Ensemble/Committee Implementation of Convolutional Neural Networks for Pattern Recognition?

Question

Neural networks

Ensembles of convolutional neural networks show the best results in terms of pattern recognition accuracy. In particular, an ensemble/committee of 35 convolutional neural networks showed the highest accuracy in recognizing digit patterns from the MNIST database (handwritten digit pattern database), showing an error rate on a test sample of 0.23% :

Source: The MNIST database of handwritten digits (yann.lec...

In one of the modern systems for image classification, a tangible advantage is obtained when using an ensemble of convolutional networks. Reflected in an article by He Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun ( Microsoft experts ):

Deep residual learning for image recognition/Deep...

The MSRA team with their Ensemble Paradigm ResNets model won first place in the "Large Scale Visual Recognition Challenge 2015" (I... in all 5 main categories:

• ImageNet Classification: 152 layer networks
  
• ImageNet Detection: 16% better than 2nd place
  
• ImageNet Localization: 27% better than 2
  
• COCO Detection: 11% better than 2nd
  
• COCO Segmentation: 12% better than 2nd
  

Source: MSRA @ ILSVRC & COCO 2015 competitions (presentation...

The article describing the model contains a table:

Error rates (%) in results with ensembles. Top-5 error metric based on the ImageNet test set .

These more or less recent examples of the successful use of an ensemble/committee of convolutional neural networks indicate the relevance of using the ensemble paradigm (in particular, the ensemble of convolutional neural networks) in pattern recognition problems. For earlier examples demonstrating the effectiveness of using an ensemble, see here: Convolutional Neural Network Ensemble for Recognizing ... recognition/classification of digit patterns or ImageNet visual patterns (for non-commercial

использования) . В идеальном случае было бы неплохо, если бы реализация нашлась на Python или C#. Кто с этим сталкивался, возможно, сможет помочь мне сориентироваться.

Answer 1

https://github.com/tensorflow/models