cs221_project

This repository contains the Tensorflow codebase of our implementation to tackle the task Facial Expression Recognition and this is the project we have done for Stanford course CS221: Artificial Intelligence. We train MLP (Multiple-layer Perceptron) and CNN (Convolutional Neural Network) models to tackle the classification task. We use the training data from Kaggle Facial Expression Recognition Challenge and our best model achieves 69.0% of testing accuracy, approaching the state of the art method as stated in the Kaggle Leadboard.

Detailed Project Report:

http://www.cs.stanford.edu/~kaichun/resume/cs221_project_report.pdf

Data Download and Pre-processing

• All data can be downloaded from the official Kaggle challenge webpage: Kaggle Facial Expression Recognition Challenge

• We convert the data into HDF5 file format by running the script

        python write_hdf5.py
  

• We use the following script to compute the mean and scale of the training images in order to perform whitening during training and testing.

        python compute_mean_scale.py
  

Code Organization

• All the codes are in the exp folder.

• exp/cnn contains the code for our CNN experiments

• exp/fcn contains the code for our MLP experiments

• The other Python scripts under exp folder are used to draw the training curves, confusion matrix and other analysis and visualization that are presented in the final report.

• Insides each of exp/cnn and exp/mlp folders, there are three main Python scripts: train.py, model.py, eval.py. Different version of the three files exist in both folders. They are different variants of training scripts and network architectures that we experimented.

• To train a model, you should run train.py like the following

        python train.py [python_model_file_prefix] --gpu [gpu_id] --batch [batch_size] 
        --epoch [training_epochs] --describe [other_descriptive_comments_for_this experiments] --wd [weight_decay]
  

Run the following command to get the detailed usage for train.py.

        python train.py -h

• Network architecture are defined in model.py.

• After training, we use eval.py script to test the model on testing split by running the following command.

        python eval.py [python_model_file_prefix] [path_to_pretrained_model_checkpoint] [output_file] --gpu [gpu_id] --batch [batch_size]
  

[output_file] stores the logs of this evaluation.

Contact

Webpage: http://www.cs.stanford.edu/~kaichun/

E-mail: kaichun [at] cs.stanford.edu