keras_emotion_classify

a keras implement of emotion classifier.

联系作者

• 邮件: youyou630@163.com
• QQ: 603722900

Test environment

• python 3.7.3
• keras 2.2.4
• OpenCV 4.0.0
• tensorflow 1.12.0
• cuda 10.0.130
• cudnn 7.5.1

DATASETs

FER2013 dataset

How to run?

1. Download dataset FER2013.
2. process dataset.

python prepare_data.py

3. train your model.

python emotion_train.py --dataset data/fer2013 --checkpoint ckpt/CapsuleNet -b 128 --network CapsuleNet

Dataset summary

Results

all networks run 70 epochs.

Network	optimizer	learning rate	acc	val_acc
VGG16	adam	0.0001	89%	64%
VGG19	adam	0.0001	89%	64%
ResNet50	adam	0.0001	75%	50%
DenseNet121	adam	0.0001	91%	55%
DenseNet201	adam	0.0001	92%	55%
MobileNetV2	adam	0.0001	65%	47%
CapsuleNet	adam	0.0001	95%	66%
CapsuleResNet	adam	0.0001	88%	63%
DenseNet201	adam	0.002	88%	63%
CapsuleNet	adam	0.0005	92%	65%
CapsuleResNet	adam	0.0005	86%	64%
VGG16	sgd	0.01	93%	62%

FERPLUS dataset

How to run?

1. Download dataset FERPlus.

git clone https://github.com/microsoft/FERPlus.git

2. process dataset.

cd FERPlus/src
python generate_training_data.py -d /dst/dataset/path/ferplus -fer you/fer2013/dataset/path/fer2013.csv -ferplus ../fer2013new.csv
cd you/own/path/keras_emotion_classify
python prepare_ferplus_data.py --image /dst/dataset/path/ferplus --label FERPlus/data --dst final/dst/path

3. train your model.

python emotion_train.py --dataset final/dst/path --checkpoint ckpt/CapsuleNet -b 128 --network CapsuleNet

Dataset summary

Results

all networks run 70 epochs.

Network	optimizer	learning rate	acc	val_acc
CapsuleNet	adam	0.0001	96%	82%
VGG16	adam	0.0001	95%	81%

mAP

use CapsuleNet model

emotion	angry	disgust	fear	happy	sad	surprise	neutral	contempt
AP	65%	0.97%	26.4%	89.9%	25.4%	77.8%	74.9%	26.3%

mAP = 48.3%