Facial Emotion Recognition on fer2013 dataset using TensorFlow! (Accuracy ~ 65%)

DATA SET:

• Download Data Set: fer2013.bin (63M) and test_batch.bin (7.9M) from https://goo.gl/ffmy2h

  Image Properties: Size of an image - 48 x 48 pixels (2304 bytes), Size of a label - number in (0..6) (1 byte) (0=Angry, 1=Fear, 2=Happy, 3=Sad, 4=Disgust, 5=Surprise, 6=Neutral).

  Data Set Format: 1st byte is the label number and the next 2304 bytes are the image pixels.

• Create a data directory in your system: /tmp/fer2013_data/

• Put the training data set (28,709 images) in: /tmp/fer2013_data/fer2013-batches-bin/fer2013.bin

• Put the testing data set (3,589 images) in: /tmp/fer2013_data/fer2013-batches-bin/test_batch.bin

HOW TO TRAIN:

• Install TensorFlow: https://www.tensorflow.org/versions/r0.7/get_started/os_setup.html#pip-installation
• Run python fer2013_train.py
• Run python fer2013_eval.py on fer2013.bin data (Training Precision)
• Run python fer2013_eval.py on test_batch.bin data (Evaluation Precision)

HOW TO DEMO:

• From https://goo.gl/ffmy2h download the checkpoint files checkpoint, model.ckpt-6000, model.ckpt-6000.meta located in 65acc-checkpoint dir.

• Copy these files into /tmp/fer2013_train/

• Inside the demo folder of the emotion-recognition project, run ./demo.sh IMG#. Provide an IMG#, which is the row number in the private-test-150.csv, where each row corresponds to an image. There are 150 such rows/images.

• Executing ./demo.sh outputs the label predicted by the trained model. This can be cross checked with the first value in the row of the csv file.

• To actually view the image and visually cross check the emotion, run uint8-to-image.py script on private-test-150.csv. This generates 150 .png image files with appropriate IMG# in the image file name.

STATS DASHBOARD:

• Run tensorboard --logdir "/tmp"
• Go to http://0.0.0.0:6006/
• This displays events, images, graphs and histograms for the train and eval runs on the model.

SCREENSHOTS:

• https://goo.gl/rJfjYL

REFERENCES:

• Code references and examples from https://www.tensorflow.org
• Data Set from https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data