Deep Learning Utilities

Date: May 25, 2018
Last Updated: May 27, 2018
Categories:
Projects python
Tags:
python python-c-api numpy deep-learning

Contents


Introduction

Processing data with tensorflow (python) could be accelerated with utilizing advanced CPU drivers and GPU. In this case, the pre-processing and post-processing may delay the fast GPU processing. For example, one of the most ordinary application is converting the segmentation mask picture to the multi-channel mask.

Matrix projecting

As is shown below, in the segmentation picture, we use each color to represent a label. But in real application, to avoid the overlapping of the label value, we need to separate each label in one-hot format, i.e. the multi-channel mask. After processing, we may need to convert the one-hot mask back to segmentation picture so that people could see the result clearly. If we use the python-numpy-API to complete this task, the efficiency may be extremely low. Because we need to use the common dictionary API of python to build the link of the label and the segmentation picture value.

Segmentation picture Muti-channel mask

In this project, we use the unordered map API to rewrite this process, the efficiency of this tool is extremely high. You could see the description of projector to learn more.

IO tool

This project also include tools for IO purpose. It provides a fast approach of batch reading. Now we only support reading the seismic raw data. In the future it may support more kinds of data.

Project Page

See here to view the main page of this project. You could also download the compiled binary files in the release page.

Git Page