --list_root yyy --pic_root_dir yyyy
Explanation of parameters:
pretrained_vgg_path: The path of the pre-trained age estimation model, which is pre-trained on IMDB-WIKI dataset.pretrained_resnet_path: The path of the pre-trained face recognition model, which is the ResNet-50 version of VGGFace2.list_root: The path of the dataset image list files, should contain training and validation dataset files.pic_root_dir: The root path of the dataset images.
- Age estimation pre-trained model is needed, which is pre-trained on IMDB-WIKI dataset.
- Face recognition pre-trained model is needed, used for identity features extracting. We employ the ResNet-50 version of VGGFace2.
- FC_LR = 1e-4 (learning rate of fully-connected parameters)
- NET_LR = 1e-4 (learning rate of other parameters)
- BATCH_SIZE = 64 (training batch size)
- OPTIMIZER = 'adam' (optimizer choice, adam or sgd)
- WEIGHT_DECAY = 1e-4 (weight decay, applied only when using SGD)
- MOMENTUM = 0.9 (momentum, applied only when using SGD)
- DECAY_RATE = 0.1 (decay rate of learning rate every 10 epoches)