How to Train TensorFlow Lite Object Detection Models Using Google Colab | SSD MobileNet
ฝัง
- เผยแพร่เมื่อ 20 พ.ค. 2024
- Let's train, export, and deploy a TensorFlow Lite object detection model on the Raspberry Pi - all through a web browser using Google Colab! We'll walk through a Colab notebook that provides start-to-finish code and instructions for training a custom TFLite model, and then show how to run it on a Raspberry Pi. The notebook uses the TensorFlow Object Detection API to train SSD-MobileNet or EfficientDet models and converts them to TFLite format.
Click this link to the Colab notebook to get started: colab.research.google.com/git...
-- Other Links --
📸 How to capture and label training data for object detection models: • How to Capture and Lab...
🏅 TFLite model comparison article: ejtech.io/learn/tflite-object...
🍓 Instructions to set up TFLite on the Raspberry Pi: • How To Run TensorFlow ...
💻 Instructions to run TFLite models on Windows: github.com/EdjeElectronics/Te...
🐜 How to quantize your TFLite model: Still to come!
📄 TFLite GitHub repository: github.com/EdjeElectronics/Te...
-- Chapters --
0:00 Introduction
1:06 Google Colab
1:41 1. Gather Training Images
3:22 2. Install TensorFlow
4:43 3. Upload Images and Prepare Data
8:41 4. Set up Training Configuration
11:20 5. Train Model
13:48 6. Convert Model to TFLite
14:20 7. Test Model
17:50 8. Deploy Model
22:07 9. Quantization
22:30 Conclusion
-- Music --
- Blue Wednesday - Japanese Garden
- Provided by Lofi Records
- Watch: • Blue Wednesday - Japan...
- Download/Stream: fanlink.to/Discovery - วิทยาศาสตร์และเทคโนโลยี
Thank you for a wonderful breakdown of all the needed steps to do training on colab ❤❤❤ Your follower from Egypt 🥰
Thank you for a wonderful breakdown of all the needed steps to do training on colab.
Ive been banging my head on my desk trying to find a good tutorial. Big thank
Amazing tutorial! One of the most useful videos I've ever watched. Keep up the good work!
Thanks a lot, Edje!
Really excited your back! Thanks for the video.
Really amazing video with detailed and well-explained steps. Great job!
Your tutorial is the best on the internet. Exactly what I needed! Thank you! Thank you! Thank you!!!
Thanks for making this amazing video! It is really helpful. Look forward to the "quantize TFLite model:" video too!!!
That was so detailed and easy process! Thank you so much for the effort you have been given for all those videos. Quick question, what if somebody wants to use rstp streams instead of raspberry or USB camera on the Raspberry Pi, what should be change and where? Thanks again.
Thankyou so much you made my day, I was searching all internet for working colab notebook for this model.
Thankkk youu sooo much for the tutorial, I've been researching on the internet but didn't get the solution. Great work you've earned a subscriber.
Really nice video. Anxiously waiting for an Android deploying tutorial!
Wonderful, thanks for the efforts for explaining step by step
Amazing!! Incredible to see you using the model on another device.
thank you so much for uploading this video. Helped a lot!
:D
Thank you! Clear and easy to understand while having sprinkles of humor made this video really educational AND enjoyable. Easiest like and subscribe.
Thanks for the kind words! 😸
@@EdjeElectronics There are a few persistent warnings in the notebook now because of which it does not run anymore. Kindly check the comment sections and reply. The training does not commence in step 5. It worked once when we tried the notebook on a smaller dataset but then now we are trying to deploy the full dataset it shows this issue and does not run.
I struggle a week trying the repo from Google and this video save my life thank you I would like to see a video of posenet in Raspberry Pi
Thankyou somuch for your helping with this video, now can implement this tutorial in my final task
This tutorial was super helpful, thank you a lot! It is very easy to follow along and understand with step-by-step instructions
Please keep making great videos. Awesome content.
Still is one of the best tutorial for model training in TH-cam. But I have a suggestion, could there be an option to add metadata scripts to the notebook? this could be highly useful since of the release of googles ML kit for android & ios development.
gracias por los vídeos, me ayudo mucho a entrenar modelos.
Thank you for the great content. Waiting for the deployment of tflite on the android studio part. Thank you
Good to see you again!
Thanks! I'm going to try and hold myself to a more regular release schedule now... we'll see how it goes! 😁
Great video tutorial. Thank you!!!
thank you broo, you're amazing💣
Why does the model only detect one class during the testing process even though the dataset has three classes?
Great videos. It works well thanks.
LIFE SAVER MAN THANKS
Thank you. Very nice guide for beginner. Please do video for audio classification using microphone with raspberry pi.
for anyone who noticed that not all images are getting moved to its respective folders, and that later it gives some error that a file is missing: make sure all your images end with .jpg instead of .jpeg since all the scripts on the github page only takes into account those files
Great tutorial! I have a question though. Do I need to pre process my images to become accurate or no need? I've been training my model for a weeks and I already have a 1000+ datasets but some of the datasets are not accurate. Help please.
Great tutorial!! Great introductory experience while also providing everything needed for learning rabbit holes.
If you come across Step 5 ending with ^c, it is because you are running out of memory. Either resize your images to be large/medium (~200-300KB per image), upload less images, or both.
Thank you for the tutorial video! And I have a question how can I export the classification output data and save in the txt file?
Hello, first thank you for sharing your workflow with us. How can I get the detected objected (as text) scripted when i run the model?
Any tips?
No fucking way u are givin us for free this kind of information, i love u so much, i have spend the last 2 months trying do a tensorFlow Object Detecion works, u deserve the best things of world can provide thanks thanks thankssssssssssssssssssssssssssssssss
Hey all, just a quick update (April 10 2024). I ran through the full notebook today with my coin dataset, and everything worked without errors. A "Restart Session" option appears during Step 1 after running the last set of install commands. When it appears, click the "Restart Session" option, and then keep working through the steps. If you're getting errors, try using my coin dataset and seeing if it works (it should). Then, compare your annotation files with the annotation files from the coin dataset. It's likely there's a difference in your annotation files that are causing problems.
Hello Edje, thanks for this great tutorial! While following exactly the steps in your notebook, I ran into an error repetitively when running model_builder_tf2_test.py: "Could not load dynamic library 'libcudart.so.11.0'; dlerror:.." Any idea of how to solve it?
i unable to make labelmap.pbtxt on step 3.3, how do i fix this?
Really thank you very much for the tutorial I was able to do the inference on a Raspberry Pi. However, just after the entire model I evaluated the precision of the model on the training and test data but the script that I I wrote to myself about errors since the images were saved in tfrecord format. Could I have a hand lesson?
Please make a video for android studio version. Thank you. Great tutorial by the way
Thank you for the wonderful breakdown !! Can you do a tutorial on how to deploy on android please
Thanks for the video!!. I have some doubts, if I would train a model for detecting no ok parts, the way you showed us will fit for my model?, Must I take a picture for any defect and label it? Must take pictures for ok parts and label it also?. I'm trying to detect defects like scratches, marks, paint defects, etc...any help will be much appreciated!!
Excellent video. I have a query regarding the application of the model. If, for example, I require that the detected elements be registered, what script would you recommend I use to do this? Thanks.
great tutorial thanks, can you explain how to use the TF lite model in real-time detection using a webcam and how to deploy the model on a mobile application
thanks for best video ever please make the application we are waiting you ...thanks
Can we train for image segmentation model too with this codelab?
Really detailed. A query though, where in the above procedures should I rectify to train a face detection model??
The best tutorial in the world! :) Please could you modify the code on Google Colab in the future so that training also works for negative images and xml files without objects in them? Now it gives errors. Thank you so much, we love you!
First of all, thank you very much for the great work. I have a question for you. I can't convert tensorflow lite version to tensor flow js how should I write a code block about it
Thanks for the helpful tutorial, but I want to make an inference on the esp32 cam, can you provide a solution?
Great tutorial...............
Hello! Thanks for the tutorial. It really helped a lot. But, I'd like to ask, besides the mAP evaluation metrics, how do I include other metrics such as Kappa, F1, confusion matrix, etc in order to provide more information about the model.
Thanks!
i need to
kkkkk
worth video Thank you so much
Hello, can we have a tutorial about implementing the tensorflow lite model that is made in this video to a mobile object detection application. Thank you very much and looking forward to it
Your video have been so useful to me, i just would like you make a tutorial to run it on a mac machine
Hello, to increase performance, you can use multithreaded computing. To do this, when loading the model, specify the "num_threads" argument, which must contain the number of threads that the processor supports. Basically, TFLite uses only one core.
The code:
import multiprocessing
tf.lite.Interpreter(model_path=PATH_TO_CKPT, num_threads=multiprocessing.cpu_count() )
On my old laptop with dual-core CPU, this gave a double increase in performance.
Basic - 140 ms per frame.
With two cores - 85 ms per frame
Thanks, I'll have to try that!
This works like a charm bro! On a pi 4, My fps went from 8fps to 20fps when using the google sample tflite model!🤯
@@khamismuniru5188 how
@@sumit_ux Section 7.1. In the line -> interpreter = Interpreter(model_path=modelpath)
@@fe_nik_s776 you mean like this?
import multiprocessing
tf.lite.Interpreter(model_path=PATH_TO_CKPT, num_threads=multiprocessing.cpu_count() )
interpreter = Interpreter(model_path=modelpath)
Hello, friend!! Many thanks for this utterly useful video. I use trained model in Home Assistant addon DOODS and can't deal with error: AttributeError: 'TensorflowLite' object has no attribute 'input_data''`
What do you think what i need to change in google colab sequencing in order to solve this issue?
Hi wonderful video, I'm working on a custom dataset with more than 4 classes what file or code would I change?
Hello! Really informative video, I just wanted to know the process stays or less the same even if im on a Mac while making the project right? Is there anything I should be aware of?
Furthermore its possible for me to make it so that the software detects when I am bending/bent towards the right/left and give an output, or because its the same person it wont be able to recognise that?
thanks!
Thanks! It should all work inside the web browser, even if you're on a mac. However, I still haven't written the guide for how to take the downloaded model and run it on macOS. You should take a stab at it though! In fact, I'd pay you $50USD if you can write a macOS deployment guide similar to the one I wrote for Windows (github.com/EdjeElectronics/TensorFlow-Lite-Object-Detection-on-Android-and-Raspberry-Pi/blob/master/deploy_guides/Windows_TFLite_Guide.md). Email me if you're interested in that - info@ejtech.io !
For detecting which way you are bent, I would try using an OpenPose model (look it up on TH-cam to see what I mean).
Good news! Training on GPU is working again. Thank you @markwassef8643 for finding the solution.
great video! thanks!
Hello.
Please, is there a way to change the label style (e.g., font and size) and the colour of the bounding boxes for each of the classes drawn on the test dataset (the detection results)?
Thanks
Hi, may I ask if the Tensorflow Lite can be deployed on the Nvidia Jetson Nano?
Really good tutorial, I am following along for my school project, I can get as far as to the training, then it won't start training in Colab, in the log, it shows that the TensorFlow Addons needs tf version 2.12 - 2.15.0, would that be the case? I can see your video using python 3.8, currently the colab is run on python 3.10, could that be the reason? Could you please help? Thank you
Thank You for your tutorial, it has been of great help to me but, I when upload the images folder and run it to distribute data for training, validation and test it show me zero images in total
Please guide me in this regard
How to add images with null annotations to it? Should we just add the images or should we have to create XML files for it?
Hello, can you please demonstrate how to configure the model output with classes as TensorBuffer, locations as TensorBuffer, scores as TensorBuffer, and numberofdetection as TensorBuffer?
Can you do this with screenshots from a website to train it to identify certain images on a website?
Great video intesetd if you have
Instructions to run TFLite models on mac os
does this work on raspberry pi camera module together with the rpi 4b? thanks
which model is best for human detection
Thank you so much a value able content. Could you please implement the tflite modal in android application
Love your tutorial!! good job with the instruction. I kind of using your material as my object detection experiment for my own dataset. But I got stuck at the training. seems like there is a version deprecation issue due to the colab upgrading TF version. I am wondering if there is an update to the code that helps at stage 5. TFLite model training :)
Thanks! Sorry, I'm not sure what's going on, and I don't have time currently to look in to it. Can you try Google searching the error to see if there's a resolution?
Hey, I was facing some issues as well. Make sure to run the session with the GPU. When I was training a larger model without the GPU, it hung on the training initialization. I also decreased the batch-size but I'm not convinced that it's actually necessary. Btw, thanks @EdjeElectronics for the tutorial, I really appreciate it!
Yeah thanks you two for replying, and can’t thank enough to @EdjeElectronic for providing the tutorial. So my issue turns out to be that my dataset contains .csv for annotation. So I had to ask GPT to write me a code to convert the labeling datasets back to .xml w/ exactly the same format as the dataset you provide for the coin dataset. I did it offline with Jupyter Notebook and re-upload them back to Gdrive and continue using your code to train, and after 2 weeks of tweaking, I finally made it to train with your magical instruction:)
I looked into every line of your code and went down the rabbit hole with even the extension .py you linked and trying to study the logic behind the code w/ the help of chatGPT :D
Long story short, I got it to train and it’s training. Yay!!!
Ok so i figured it out. I realize that the entire process requires xml and image dataset. i had csv and image as my dataset. So i converted the csv labeling dataset back into xml format locally and re upload the dataset and run the code and worked!! yay~✌😇
@@PeterPan-hs5tu Awesome, glad you were able to figure it out!
Hello and thank you again for this tutorial!
Today I was successfully able to get to the training step with my own dataset but I unfortunately forgot to turn my computer's sleep mode off. When I went to go back through the steps I keep getting an error Step 4 Cell 4 commented: "# Set file locations and get number of classes for config file". The error says "NotFoundError: /content/labelmap.pbtxt; No such file or directory" I can share a screenshot if you like. I've ran through the notebook several times and have also reset it with your most current version and I keep getting this. And in fact, there is no labelmap.pbtxt in my files. There is a labelmap.txt though!
Thank you!
yeah, the .pbtxt does not get created in my case either.
I ran this in a seperate cell to create the .pbtxt file:
path_to_labeltxt = os.path.join(os.getcwd(), 'labelmap.txt')
with open(path_to_labeltxt, 'r') as f:
labels = [line.strip() for line in f.readlines()]
path_to_labelpbtxt = os.path.join(os.getcwd(), 'labelmap.pbtxt')
with open(path_to_labelpbtxt,'w') as f:
for i, label in enumerate(labels):
f.write('item {
' +
' id: %d
' % (i + 1) +
' name: \'%s\'
' % label +
'}
' +
'
')
have you fix this problem yet?
Its great 👍
Please tell me about Python. It installed on your Raspberry pi?
Good afternoon, great contribution that you really make, congratulations, a question when doing the training is paralyzed and does not continue on API in Colab, what could happen?
Sorry, I haven't encountered that problem before. My guess is to restart the Colab from scratch and try to run through it again. Otherwise, I'm not sure how to solve it.
if you do everything right and it stuck on "use tf.cast instead" without any error. The problem is your train.tfrecord is empty. you need to change your xml files to make it work. it should look like this without any extra rows or parameters. Dont worry only important thing in this code is object name, min/max values and filename. If you have any additional row in your xml just manually delete it or use different label tool
datasetjpg
20.jpg
C:\Users\Kaan PC\Desktop\datasetjpg\20.jpg
Unknown
1920
1080
3
0
car
Unspecified
0
0
1205
38
1900
387
Thanks for posting this solution!
@@EdjeElectronics I want to train my own model and there is only one class for example car. Just one. But section 3.3 doesn't accept it. And it give a error masege " penny, nickel, dime, quarter use " like this. What should ı do?
I did this but it's still stuck in there? What else can be the solution?
Hi! I'm getting an error when training a custom model (Train Custom TFLite Detection Model) and a warning:
TensorFlow Addons (TFA) has ended development and introduction of new features. TFA has entered a minimal maintenance and release mode until a planned end of life in May 2024. Please modify downstream libraries to take dependencies from other repositories in our TensorFlow community (e.g. Keras, Keras-CV, and Keras-NLP).
Just wanna ask! If i just want the tensorflow model and not the tflite one? What should i do?
will the datasets from roboflow work with this colab notebook?
Thanks for the great content. I have some doubts, the video displaying is running at 2.7 FPS with Raspberry Pi, with you it reaches about 5 FPS, why my one is slower?
Hey bro, were you able to complete the custom object detection... I had an issue during the training process whereby it will stop after 50s. Usually it should take several hours. I also noticed that my system Ram reached about 12.1Gb before it stopped... did yours do the same? I would be really grateful for some guidance...
Hello, i was wondering what framerate I could get on an android phone using this model ? I need to train my model to detect potholes and road hazards
hello good sir, this is a very helpful vid. Currently waiting for the how to capture and label training data video
It will be out on Monday! Just finishing up the editing now.
@@EdjeElectronics happy to hear that sir, thank you for uploading such helpful videos
Can you make video or provide codelab to train for image segmentation with cnn
Dude can you update the tensorflow lite on raspi for Obj detection. a good chunch of the requirements on the get-req bash are deprecated. Thanks!
can you do one with medical image ? annotate and train
Are there any companies that offer training services where you just upload images and they send back to you the files needed to run image detection?
Hey, how could I make it so that when running the program with the pre-trained files, the camera detects 50% of something and saves what it has detected in a text file?
Great video. please can you add the video of TFlite model deployment in android devices?
what is inside the contents/training/train folder it is not generated in my case, getting problems while loading the tensorboard
So everything works great! Just wondering how to use it with Coral USB accelerator? Tried to run same code with -edge tpu
Hey there. I am facing issues with training part. It showing dependencies issues with keras and tensorflow. Could you please help me with this problem
Hello @Edje Electronics Thank you for this amazing project, I must i has really been helpful and a life saver. I have question though. Is there like a script that i can use to resume training assuming i have the model saved in directory that is not temporary? Thank You in anticipation
Hi Abraham, yes, there should be a way to do that, but I don't have time to remember it right now. Hope you can figure it out!
@@EdjeElectronics Thank you for your response. Yes i figured it out.
Also, do you have a video for how to auto launch the python scripts at boot on the raspberry pi, that i don't have to login to the desktop environment before starting my script? Thank you.
hello sir can you tell me how did you resume the training? I am having a really big trouble training my dataset since I have 6452 images on my dataset and I was training it for 40k steps just to have atleast good accuracy but it always get timeout before it even done training into 30k steps I wonder how did you make to resume it I really appreciate hope you share it
I had a doubt by the way.
1) What was the reason you've used the tensorflow 2.8 model. Is it like only due to compatibilty issue in Colab.
2)Could i run the same code in my local PC, if yes then how should i Install the CUDA Libs I have and RTX 4060 Gpu. So could I run the code normally depending upon the latest version of tensorflow
Please help me out with this
Thank you...
Caused by: java.lang.IllegalArgumentException: Cannot copy from a TensorFlowLite tensor (StatefulPartitionedCall:1) with shape [1, 10] to a Java object with shape [1, 4].
I am getting this error using custom model in my flutter app.
Can i run the whole program of the collab in visual studio code ?
hello, thanks for your tutorial, how can increase number of boxes in model config(not detect more than 10 object per image) can everybody help me please?
Great tutorial! I've been encountering a problem though with my dataset. All file names are correct, also the file extensions, and have tried all options of adding a dataset (upload, gdrive mounting, dropbox) I even checked the csv file and all are intact. However, when its time to create the tfrecord files, there appear to be some files missing in both the train and validation folders, thus I can't generate the pbtxt file. Any tips?
Thank you! Hmm, usually if there's an issue with the pbtxt file, it's because there's a problem with the annotation data (like a typo in a class name). Can you try using my coin dataset and see if you get the same issue?
Thankyou
Great tutorial!
I have just one question.
I would like to use my model with opencv and use the x,y,h,w of the targeted object for other stuff..
is there an easy way to do that?
did you ever figure this out? I am doing the same
Good Job !