description EMC2 Model Compression Challenge (EMCC)

Deep learning has recently pushed the state of the art boundaries in many computer vision tasks. However, existing deep learning models are both computationally and memory intensive, making their deployment difficult on devices with low compute and memory resources. To fit these emerging models on such devices, novel compression techniques are needed without significantly decreasing the model accuracy.

The EMC2 Model Compression Challenge (EMCC) aims to identify the best technology in deep learning model compression. To win a prize in EMCC, the solution will be evaluated to meet the two metrics in two tracks below:

  • Achieve highest accuracy within the target model size. The submission will not be evaluated if the model size is outside of the target range.
  • Achieve smallest model size within the target accuracy. The submission will not be evaluated if the accuracy is outside of the target range.

A participant (or a team) can submit a single model in Tensorflow ( or PyTorch ( Final scores will be computed after submission closes.

There are two categories, computer vision and NLP. Computer vision has three tracks, each participant can submit to either or all tracks.

Computer Vision Category:

  • ImageNet Classification: This category focuses image classification models.
  • COCO Object Detection: This category focuses on object detection models.
  • PASCAL Object Segmentation: This category focuses on object segmentation models.

NLP Category:

  • NLP Transformer: including but not limited to models such as BERT, GPT, XLNET, etc.

Timeline and Submission

Please see the for dates and additional details.