Quantization of Fully Convolutional networks for semantic segmentation

Project Title

Quantization of Fully Convolutional networks for semantic segmentation

Project Area of Specialization Artificial IntelligenceProject Summary

Semantic Segmentation is one of the key problems in Computer Vision. ‘Fully Convolutional Networks’ are one of the tools to achieve this task. These networks exceed the state of the art ‘Convolutional Neural Networks’ in terms of performance using decoder networks through upsampling and skip connections. However, the addition of upsampling ,as a decoding method, at the tail end of the network introduces increased computational complexity and cost. Reducing the complexity and cost is essential for implementation of these networks on hardware such as FPGAs, VLSI, and embedded systems. This project suggests Quantization as an optimization method for Fully Convolution Networks to deal with the mentioned problem. To achive this task we  convert the normally used floating point parameters/weights into Fixed point.

Project Objectives

Our objective is to learn implementation of FCNs for semantic segmentation and then obtain a quantized version of it. Quantization will help us achieve the objective of reducing the computational complexity along with implementation cost. We will try to obtain a software prototype by the end of this project. This prototype can further be worked with for Hardware implementation.

Project Implementation Method

To test the optimzation method we used a Fully Convolutional Network trained on PASCAL VOC12 dataset for Semantic Segmentation. The proccess was conducted on Matlab. This involved passing the every parameter/weight from  the network through a quantizer. The level at which parameters are quantized to are specified for each layer. Retraining of quantized network is different from conventional training .After training a network we quantize it’s parameters/weights.Then,we use these parameters in forward pass to compute the output.Back pass also uses the same quantized parameters to compute error.The changes/updated are then made to Floating Point parameters rather the quantized parameters.The Floating Point parameters are again quantized and same steps are repeated for every batch in case of batch training for every image in other cases.

Benefits of the Project

The implementation of Nueral Networks are extremely common on resourse limited devices. Therefore, Our projects suggest an optimzation method for high computationaly expensive networks. Such as, the fully convolutional network trained for semantic segmentation  was extremely computationaly expensive we were able to achive quantization.

Technical Details of Final Deliverable

For conclusion of this project we have been able to obtain the desirable results. Firstly, we conducted training and quantization of a CNN for MNIST data set to test the optimization method. After sensitivity and sparisty analysis, we had exellent resullts to show. The sparisty induced was 50.4% with accuracy of 99.1% approximatly. These compared well to the orignal results. After successfully optmizing the test network, we achived similar level of performance for the main task ,Semantic Segmentation using FCN, mentioned in the results section.

Final Deliverable of the Project Software SystemType of Industry IT , Medical , Transportation Technologies Artificial Intelligence(AI)Sustainable Development Goals Industry, Innovation and InfrastructureRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	70000
GPU	Equipment	1	70000	70000