Everything is in progress

Milestone checking

SCHEDULE FOR COMING WEEKS

11.21 -  11.24  Implement convolutional layer on CUDA     Yihao Qian

                      Implement max pooling layer on CUDA        Yiqing Cai

11.25 – 11.28  Implement ReLU on CUDA                               Yihao Qian

                      Implement fully connected layer on CUDA   Yiqing Cai

11.29 - 12.02 Implement softmax layer on CUDA                 Yihao Qian

                      Implement weight decay                                   Yiqing Cai

12.03 – 12.06 Integrate all the layers into a network structure on CUDA         Yihao Qian 

                       Implement training and testing script           Yiqing Cai

12.07 – 12.10  Test and verify speedup                                    Yiqing Cai & Yihao Qian

 ​

Progress so far

We have successfully set up the environment and all the required dependencies for the project, and make sure that our cuda program can be built successfully. For the network part, we have implemented the sequential version to classify a smaller dataset MNIST, which consists of  handwritten digit numbers.

​

Updated Goal

Our goal is to achieve at least 4x speedup, we think we should be able to deliver that.

And if everything goes well, we are considering a life demo for image classification, for now we will just consider it as the bonus part. If there is time remaining, we will try to get that done.

​

Delivery Plan

Speedup curve compared to sequential program as a graph.

Life demo or not depending on the progress.  

​

Preliminary results

We have implemented sequential program. And here is the initial result of our sequential code. We implemented a CPU sequential convolutional neural network that predict the class for handwritten digits. As you can see from the image, the test accuracy reaches around 98.15%.

​

Issues

It is a little bit difficult to install software and dependencies on cluster. So we would first tried to implement the code on our own computer. But it would be great if the cluster can be installed with opencv, boost, etc.