Find file History
EmmaNingMS and ankkhedia Add demo for resnet50, squeezenet, mnist and emotion_ferplus (#118)
* Add demo link

* deleting the title by mistake

* Update the link

* Add the demo link

* Add demo link

* add demo link

* correct a mistake

* update the description

* Update

* Update

* Update

* Update

* Update

* Update
Latest commit 0bb8cc6 Nov 30, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information. Add demo for resnet50, squeezenet, mnist and emotion_ferplus (#118) Dec 1, 2018

FER+ Emotion Recognition


This model is a deep convolutional neural network for emotion recognition in faces.


Model Checksum Download (with sample test data) ONNX version Opset version
Emotion FERPlus MD5 31 MB 1.0 2
MD5 31 MB 1.2 7
MD5 31 MB 1.3 8


"Training Deep Networks for Facial Expression Recognition with Crowd-Sourced Label Distribution" arXiv:1608.01041


The model is trained on the FER+ annotations for the standard Emotion FER dataset, as described in the above paper.


The model is trained in CNTK, using the cross entropy training mode. You can find the source code here.


Run Emotion_FERPlus in browser - implemented by ONNX.js with Emotion_FERPlus version 1.2



The model expects input of the shape (Nx1x64x64), where N is the batch size.


Given a path image_path to the image you would like to score:

import numpy as np
from PIL import Image

def preprocess(image_path):
  input_shape = (1, 1, 64, 64)
  img =
  img = img.resize((64, 64), Image.ANTIALIAS)
  img_data = np.array(img)
  img_data = np.resize(img_data, input_shape)
  return img_data


The model outputs a (1x8) array of scores corresponding to the 8 emotion 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}


Route the model output through a softmax function to map the aggregated activations across the network to probabilities across the 8 classes.

import numpy as np

def softmax(scores):
  # your softmax function
def postprocess(scores):
  This function takes the scores generated by the network and returns the class IDs in decreasing 
  order of probability.
  prob = softmax(scores)
  prob = np.squeeze(prob)
  classes = np.argsort(prob)[::-1]
  return classes

Sample test data

Sets of sample input and output files are provided in

  • serialized protobuf TensorProtos (.pb), which are stored in the folders test_data_set_*/.