Pytorch-implementation-of-DeepRegressionForests

This is a pytorch version implementation of the caffe-DeepRegressionForests for age estimation task.

Original Paper and Code

Original Paper: https://arxiv.org/abs/1712.07195

Original Code: https://github.com/shenwei1231/caffe-DeepRegressionForests

The original version utilize the caffe to implement the whole project, which needs a lot of effort to implement the conv functions. I decided to reimplement the algorithm with pytorch version code

@inproceedings{shen2018DRFs,
    author = {Wei Shen and Yilu Guo and Yan Wang and Kai Zhao and Bo Wang and Alan Yuille},
    booktitle = {Proc. CVPR},
    title = {Deep Regression Forests for Age Estimation},
    year = {2018}
}

Initialize

Firstly, download the codes with following command:

git clone --recursive https://github.com.cnpmjs.org/Kasumigaoka-Utaha/Pytorch-implementation-of-DeepRegressionForests
cd Pytorch-implementation-of-DeepRegressionForests

Create environment

You can follow the following code to create environment:

conda create -n drf python==3.7.9
conda activate drf
pip install -r requirements.txt

Dataset

In this project, I decided to use FGNET for a light training process. You can download the dataset via

wget http://yanweifu.github.io/FG_NET_data/FGNET.zip
unzip FGNET.zip

You can preprocess the dataset with

python readData.py

The results were saved in the info.csv and the split results were saved in imgs_train.csv and imgs_test.csv.

As for splitting the datasat to the training set and test set, I choose to directly split the images directly. The training set contains 839 images and the test set contains 163 images.

Training

You can train the model by

python train.py --data_dir ./FGNET/images

Experiment Training Results

Test Dataset: FGNET

Train.opt = 'sgd'

Experiment 1: modify the initial learning rate

Basic hyperparameters:

TRAIN.LR_DECAY_STEP = 20
TRAIN.LR_DECAY_RATE = 0.2
TRAIN.MOMENTUM = 0.9
TRAIN.WEIGHT_DECAY = 0.0

Initial learning rate	Best_cs	Best_mae
0.01	0.356	7.18
0.1	0.675	4.68
0.09	0.724	4.538
0.085	0.692	4.585

Experiment 2: modify the momentum

Basic hyperparameters:

TRAIN.LR_DECAY_STEP = 20
TRAIN.LR_DECAY_RATE = 0.2
TRAIN.LR = 0.09
TRAIN.WEIGHT_DECAY = 0.0

Momentum	Best_cs	Best_mae
0.9	0.724	4.538
0.85	0.711	4.551
0.8	0.712	4.594

Current best hyperparameters:

TRAIN.LR_DECAY_STEP = 20
TRAIN.LR_DECAY_RATE = 0.2
TRAIN.LR = 0.09
TRAIN.WEIGHT_DECAY = 0.0
TRAIN.MOMENTUM = 0.9

Further Plans

These are some of my further plans:

1. Modify the other hyperparameters, lr_decay_rate, lr_decay_step
2. Try to use another dataset instead of FGNET (current dataset is too small)
3. Try to modify the DeepRegressionForest structure 

Details

Further information will be updated soon.