- Notifications
You must be signed in to change notification settings - Fork37
usef-kh/fer
Folders and files
Name | Name | Last commit message | Last commit date | |
---|---|---|---|---|
Repository files navigation
This work is published onarXiv
Our final model checkpoint can be foundhere
In this work, we achieve the highest single-network classification accuracy on FER2013. We adopt the VGGNet architecture, rigorously fine-tune its hyperparameters, and experiment with various optimization methods. To our best knowledge, our model achieves state-of-the-art single-network accuracy of 73.28 % on FER2013 without using extra training data.
In tuning, we experiment with several deifferent optimizers, learning schedulers and run a grid search over all parameters. Some of our results are shown below.
Optimizers | Schedulers |
---|---|
![]() | ![]() |
Visualizing the information captured inside deep neural networks helps describe how they differentiate between different facial emotions. A saliency map is a common technique used in visualizing deep neural networks. By propagating the loss back to the pixel values, a saliency map can highlight the pixels which have the most impact on the loss value. It highlights the visual features the CNN cancapture from the input; thus, allowing us to better understand the importance of each feature in the original image on the final classification decision.
To use this repo, create a conda environment usingenvironment.yml
orrequirements.txt
# from environment.yml (recommended)conda env create -f environment.yml# from requirements.txtconda create --name <env> --file requirements.txt
Download the officalfer2013 dataset, and place it in the outmost folder with the following folder structuredatasets/fer2013/fer2013.csv
To train your own version of our network, run the following
python train.py network=vgg name=my_vgg
To change the default parameters, you may also add arguments such asbs=128
orlr=0.1
. For more details, please refer toutils/hparams.py