2020 Fiscal Year Research-status Report
Scaling up CNN computations for data-intensive scientific applications
| Project/Area Number |
20K19823
|
|---|
| Research Institution | Kobe University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | 計算の最適化 / Computer Vision |
|---|
| Outline of Annual Research Achievements |
On the theoretical side, my research activity has focused on the computational optimization side: High performance GPU kernels optimized for high resolution image processing have been optimized for RTX GPUs using TVM. These kernels include three dimensional convolutions to be used for neuron segmentation applications as well as differentiable block matching operators to be used for stereo reconstruction from Scanning Electron Microscopy (SEM) images of copper surfaces. On the practical application side, steady progress has been made on two fronts: The development and optimization of the insect detection system at high resolution and the analysis of copper surfaces imaging. Regarding the copper surface analysis, two different systems have been proposed: First, a semi-supervised regression system has been optimized to assess the roughness (strictly speaking, to assess the adhesive potential strength) of copper surfaces from SEM images of their surface. Secondly, a stereo reconstruction approach has been developed to infer the 3D structure of the surface from pairs of tilted images. Finally, in an additional practical application project, a system for the analysis of road pavement damages was presented at the IEEE Big Data Cup Challenge workshop.
|
| Current Status of Research Progress |
Current Status of Research Progress
3: Progress in research has been slightly delayed.
Reason
The status is slightly delayed as the development of the curvature surgeon has been paused for a few months for lack of human resources as one of my student graduated and a suitable replacement could not be found. The development of TVM operators has proved efficient, and small scale experiments on insect and copper surfaces are showing high enough accuracy for practical applications. On the other hand, neuron segmentation is proving more difficult due to high level of noise in the data.
|
| Strategy for Future Research Activity |
Future work will focus on two aspects: First deploying the optimized operators for large scale experiments. This task will require the integration of our custom kernels into Pytorch for the training phase, as well as deploying a hybrid computational graph containing both custom kernels and Cudnn kernels for the inference step. Second, further optimizations are still needed to improve on the accuracy of the different practical applications. Special care should be given to the neuron segmentation data for which the accuracy is still too low. Finally, I believe it would be beneficial to make these procedures easy to replicate for the broader community to be able to perform these kinds of optimizations without so much development effort. Further thoughts should be given on how to do so.
|
| Causes of Carryover |
Due to the ongoing pandemic, no money was spent on business travels. Instead almost all of my spending focused on computational equipment: The main buying consisted in Lambda server equipped with high performance NVIDIA GPUS. This server has allowed me to run previously unfeasible computations. Furthermore, imaging equipment was purchased for road damage detection and a minor expenses were made to 3D print large scale reproductions of copper surfaces. Finally software needed for research presentation was purchased.
|
