TensorFlow.js Object Detection Made Easy

Подписаться 3,1 тыс.

Просмотров 13 тыс.

50% 1

TensorFlow.js Object Detection Made Easy
In this video you're going to discover an easy way how to train a convolutional neural network for object detection and use the created TensorFlow.js machine learning model afterwards to detect Azure logos locally in your web browser.
Here's what I cover:
00:00 | Intro
00:55 | Preview what I will build
01:35 | Create a Custom Vision project
03:01 | Upload & tag the images
05:41 | Train the model
06:30 | Evaluate / test the model
07:42 | Export & download the model
09:10 | A first look inside the web app
10:32 | Integrate the model with TensorFlow.js
15:45 | Check the results
16:12 | Performance Optimizations
TensorFlow.JS is a great machine learning javascript-based framework to run your machine learning models locally in the web browser as well as on your server using node.js.
But defining your model structure and training it, is way more complex than just using a trained model.
Azure Custom Vision - one of the various Cognitive Services - offers you an easy way to avoid this hassle.
If you wan't to have a look at the final code, please check out my GitHub repository:
github.com/SaschaDittmann/tfj...
The dataset I used to train the model, can be downloaded here:
github.com/microsoft/AIVisual...
🔔 Subscribe for more free data analytics videos: ru-vid.com...
And don't forget to click the bell so you don't miss anything.
💪 Share this video with a RU-vidr friend: • TensorFlow.js Object D...
📢 If you enjoyed this video help others enjoy it by adding captions in your native language -
ru-vid.com_vid...
Watch my most recent upload: bit.ly/2OihAlj
• Uploads from Sascha Di...
🚀 Tools and services I used in this video:
TensorFlow.JS
www.tensorflow.org/js
Azure Custom Vision
www.customvision.ai/
Visual Studio Code
code.visualstudio.com/
✅ Recommended playlists:
TensorFlow.JS
• TensorFlow.JS
Data Science
• Data Science
✅ For business inquiries contact me at CloudBlog@gmx.de
✅ Let's connect:
Watch me Live on Twitch: / saschadittmann
Join our Discord: / discord
Twitter: / saschadittmann
Facebook: / datadrivendev
Instagram: / saschadittmann
LinkedIn: / saschadittmann
GitHub: github.com/SaschaDittmann
DISCLAIMER: This video and description contains affiliate links, which means that if you click on one of the product links, I’ll receive a small commission. This helps support my channel and allows me to continue making awesome videos like this. Thank you for the support!
#TensorFlowJS #CustomVision #ObjectDetection

Опубликовано:

2 авг 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 36

@SaschaDittmann 3 года назад

Hope you enjoy this video! Here's what I cover: 00:00 | Intro 00:55 | Preview what we will build 01:35 | Creatung a Custom Vision project 03:01 | Upload & tag the images 05:41 | Train the model 06:30 | Evaluate / test the model 07:42 | Export & download the model 09:10 | A first look inside the web app 10:32 | Integrating the model with TensorFlow.js 15:45 | Checking the results 16:12 | Performance Optimizations … and more info here: The GitHub repository github.com/SaschaDittmann/tfjs-cv-objectdetection The dataset I used to train the model, can be downloaded here github.com/microsoft/AIVisualProvision/tree/master/Documents/Images/Training_DataSet TensorFlow.js Image Classification Made Easy ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-WFUL_oRT3nk.html

@krzusiekniwazne 3 года назад

And that's what I need. Thank you very much Sascha.. And I am asking for more.

@SaschaDittmann 3 года назад

You're welcome! I'll be sending out a poll later this week on my community tab, what I should create next. 😉

@ahmedrajaspeaks 2 года назад

Great Job Sascha!

@heyyy4987 2 года назад

awesome...

@dalandan8300 Год назад

He sascha, nice tutorial! Thank you, that's what I actually needed. Too bad azure don't allow my debit card for the free trial. Do you have another approach?( Without azure)

@RudfaanMaimahad 2 года назад

is there anyway to 1. change box style (which one around image) 2. crop and save detected area to image file

@VogelVlug 2 года назад

Hey Sascha, I'm struggling to get you project running correctly with an own model. I trained a new model in CustomVision and just replaced your model in the project with my own new one, and I do get results, but the probabilty is a lot different from when I quicktest the exactly same image directly in Custom Vision. A prediction that gets a probability of 92.1% via the quicktest only gets a probability of 55%.

@VogelVlug 2 года назад

Nevermind, I found the problem: if (is_new_od_model) { console.log( "Object Detection Model V2 detected." ); image = is_new_od_model ? image : image.reverse(-1); // RGB->BGR for old models } With the wrapping if statement the image never gets converted to BGR for an old model. I moved the line and now the precision ist a lot better. (Still not exactly as good as directly within CustomVision, but they probably do even more pre-processing.) I created a Pull-Request in your github repository moving that line 😊

@asddasd5262 2 года назад

Getting this error in console. Model is loaded. Did not changes to the clone git Uncaught (in promise) Error: Cannot infer the missing size in [-1,0] when there are 0 elements

@elizabethcastrovillegas3231 3 года назад

Hello Friend. tanks for the tutorial. I have an error to tray to run you proyect, in the console appeear this Uncaught Error: Based on the provided shape, [3,3,3,16], the tensor should have 432 values but has 363 and the proyect dont work. how i can solve this. Tanks and regards.

@SaschaDittmann 3 года назад

Did you clone the repo or download it as zip file. I use Got LFS which messes up the model bin files when using Download as ZIP

@MajinVegeto89 2 года назад

I can't get past the error: "Error: The shape of dict['Placeholder'] provided in model.execute(dict) must be [-1,416,416,3], but was [1,2560,1440,3]"

@FrewqasZ 3 года назад

Hi! sorry to reach out to you in this rather awkward way but im really at the end of my rope and cant find help anywhere. Ive got a problem with azure customvision regarding your video about simple image recognition. I am adding my images (of fruits) and after training my model has 100% precision/recall and ap ... this makes the model completely useless. Do you have any idea what im doing wrong or where to find help?

@SaschaDittmann 3 года назад

Hi Jeremiasz, How many images are you using? Do you have a fair amount for each kind of fruits? Are you using any kind of Image Data Augmentation?

@FrewqasZ 3 года назад

@@SaschaDittmann Hello! thank you so much for answering. 1. 400 images total, 2. 200 images per category (banana/apple) 3. I am not aware if i do, so probably not. Images of fruits are downloaded from www.kaggle.com/moltean/fruits

@xhancexd 2 года назад

Hello, Do you have a tutorial to implement this to node-red?

@SaschaDittmann 2 года назад

I‘m sorry but no. But I‘m happy to add that to my list of ideas

@ahmadadib51 3 года назад

Thanks for this tutorial. But I had a little problem, the predictions could not be made, localhost continued to show "running predictions" but it never finished. I don't know why, is it because of my big model? My model reaches 61 bin files. Please help me

@SaschaDittmann 3 года назад

Might be the large model. Please check with the developer tools in your browser, if you get an error message.

@ahmadadib51 3 года назад

@@SaschaDittmann In console I got "Error : Activation leakyrelu has not been implemented for the WebGL backend"

@beautymates_de 3 года назад

@@SaschaDittmann same here as Arib Bachmid I get I got "Error : Activation leakyrelu has not been implemented for the WebGL backend" can you help? I get this after optimization. In the json file it seems that leakyrelu changed to leakyrelu_alpha, maybe this is the problem? Thanks in Advance and for this awesome video

@beautymates_de 3 года назад

got this: replace the thensflow script with

@song705 2 года назад

Hello Thanks for your tutorial can you help me with one little problem i am getting this error with my own Model Uncaught (in promise) Error: TensorList shape mismatch: Shapes -1 and 3 must match at lv (util_base.js:108:11) at OO (tensor_utils.js:28:3) at e.t.setItem (tensor_list.js:229:5) at control_executor.js:241:18 at u (runtime.js:45:40) at Generator._invoke (runtime.js:274:22) at Generator.forEach.e. [as next] (runtime.js:97:21) at Um (runtime.js:728:43) at o (runtime.js:728:43) at runtime.js:728:43

@matteoscoccia1634 Год назад

Hi, i got the same error, did you find a solution?

@manfyegoh 3 года назад

what is version 1 or version 2 @_@" how do we know the version?

@SaschaDittmann 3 года назад

There are two different model versions, which might be created by the Custom Vision service. I use the output of the prediction (length of the array) to check which model version was used.

@mirzaabdulrehman428 3 года назад

make a video on runtime detection on webcam

@SaschaDittmann 3 года назад

Great idea! Thank you!

@chings228 3 года назад

is it free for azure training model ?

@SaschaDittmann 3 года назад

Hi Tom, there is a free, as well as a paid, offering for Azure Custom Vision. The free instance currently contains: - Up to 2 projects - Up to 1 hour training per month - Up to 5000 training images per project For more details about both offerings, please have a look here azure.microsoft.com/en-us/pricing/details/cognitive-services/custom-vision-service/

@user-sq2cx6qc3y 3 года назад

Херня какая то, при регистрации пришлось засветить карту банка. И там нет тарифа бесплатного, есть тариф "с оплатой по мере использования". В ролике много вырезано. Интерфейс отличается координально. После регистрации сразу начнутся проблемы с пониманием и попытками повторить. Забейте, не тратьте время...

@SaschaDittmann 3 года назад

If you are looking for the free offer details, checkout this link: azure.microsoft.com/en-us/pricing/details/cognitive-services/custom-vision-service/

@user-sq2cx6qc3y 3 года назад

@@SaschaDittmann Ты на своём "бесплатном" аккаунте демонстрируешь тренировку 24 часа, хотя бесплатно только 1 час в месяц, судя по данным из твоей ссылки.

@SaschaDittmann 3 года назад

@@user-sq2cx6qc3y I'm sorry that it came across that way. I should have emphasized this more clearly.