jacobpolloreno/emotion_classifier

Facial Emotion Classifier

Emotion Classifier

Description

I created this project to explore various architectures aimed at image classification.

Models

• Basic CNN
• Basic CNN with Batch Norm
• Deep CNN
• Fully Convolutional Network
• VGG16 Transfer Learning Network

** To see model details run python model_and_train.py --arch <NAME> --summary

Results

• Deep CNN model had highest test accuracy of 65%
• VGG16 model reached accuracy of 55%

Getting Started

1. Download data from Kaggle competition and unzip into 'data/'
2. Install python dependencies with pip install -r requirements.txt
3. Run python model_and_train.py --help to see available options
4. Try python model_and_train.py --arch basic --summary to print the Keras architecture layout of 'basic CNN' architecture
5. Train with python model_and_train.py --arch <NAME> --run_name <TEXT>
6. Model training is logged by default so you can view it in Tensorboard: tensorboard --logdir=log

CLI Options

$ python model_and_train.py --help

Usage: model_and_train.py [OPTIONS]

Options:
  --arch [basic|basic_with_batch_norm|deep_cnn|fully_conv|vgg]
		                  model architecture to train with
  --summary                       PRINT model summary for architecture
  --save_dir TEXT                 directory to log to
  --run_name TEXT                 name of run for Tensorboard
  --lr FLOAT                      learning rate for training  [default: 0.001]
  --dropout_rate FLOAT            dropout rate for training  [default: 0.3]
  --num_epochs INTEGER            number of epochs to train for  [default:
		                  1000]
  --batch_size INTEGER            batch size for training  [default: 32]
  --patience INTEGER              Early stopping training patience  [default:
		                  10]
  --validation_split FLOAT        [default: 0.1]
  --test_size FLOAT               percentage of date to use for
		                  testing/validation  [default: 0.2]
  --reduce_lr                     Reduce learning rate when val loss has
		                  stopped improving
  --help                          Show this message and exit.

Transfer Learning Model

NOTE: Running the VGG model for the first time will create the bottleneck features from VGG. This will take several minutes to compute and use a lot of system resources.

Dependencies

• Keras
• Click
• Tensorflow
• Numpy
• Sklearn
• OpenCV

TODO

• Add progress bar for events
• Add CLI image classify method
• Add data augmentation to images
• Experiment with different methods of using gray images instead of OpenCV gray2color conversion
• Add a deep residual network