myDRGAN

Pytorch implimentation of《2017CVPR-Representation Learning by Rotating Your Faces》, which is relied on kayamin's implement of DR-GAN.

CVPR2017: Representation Learning by Rotating Your Faces

we trained the network using multiPIE and evaluated it for face recognition.

Usage

look at document tree and see each file for more details, we can modify dataset and training parameters to run it:

myDRGAN
│  continue_scripts.sh
│  generate_multi.sh
│  generate_single.sh
│  iden_multi.sh
│  iden_single.sh
│  main.py
│  README.md
│  train_multi.sh
│  train_single.sh
│
├─.vscode
├─data
│      data.py
│      datdaset guide.pdf
│      mydataset.py
│      __init__.py
│
├─model
│      model.py
│      multiple_DR_GAN_model.py
│      multiple_DR_GAN_model_old.py
│      single_DR_GAN_model.py
│      single_DR_GAN_model_old.py
│      weights.py
│      __init__.py
│
├─run
│      generate_image.py
│      representation_learning.py
│      run.py
│      scheduler.py
│      train_multiple_DRGAN.py
│      train_single_DRGAN.py
│      __init__.py
│
├─snapshot
│  └─bestmodel
│          goodgen_multi_G.pth
│          goodgen_single_G.pth
│          goodiden_multi_G.pth
│          goodiden_single_G.pth
│
└─util
        myacc.py
        mybuffer.py
        mylog.py
        mytranmodel.py
        __init__.py

there are scripts to run DRGAN, we can modify some parameters in scripts:

generate images

generate_multi.sh
generate_single.sh

identity recognition

iden_multi.sh
iden_single.sh

train model

train_multi.sh
train_single.sh

Experiments

The experiments in following are preliminary(something maybe wrong), we are supposed to explore more details. But I am sorry to leave this reposity to start carrying out other assignments.

single DRGAN

batch_size:

64

data prepeocess:

transforms.CenterCrop(150)
transforms.RandomCrop(96)

policy:

if epoch<2**1+3: ratio=1+1 # 1:1
elif epoch<2**2+3: ratio=2+1 # 1:2
elif epoch<2**3+3: ratio=3+1 # 1:3
elif epoch<2**4+3: ratio=4+1 # 1:4
elif epoch<2**5+3: ratio=5+1 # 1:5
elif epoch<2**6+3: ratio=6+1 # 1:6

total epoches:

40

plot loss

generate images

identification rate

degree	0	15	30	45	60
goodiden	86.8	91.4	91.1	87.4	81.8
goodgen	88.7	91.8	90.1	86.1	81.2

we have higner iden rate but the generated images are really ugly.

multi DRGAN

imagesperID:

6

batch_size:

60

preprocess:

transforms.CenterCrop(150)
transforms.RandomCrop(96)

policy:

if epoch<2**1+1: ratio=2**0+1 # 1:1
elif epoch<2**2+1: ratio=2**1+1 # 1:2
elif epoch<2**3+1: ratio=2**2+1 # 1:3
elif epoch<2**4+1: ratio=2**3+1 # 1:4
elif epoch<2**5+1: ratio=2**4+1 # 1:8

total epoches:

40

plot loss

generate images(really poor)

identification rate

degree	0	15	30	45	60
goodiden	89.5	93.5	92.7	89.4	85.1
goodgen	88.6	91.0	89.4	83.3	76.4

we have higner iden rate but the generated images are really ugly.

Good iden rate doesn't mean good generatation, we can see the comparision in snapshot file.

model

some models to download: bestmodel in baidudisk password: bc97