Used Google Cloud Platform

31 People have highlighted this activity

Activity Feed

DevSecOps Specialist, Xurya Daya Indonesia
Feb 05, 2021
Enhancement completed, than let's see what happen if we implement it into web-based. 
Read more
DevSecOps Specialist, Xurya Daya Indonesia
Jan 25, 2021

Machine Learning Pitfalls (2)

If you ever think about these, then you're not ready to develop machine learning.
Read more
DevSecOps Specialist, Xurya Daya Indonesia
Feb 02, 2021
Well this photo explains everything.
Read more
DevSecOps Specialist, Xurya Daya Indonesia
Jan 04, 2021
Happy new year for all the readers! May your wishes come true this year!

Machine Learning Pitfalls (1)

Quick story about my experience while training datassets for machine learning (image processing). Please, choose all datasets you want to train wisely. Do recheck and filter first before all datasets move into training phase. 

As you can see in the photo, this is my report to my boss. After having trial and error to enhance the confidence level, all i get from the predicition for same photos (before 2nd training --> after 2nd training) are below than what i expect. In this case, i use photos as my datasets. Different position, different angles are affecting the confidence level. Datasets that you trained might turnback against your expectation because it confuses them to predict. 
Read more
DevSecOps Specialist, Xurya Daya Indonesia
Dec 24, 2020

Image Processing with AutoML


Quick short explanation about the implementation more focused on the customer part. I use Google Vision & AutoML by Google Cloud Platform. The main goal is the system able to differentiate which parts are factory default (original) and which parts are customized (non-original). And I expect the machine learning able to detect the brand and the type of the motorcycles too. Therefore, I trained various customized motorcycles and original motorcycles with different types and brands.

Labels were made to differentiate the parts and brands/types of the bike. There are 300++ photos trained, and the results are quite amazing. The system itself able to detect the labels correctly. However, since the system trained with 300++ photos only, the confidence prediction level might be low sometimes. And even resulted positive false, or negative true. 

So that's why more datasets needed, and more training needed. 
Read more
DevSecOps Specialist, Xurya Daya Indonesia
Dec 07, 2020
Currently I'm developing an Artificial Intelligence or Machine Learning that focuses on Image Processing. It's more like a research and development. After having a discussion with external teams in choosing tools, i decided to train all the datasets in AutoML service from Google Cloud Platform. 

Besides the r&d phase, all implementation that related to the machine learning needed for further like web development, so user able to scan it via web-based i guess. Or even API that connects the machine learning system with a mobile app. Let's settle my goals at first for this development.

Expectation

  • The system able to detect the brand and the type of the bike, the machine learning itself detect a brand of a bike, or even type of the bike (release year too. it'll be awesome if its able. lol). Farther image capture would be recommended for this kind of prediction to ensure it achieves accurate prediction. (example -- left image).

  • The system able to detect the quality of the parts. In this case. close-up photo-angle needed like right side images. It prevents the misleading label-detection. Besides that, it helps the condition of the part are seen enough clearly for the prediction such as the scratches, dents, etc.

  • The system able to differentiate customized and original parts. Since we're going to value every single part of the bikes, original parts that still attached on the bikes will get better values than customized parts (in the market). So that's why it's quite important to be able detect those. 
Read more
Loading...