FER+ Emotion Recognition 1.2

By for July 3, 2018

Report Abuse
This model is a deep convolutional neural network for emotion recognition in faces.
**Paper** [Training Deep Networks for Facial Expression Recognition with Crowd-Sourced Label Distribution.][1] **Dataset** The model is trained on the FER+ annotations for the standard Emotion FER [dataset][2], as described in the above paper. **Source** The model is trained in CNTK, using the cross entropy training mode. You can find the source code [here][3]. **Model input and output** **Input** * The model expects a grayscale input image of the shape (1x64x64) **Output** * Output is a (1x8) array **Pre-processing steps** Resize the input image to a (1x64x64) array of type float32. **Post-processing steps** Route the model output through a softmax function to map the aggregated activations across the network to probabilities across the 8 classes, where the labels map as follows: **emotion_table**= {'neutral':0, 'happiness':1, 'surprise':2, 'sadness':3, 'anger':4, 'disgust':5, 'fear':6, 'contempt':7} **Sample test data** Sets of sample input and output files are provided in .npz format (test_data_*.npz). The input is a (1x64x64) numpy array of a test image, while the output is an array of length 8 corresponding to the output of evaluating the model on the sample input. **Model size: 34 MB** **License** [MIT][4] [1]: https://arxiv.org/abs/1608.01041 [2]: https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data [3]: https://github.com/ebarsoum/FERPlus [4]:https://github.com/onnx/models/blob/master/LICENSE