Repository files navigation

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

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

• 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

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

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

• 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"

all trainned model saved in dir named "model"

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

• 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

• 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

download imdb/wiki dataset and thenextract those data to the "./dataset/"
download wikidownload imdbdownload asiadownload utkdownload afad

>>> 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

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

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

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

• 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

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

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