DL_FS2K

Comparision of several CNN based networks on human face attribute classification. 6 methods related to light-weight convolutional neural networks are implemented using pytorch, and results on FS2K are presented in table format as below:

1. Accuracy of 6 attributes

method	hair	hair color	gender	earring	smile	frontal face
MobileNetv3 Small	94.83%	51.72%	81.26%	82.12%	66.82%	81.16%
MobileNetv3 Large	94.45%	50.09%	83.55%	78.79%	66.54%	81.54%
SuffleNet	95.03%	39.96%	60.42%	82.12%	64.05%	83.37%
EfficientNetv2	95.03%	51.24%	84.70%	82.12%	66.25%	83.37%
MB3 Small+conv+softmax	95.02%	40.15%	81.17%	82.50%	64.15%	83.37%
MB3 Small+linear+softmax	95.03%	39.96%	78.01%	82.12%	66.25%	83.36%

2. Average accuracy

method	average accuracy
MobileNetv3 Small	76.32%
MobileNetv3 Large	75.83%
SuffleNet	70.83%
EfficientNetv2	77.12%
MB3 Small+conv+softmax	74.39%
MB3 Small+linear+softmax	74.12%

Preparation

1. clone the repo with command

git clone https://github.com/Eiuyc/DL_FS2K.git

2. download FS2K dataset with command

git clone https://github.com/DengPingFan/FS2K.git

3. move FS2K under DL_FS2K/data/

The structure of the repo should look as follows:

DL_FS2K
├─data
│  └─FS2K # FS2K dataset root directory
│      ├─photo
│      └─sketch
├─doc
├─save # saved models
│ ├─model
│ │ └─ENV2.pth
│ ├─MB3_L.ckpt
│ ├─MB3_S.ckpt
│ ├─MB3_S_conv_softmax.ckpt
│ ├─MB3_S_linear_softmax.ckpt
│ └─SF.ckpt
├─effnetv2.py # efficientNet
├─MB3_L_S.ipynb # mobileNetv3 Large and Small
├─MB3_S_conv_softmax.ipynb # mobileNetv3 Small with convolutional layer branches
├─MB3_S_linear_softmax.ipynb # mobileNetv3 Small with linear layer branches
├─SF.ipynb # SuffleNet
└─train_ENV2.py #  trainning entrypoint

Train

• ipynb

  1. open the corresponding *.ipynb file
  2. the train func is defined as below:

  def train(m, # model
         d, # device
         train_dl, # train dataloader
         val_dl, # validation dataloader
         saveDir=Path('save'), # save directory
         resumePath=None, # resume the last training from this checkpoint
         lr=0.001, # learning rate
         e=50, # total epoch
         s=10 # save checkpoint every s epochs
        ):

  3. change the saveDir to the actual save directory when you execute
  4. execute the code blocks in order
  5. the best.ckpt file will be written under saveDir after the training

Note the MB3 Small and Large are inplemented on the single file MB3_L_S.ipynb, and mode is set to small as default. Change the mode=large in the script when you train MB3 Large:

# for MB3 Small:
class MobileNetV3(nn.Module):
    def __init__(self, nclass=1000, mode='small', width_mult=1.0, dilated=False, norm_layer=nn.BatchNorm2d):

• py

  EfficientNetv2 can be trained with command:

  python train_ENV2.py

  the .pth file will be written under ./save/model as default

Validate

• ipynb

  execute the val function as below to validate the .ckpt file

  em, epo, acc = load('save/best.ckpt')
  val(Model().to(d).eval(), em.state_dict(), val_dl, d)

• py

  model will be validated automatically while training

Acknowledge

All the backbone network implementations above are from github. MobileNet is from Tramac. EfficientNet is from d-li14. SuffleNet is from jaxony.

License

GNU General Public License v3