C3AE

This is a unofficial keras implements of c3ae for age estimation. welcome to discuss ~

--------[result]-----------------

source	version	IMDB(mae)	WIKI(mae)	extra change	model
from papper	--	6.57	6.44	--	--
our implement	c3ae-v84	6.77	6.74	change kl to focal loss without se_net	model/imdb_focal_loss_c3ae_v84.h5
our implement v2	c3ae-v89	6.58	--	SE_NET + focal_loss	model/c3ae_imdb_v89.h5
our implement v3	c3ae-v90	6.51	--	white norm + SE_NET + focal_loss	mail to geekpeakspar@gmail.com

U can add gender prediction to the task if you want reach a lower mse. It cant decrease mae from 6.51 to 6.41.

structs

• assets
• dataset (you`d better put dataset into this dir.)
• detect (MTCNN and align)
• download.sh (bash script of downloading dataset)
• model (pretrain model will be here)
• nets (all tainging code)
  • C3AE.py
• preproccessing (preprocess dataset)

Pretrain model(a temp model)

all trainned model saved in dir named "model"

required enviroments:

numpy, tensorflow(1.8), pandas, feather, opencv, python=2.7

pip install -r requirements.txt

test

• for image

  python nets/test.py -i assets/timg.jpg

• for video

  python nets/test.py -v

Preparation

download imdb/wiki dataset and then extract those data to the "./dataset/"
download wiki download imdb

Preprocess:

>>>  python preproccessing/dataset_proc.py -i ./dataset/wiki_crop --source wiki -white -se
>>>  python preproccessing/dataset_proc.py -i ./dataset/imdb_crop --source imdb -white -se

training:

plain net
>>> python C3AE.py -gpu -p c3ae_v16.h5 -s c3ae_v16.h5 --source imdb -w 10
with se-net and white-norm (better result)
>>> python C3AE.py -gpu -p c3ae_v16.h5 -s c3ae_v16.h5 --source imdb -w 10 -white -se

DETECT:

[mtcnn] (https://github.com/YYuanAnyVision/mxnet_mtcnn_face_detection): detect\align\random erasing

net struct

Q&A:

• only 10 bins in paper: why we got 12 category: we can split it as "[0, 10, ... 110 ]" by two points!\
• Conv5 1 * 1 * 32, has 1056 params, which mean 32 * 32 + 32. It contains a conv(1 * 1 * 32) with bias
• feat: change [4 * 4 * 32] to [12] with 6156 params.As far as known, it may be compose of conv(6144+12) ,pooling and softmax.
• the distribution of imdb and wiki are unbalanced, that`s why change the KL loss to focal loss

puzzlement:

• the result of the feature layer(W2) is far from expected. Maybe our code exists some error.

Reference

• focal loss: https://github.com/maozezhong/focal_loss_multi_class/blob/master/focal_loss.py
• mtcnn: https://github.com/YYuanAnyVision/mxnet_mtcnn_face_detection