IURIX
Begin main content

{Hello Yarn 2, Goodbye Node_modules.}



09:10 PM, 27 Aug 2021 by Iuri Sampaio Permalink | Comments (0)

{Optimize DL models via quantization + OpenCV For Beginners price goes up on Sep 1}


Hello iuri,

This is Satya Mallick from LearnOpenCV.com.

In our consulting business, we are occasionally approached by prospective clients who hired a Deep Learning "expert" to implement a solution inexpensively, but are now paying thousands of dollars in cloud computing costs.

In one instance, the cost was $100k per month!

This is often the result of a rookie deep learning developer not taking the time to optimize the model for production. You don't want to be that developer!

After we optimized the models for the client, their costs dropped to under $50k per month!

A few weeks back we shared an Introduction to OpenVINO.

Today, we will dig deeper into a concept called Quantization for optimizing model size and performance.

After we train a deep learning model, we typically store the weights in FP32 (Floating Point 32-bit) format.

When we want to deploy this model in production, using 32-bit precision can be an expensive choice.

Converting a model to 16-bit precision (FP16) reduces the size of the model by 50%, substantially improves speed on many hardware, and usually has negligible effect on accuracy.

But why stop at FP16?

Storing the weights in INT8 (Integer 8-bits) can lead to even smaller models, and blazingly fast inference time. But you have to be careful, and make sure the accuracy does not drop.

In today's post, we will start with an overview of Deep Learning model quantization, understand what is post-training optimization, learn about different quantization methods, and see a real world example by converting YOLOV4 from FP32 to INT8.

​https://learnopencv.com/post-training-quantization-with-openvino-toolkit/​

You will find all the necessary code at the link below.

​https://github.com/spmallick/learnopencv/tree/master/Post-Training-Quantization-with-OpenVino-Toolkit​

OpenCV For Beginners : Pre-Sale Ends Soon ($87)

We have been working very hard to bring OpenCV for Beginners to life! As you may know, this course was backed by 1,636 backers in our recent crowdfunding campaign!

The course will open on Sep 1 and the price will go up to go up to the standard retail price of $117. But you have an option to buy it now at a discounted price of $87.

We are also offering a FREE 3-hour Python Crash Course when you back OpenCV for Beginners before Sep 1.

Click on the button below and reserve a seat.

Satya

​Code for ALL blog posts​

​Courses / Webinar / Jobs​
​Facebook / LinkedIn / Twitter / Instagram
​

​

​

09:10 PM, 27 Aug 2021 by Iuri Sampaio Permalink | Comments (0)

{Fwd: Why You Should Probably Never Use pandas inplace=True}