C3AE

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

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

Update History:

• 2019-9 C3AE org
• 2020-7 transfer to tensorflow2.1 and exposing gender branch.
  • 1.add gender prediction
  • 2.change neck
  • 3.add Mish6, GeM, Smooth label and so on.
  • 4.add utk, afad, asia dataset
  • 5.add tflite freezing
• 2020-8 fixed bugs and add fixed-precision model
  • 1. remove nan weights (it may case predict as unpredict)
  • 2. add -fp16

Part1 age branch

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 [email protected]

Part2 add gender branch

Triple-Boxes show much influence with different dataset, meanwhile the distribution plays an important role.

source	version	asia	utk	afad	model
our implement v4	asia	age: 5.83 gender 0.955	--	--	./model/c3ae_model_v2_117_5.830443-0.955
our implement v4	asia+utk	--	age: 5.2 gender 0.967	--	./model/c3ae_model_v2_91_5.681206-0.949
our implement v4	asia+utk+afad	age: 5.9 gender 0.9234	age: 5.789 gender: 0.9491	age: 3.61 gender: 0.9827	./model/c3ae_model_v2_151_4.301724-0.962
our implement v4	fp16+white-norm asia+utk+afad	age: 6.0 gender 0.97	age: 5.6 gender: 0.942	age: 3.72 gender: 0.987	./model/c3ae_model_v2_fp16_white_se_132_4.208622-0.973

cation: Gender annotaion of utk is opposite to wiki/imdb/asia.

**You can change weights of loss to improve age mae, [1, 20, 20].

python nets/C3AE_expand.py --white -se --source "afad" -gpu -m "fp16_white_se" -p ./model/c3ae_model_v2_fp16_white_se_132_4.208622-0.973 -fp16 -test

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), which contains "wiki" "imdb" "afad" "asia" "utk"

Pretrain model(a temp model)

all trainned model saved in dir named "model"

required enviroments:

numpy, tensorflow(2.1), pandas, feather, opencv, python=3.6.5

pip install -r requirements2.1.txt

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

pip install -r requirements.txt

test

age and gender branch(only for py3 and tensorflow2+)

• for image

  python nets/test.py -g -white -se -i assets/timg.jpg -m ./model/c3ae_model_v2_fp16_white_se_132_4.208622-0.973

• for video

  python nets/test.py -g -white -v -se -m ./model/c3ae_model_v2_fp16_white_se_132_4.208622-0.973

age branch

• for image

  python nets/test.py -se -i assets/timg.jpg -m model/c3ae_imdb_v89.h5

• for video

  python nets/test.py -v -se -m model/c3ae_imdb_v89.h5

Preparation Datasets

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

Preprocess:

>>> python preproccessing/dataset_proc.py -i ./dataset/wiki_crop --source wiki
>>> python preproccessing/dataset_proc.py -i ./dataset/imdb_crop --source imdb
>>> python preproccessing/dataset_proc.py -i ./dataset/AFAD-Full --source afad 

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
for gender and age prediction:
>>> python nets/C3AE_expand.py -se -white --source "afad" -gpu -fp16 -p ./model/c3ae_model_v2_fp16_white_se_132_4.208622-0.973 

freeze tflite

python nets/C3AE_expand.py -se -white --source "asia" -gpu -fp16 -p ./model/c3ae_model_v2_fp16_white_se_132_4.208622-0.973 -fz

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
• gender prediction: detail in nets/C3AE_expand.py

To-Do:

- 1.anchor free boundbox
- 2.add another new feathers

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