How To Train an Object Detection Neural Network Using TensorFlow (GPU) on Windows 10 

Hey everyone! I recently updated the written version of this guide to work with TensorFlow versions up to 1.13.1. If you are encountering errors following this video, please check out the guide and make sure you are using the most up-to-date commands. Here are answers to some common questions: - Where is the train.py file? The train.py file is now located in the /object_detection/legacy folder. You can copy it to the /object_detection folder and use it as normal. - Images aren't appearing when testing the Jupyter notebook, even though there were no errors? Try this: go in to object_detection/utils/visualization_utils.py and comment out the import statements on line 25 and 26 that include matplotlib. Then, re-run the script. - Can I run this just using CPU and not GPU? (Yes, just use "pip install tensorflow" instead of "pip install tensorflow-gpu".) - How to resolve errors related to _pb2 files? See Step 2f of the guide linked below. Check out the guide at github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10 !

13:38 -> you will find the setup.py in the slim folder copy that into research folder 13:59 -> Object_detection_tutorial.ipynb - Is located in the "object_detection\colab_tutorials" for anyone who can't find it 24:19 -> according to fixed comments, train.py file is now located in the /object_detection/legacy folder. fpn_pb2 -> protoc --python_out=. .\object_detection\protos\fpn.proto center_net_pb2 - > protoc --python_out=. .\object_detection\protos\center_net.proto official -> pip install tf-models-official etc... ㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡㅡ If these error were found, I think you should start again from the beginning

weeks on the search for a good start tutorial that doesn't have 50+ parts and this is where I land. One of the best clear and concise tutorial/introduction video I've seen so far! Keep it up!!

This is the exact explanation that I needed it about how the training process gets "stored" with the 5 minutes checkpoints by tensorflow. I'll try to replicate this in a Mac environment and then somewhere in the cloud. Thanks for this video, you rocks!

You deserve a lot of credit for making this video. There are many tensorflow/machine learning tutorials online but few that really show you this detailed from getting started to creating your model and training it.

Here are the issues that have been found so far with this tutorial: - TensorFlow moved the train.py file used in Step 6 to the "legacy" folder inside the /object_detection folder. You can still use it by moving it back into the /object_detection folder. - If the images aren't appearing in Jupyter notebook at the end of Step 2, try going in to object_detection/utils/visualization_utils.py and comment out both the import statements at the top that include matplotlib. Then, try re-running the Jupyter notebook. - Starting at 21:41, I say to change the num_classes variable to 6, but I accidentally changed it to 36. It should be 6. (Thanks Rafael!) - Google may add more .proto files to the object_detection/protos folder, so it may be necessary to add more files to the "protoc" command at 13:13. You can do this by adding ".\object_detection\protos\FILENAME.proto" to the end of the long command string for each new file. (Thanks exnaruto1!) - When running the "python train.py" command, if you get an error that says "TypeError: __init__() got an unexpected keyword argument 'dct_method'.", then remove the "dct_method=dct_method" argument from line 110 of the object_detection/data_decoders/tf_example_decoder.py file. (Thanks Evpatoria!) - When running "python train.py", if you get an error saying "google.protobuf.text_format.ParseError: 110:25 : Expected string but found: '“' ", try re-typing the quotation marks around each of the filepaths. If you copied the filepaths over from my GitHub tutorial, the quotation marks sometimes copy over as a different character type, and TensorFlow doesn't like that. - For train.py, if you get an error saying "TypeError: Expected int32, got range(0, 3) of type 'range' instead.", it is likely an issue with the learning_schedules.py file. In the \object_detection\utils\learning_schedules.py file, change line 153 from "tf.constant(range(num_boundaries), dtype=tf.int32)," to "tf.constant(list(range(num_boundaries)), dtype=tf.int32),". - If you are still getting an error complaining about "Expected int32, got range(0, 3)" or something similar, try implementing this fix: github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10/issues/11 Please let me know if you see any more errors with the video or with the commands/instructions from my GitHub repository!

very less start from scratch tutorials on youtube Thank you!!

What a fantastic Tutorial. I love the thoroughness. :)

God dude the beginning camera footage is so cool to watch

Spent a whole night with this. Totally worth it. Amazing tut man...

One of the not just best but one of the greatest video ever made in youtube for beginner .We can only say that you are the best thank you.❤❤❤

One of the cleanest videos i´ve ever seen

thanks for all your hard work you put in making this video

a shortcut while labelling the images is to make use of w - to start capturing the area (selecting) d- to move to the next image also , auto save your work

Patience required to complete, very excellent tutorial

Thank you so much! i really appreciate your tremendous contribution to this community. I managed to adapt this tutorial to tensorflow2.0 and it worked!... Thank you! Like and new sub

Bro why is the code content in object_detection_tutorial.ipynb not same with the tutorial you make?

thanks for this tutorial, was looking for something exactly like this, other videos don't go into details like, just skip here and there bam already trained data and detecting, i wanted to know how it looks like from start to finish cuz it was hard just trying to grasp text tutorials.

Thank you for making this tutorial. Massively helpful!

@Edje Electronics I would like to report using your tutorial as a starting point and a lot of googling! I was able to: 1. Translate all the required information to linux (Ubuntu 18.04) 2. Get Tensorflow Rocm working (Vega 56 AMD GPU) 3. Swap out the dataset with my own. Vega is humming along as I type this. Training now. Thanks for putting this detailed tutorial together. :)

For those of you looking for the new zoo model link: github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md The Jupyter notebook is now in: C:\tensorflow1\models esearch\object_detection\colab_tutorials Update: I got the project to working.. what you guys should really do is: Clone the git repository and then switch to the branch version of the video, this way you guys will have the same directory structure as shown in the video.

Hi, is there a way to get only the precision, recall, true positive, false positive and false negative values? I found that these values are calculated in the metrics.py file but I am unable to print/write these results to a csv file.

THis tutorial is absolutely awesome! I love it so much. I can't thank you enough.

24:15 i keep getting the following error - ImportError: No module named 'tensorflow.compat' - is there any solution to this, ive scowered the internet and cant find one

Desperately waiting for your video on implementing this on raspberry pi

how did u make the model tho? Like what loss function did u use??

hey can someone tell me which model works for tensorflow 2 ? as this video is only for tensorflow 1

For those who have been stuck with choosing versions of software and modules to use, this is my additional suggestion. I am using Windows 10 pro. - Installing Anaconda: I chose the last version of anaconda released in 2018 which supports python 3.7.1. Technically, you can use the latest released version. - Python version: 3.7.1 - Tensorflow ver 1.13.1 goes along with its Github commit: github.com/tensorflow/models/tree/r1.13.0 - NumPy version 1.16.4 - pandas version 1.1.5 - Matplotlib version 3.0 For the rest, I just follow the author's instructions and the support from other guys here. Thanks a lot.

Great Tutorial! Go to 16:20 for cat-cable ultimate battle!

To everyone having issues with tensorflow 'app' or tensorflow 'contrib' . You need to downgrade your tensorflow version to 1.14. Using tensorflow 2.0 onwards doesn't work. even if you use compat.v1, since 'contrib' was completely removed from tf 2. EDIT: for those of you wondering how to do this, go into terminal and activate tensorflow1 environment pip uninstall tensorflow Wait for it to uninstall pip install tensorflow-gpu==1.14 Wait for it to install and you should be good to go. Be aware that you will have to download the correct versions of CUDA and CUDNN

what is the different between this model and yolo model?

Hi , if i want to train a specific type of object (eg. water bottle), for my image do i include 2 water bottle and only label one of it in the image?

Does anyone know how to count the detected objects?

i couldn't get the images at 15:00 ,please help.

Hello, How do I allow continual training of my object detection model?

do you know if i can use the ibm cloud anotation tool to mark the images?

when i try to run train.py AttributeError: 'KeepAspectRatioResizer' object has no attribute 'resize_method' please help

Sir @Edje Electronics, does this work even if im not using Tensorflow-Gpu?

how can i see the accuracy graph like loss graph using tensor board? please say me.

can we use ssd_mobilenet_v2_coco instead of ssd_mobilenet_v2_quantised_coco model for this

Sir can we run this in Google colab and if possible can you help me with the steps

Whenever I run the python setup.py build /install command, it says it cant find a setup.py file in the directory of C:\tensorflow1\models esearch>

is the test set same as validation set?

Could you tell me how I print name of object as text when detected?

For Tensorflow 2.0 use following command to upgrade generate_tfrecord.py on anaconda promt tf_upgrade_v2 --infile generate_tfrecord.py --outfile generate_tfrecord_new.py Also in line 23 replace # flags = tf.app.flags flags = tf.compat.v1.flags

I am getting: AttributeError: module 'tensorflow' has no attribute 'contrib'. Any solution?

Edje Electronics Thanks for tutorial it was of great help. 🙏

what alternative tool do i use instead of labelimg to labes non linear lines (transmission lines ) ; labelimg uses boxes which isn't helpful in my case

Hi! I have no words to thank you. You really helped me. May Allah Pak bless you. Bundle of thanks :)

Hi, want to ask, can i still done this if my laptop's graphic card is not Nvidea? Using intel graphic

i need to classify my dataset of car make models images files with tensorflow but i work with more then 1600 car models ? which one i should choose it as a detector ? and i should use labelimg ? or i just work with the image ?

@EdjeElectronics Hi, I'm in the middle of Step 6 trying to run the training, but when I do so, I come up with "ValueError: No variables to save. " Does this issue sound familiar?

I have a problem while initialising the training File "train.py", line 48, in from tensorflow.contrib import framework as contrib_framework ModuleNotFoundError: No module named 'tensorflow.contrib' If any one have solution to this please help.

how can i print the detecetd objects ?

very good, I don't do windows 10 but it was very helpful for object recognition and training.

@Edje Electronics, I am getting error saying that the faster_rcnn_inception_v2 is not supported when running the train.py file

Great tutorial man! My question is does this tutorial work with the current latest versions of tensorflow (v1.12), cuda (v10) and cuDNN (v7.4.1)?

The best tutorial on youtube by far!! Thank you so much. I did everything and works great, my question is: in Tensorboard, in the Projector Embeddings, you see 1024 points, 396 dimensions. I'd like to extract the data (classifiers labels and images) to create the tsv and sprite, or whatever is necesary to see the images in the projector. How do I do that_ I'd great if you make a tutorial or upload code to do that.

faster_rcnn_inception_v2 is not compatible with tensorflow v2.x Is there any way other than uninstall current version and install 1.x older versions? Would you recommend me other model compatible to current tensorflow version?

can this isntallation and setup be done for the the CPU version as well?

So, I ran this successfully (with the occasional hick up here and there, especially around the protocs) and because I don't have a GPU, it trained over 5 days on my CPU. But I got there. Thanks to Edje! I'm recognizing parked cars, as the goal is to count the amount of parked cars in my neighborhood (for a community project) without manually counting them. How would anybody take the next step and count the amount of recognized parked cars? Suggestions are highly appreciated!

hey, just a small question .. can i use the trained model from my pc and run it on a pi?

Hi for step 5b. If I selected ssd v2 model to be used, Won't I have to edit that and not the the faster v2 interception.pet?

Hi, how can we write down the information appeared on the rectangular boxes on the console? Example: if there is a car in the video, how can we get the information in console, saying this is the car appeared at this time in the video can we do that ?

This video was very useful. Thanks a lot. I had 2 errors when training. First one TypeError: __init__() got an unexpected keyword argument 'dct_method'. Fixed this one by removing the 110th line "dct_method=dct_method" in object_detection/data_decoders/tf_example_decoder.py. Second error was this: TypeError: Expected int32, got range(0, 3) of type 'range' instead. Fixed this one like this: Go to object_detection/utils/learning_schedules.py Replace line 153 with this: tf.constant(list(range(num_boundaries)), dtype=tf.int32), So if anyone else has the same issues this is how I fixed it.

Hello, here are some tips on what helped me run this model. It includes the versions of the packages that i installed aswell as some notes on how to work arround the steps 1 and 2 of the GitHub manual. I have yet to try and run it with my own personal images so don't expect automatic results... NAME OF MY CONDA ENV: tf1.13 GPU: nvidia quadro k620 VERSIONS: cudnn 7.3.1 cuda 10.0_0 tensorflow-gpu 1.13.1 tensorflow-estimator 1.13.0 tensorflow 1.13.1 pip 10.0.1 protobuf 3.6.0 python 3.5 pillow 7.0.0 lxml 4.4.2 cython 0.29.15 contextlib2 0.6.0.post1 jupyter 1.0.0 jupyter-client 6.0.0 jupyter-console 6.1.0 jupyter-core 4.6.3 matplotlib 3.0.3 pandas 0.25.3 opencv-python 4.2.0.32 INSTALLING: pip install --ignore-installed --upgrade tensorflow==1.13.1 conda install -c anaconda cudnn==7.3.1 -> equivalent to 7.4 CHECKING: conda list NAMEOFPACKAGE SET PATHS: set PYTHONPATH=C:\tf1.13\models;C:\tf1.13\models esearch;C:\tf1.13\models esearch\slim CHECKING: echo %PYTHONPATH% PROTOBUF COMMAND EXCEPTIONS: REMOVED FROM ORIGINAL \object_detection\protos\calibration.proto . -> DOES NOT EXIST WITH TF1.13 \object_detection\protos\flexible_grid_anchor_generator.proto. -> DOES NOT EXIST WITH TF1.13 SETUP BUILD: cd C:\tf1.13\models esearch python setup.py build python setup.py install EXECUTE LOCALHOST DEMO: cd C:\tf1.13\models esearch\object_detection jupyter notebook object_detection_tutorial.ipynb LINKS: TENSORFLOW / PYTHON / CUDA / CUDNN - VERSIONS : www.tensorflow.org/install/source#tested_build_configurations PREBUILT MODELS: github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md ssd -> low CPU MAIN PROJECT: github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10#1-install-anaconda-cuda-and-cudnn I also ran into some DLL problems when executing the jupyter notebook because my Python folder had two. It only took removing one of them to work... I hope this helps you!

In which portion of the object_dectection_image.py file contains the "comparison between the trained model and test image code"? Could you please help me with the solution??

Can this be done on mac os if you run the cmd prompt commands in terminal?

facing error like this File "generate_tfrecord.py", line 23, in flags = tf.app.flags AttributeError: module 'tensorflow' has no attribute 'app' please help

python: can't open file 'C:\tensorflow1\models esearch\setup.py': [Errno 2] No such file or directory Can you help me ? I take this error.

hey, Edge electronics am able to get the output for webcam but I am not getting the output for the image I have changed the number of classes. the python code is running and displays the output but without detecting the image [like there are no rectangular boxes and on]

hi! I have a problem with detection.py, I try to do as your video but when I run detection (for image) the result is nothing, that mean do not have bonding box.

please help me with this error: (tensorflow1) C:\tensorflow1\models esearch>python setup.py build python: can't open file 'C:\\tensorflow1\\models\ esearch\\setup.py': [Errno 2] No such file or directory

I had this problem running jupyter file: "ImportError: DLL load failed importing win32api". Just type "conda install pywin32" and reaload. It works for me.

Hi. On your example,your frozen file model is about 50Mb. But mine is 200 or 300 Mb. Why is this big? Is there a problem with the construction or anything im doing? And the file size impacts on video detecting? Thanks for your help.

how to use this repo for the BDD100k dataset? Because the main problem is in BDD100k, all the training labels for all the training images are in a single Jason file unlike the one shown here where there is a separate .XML label file for each image! Thanks in advance!

I've got an error, plz help me out (tensorflow) C:\tensorflow1\models\models-master esearch>python setup.py build python: can't open file 'C:\tensorflow1\models\models-master esearch\setup.py': [Errno 2] No such file or directory

i view don't have file setup.py in locations file C:\tensorflow1\models esearch

@Edje Electronics Have you any idea how to execute the custom object detection model on tf 1.x after tensorflow Up-gradation to 2.x because when I cloned the repo it generates many issues with my model? I already run my model but after up-gradation of tensorflow (2.x) it doesn't work properly.

Can I take the model and run it from my raspberry pi 4? I am working on a project and I want my model to detect only people. I know is the same process as in your video, but how can I take the model?

I had trouble when trying to generate the TFRecord with this command: "C:\tensorflow1\models esearch\object_detection>python generate_tfrecord.py --csv_input=images\train_labels.csv --image_dir=images\train --output_path=train.record" It says: "from object_detection.utils import dataset_util ModuleNotFoundError: No module named 'object_detection'". Does anyone knows how to solve this?

21:45 : 6 classes you mean 6 and not 36 right?

Thank you, this video really helped me a lot!

Hi regards from colombia, I would like to thank you for this exceptional video, for me is very useful this content for my grade work thanks. I haved many problems during the execution of the tutorial caused by versions of the libraries but finally works perfectly.

This is the old tutorial using TF 1.x, for Tensorflow 2.0 I advise you to visit tensorflow-object-detection-api-tutorial.readthedocs.io/en/latest/training.html for further training you custom object detector.

These settings seem to work for me : Tensorflow 1.13.1 Tensorflow-gpu 1.15 Cuda version 10.0 *Important* : From tensorflow model github repo, Only clone from the branch "r1.13.0" otherwise it will not work

Excuse me, when I run the protoc --python_out, there's one warning, it said ( warning: Import object_detection/protos/image_resizer.proto is unused.) what should I do with it?

does object detection algorithms classify on size and colors too?

When I run "python setup.py build" I get the following error: "[Errno 2] No such file or directory". Can anyone please help? Would be very kind

Hello sir, Your tutorial is so great and worked for so many people in the past but now with the updated tensorflow repository i am facing many errors and many mismatch. I am able to solve many of these errors and all but yet couldn't get any result after trying so many days. Would you mind making another new updated tutorial for object detection? That would be tremendously helpful. Thanks in advance.

Hi, when I type conda install -c anaconda protobuf, it install and when I pip install anything after that, it gives me index out of boud exception even if I type pip install numpy or any package

Excellent tutorial! But I have a problem. Is there any way that during the training I can save from one step to stop the process without losing the past steps? It seems that my program wont save any progress during the training

I figured it out. If anyone gets the same error, after the final "else" in the generate_tfrecord.py change the last line to "return 0" instead of "return none" in the class_text_to_int function.

So, for anyone who is trying to use this tutorial in january 2020, with a ton of help I've finally made it to work out. Here's how: For Tensorflow-GPU: Download and install NVIDIA CUDA 10.0 Download cuDNN 7.4.2 for CUDA 10.0 -> open archieve and navigate to the 'bin' folder -> extract the only dll file from there to the CUDA v10.0/bin directory Your Tensorflow-GPU is ready to work! If you want to use Tensorflow-CPU ignore the steps from above Next: Download the latest version of models from the github repo linked in the description Download the faster_rcnn_inception archieve and extract it into the models/research/object_detection Download the Object-Detector API from the repo and extract it's files into models/research/object_detection Inside models/research/object_detection delete the inference_graph folder content, as well as the 2 csv files from images and the content of the training folder Open an Anaconda Command prompt and create a virtual env: C:\>conda create -n tensorflow1 pip python=3.6.10 //the 3.6.10 version its working with my overall setup C:\>activate tensorflow1 (tensorflow1) C:\>pip install tensorflow-gpu==1.13.1 //the 1.13.1 version is working as well (tensorflow1) C:\> conda install -c anaconda protobuf (tensorflow1) C:\> pip install pillow (tensorflow1) C:\> pip install lxml (tensorflow1) C:\> pip install Cython (tensorflow1) C:\> pip install contextlib2 (tensorflow1) C:\> pip install jupyter (tensorflow1) C:\> pip install matplotlib (tensorflow1) C:\> pip install pandas (tensorflow1) C:\> pip install opencv-python now uninstall numpy: (tensorflow1) C:\> pip uninstall numpy (tensorflow1) C:\> pip install numpy==1.16.4 //the 1.16.4 v is working (tensorflow1) C:\> set PYTHONPATH=C:\tensorflow1\models;C:\tensorflow1\models esearch;C:\tensorflow1\models esearch\slim (tensorflow1) C:\> cd C:\tensorflow1\models esearch NOW very important: open a file manager and navigate to tensorflow1/models/research/object_detection and search for the "protos" folder. Inside this folder you'll see a bunch of .proto files -> you need to make a .pb2 file instance for every single .proto file. DO NOT USE the long command from the written tutorial (protoc --python_out=. .\object_detection\protos\anchor_generator.proto .\object_detection\protos\argmax_matcher.proto......) Instead, look up for the first .proto file u see inside this folder and write the next command: (tensorflow1) C:\>protoc --python_out=. .\object_detection\protos\file_name.proto Repeat this command for every file you see inside the folder untill you make sure that every single .proto file has a .pb2 copy (it takes some time but it's the way to go) Next: (tensorflow1) C:\tensorflow1\models esearch> python setup.py build (tensorflow1) C:\tensorflow1\models esearch> python setup.py install BEFORE RUNNING THE JUPYTER COMMAND download the models file for version 1.7 from github -> open the archieve and search for /research/object_detection/object_detection_tutorial.ipynb -> extract that specific file into YOUR ACTUAL MODEL and replace it's own object_detection_tutorial.ipynb (otherwise it will mess your tensorflow version) Also, I'm not sure if this whole jupyter step is necessary since I'm still having an import error and the images are not showing, but it's still working out. Next: (tensorflow1) C:\tensorflow1\models esearch\object_detection> python xml_to_csv.py Then, generate the TFRecord files by issuing these commands from the \object_detection folder: python generate_tfrecord.py --csv_input=images\train_labels.csv --image_dir=images\train --output_path=train.record python generate_tfrecord.py --csv_input=images\test_labels.csv --image_dir=images\test --output_path=test.record Do the labelmap thing just as in the tutorial, it's nothing to explain here !! Make sure to modify those specific lines inside the config file just as in the written tutorial And now you can run the training : python train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/faster_rcnn_inception_v2_pets.config Hope it works ! Let me know if it is woking for any of you ! Have a good day!

what version of tensorflow-gpu was used for this tutorial?

How do you cropped the image of the object being detected and show it with another frame?

from object_detection.protos import calibration_pb2 as object__detection_dot_protos_dot_calibration__pb2 ImportError: cannot import name 'calibration_pb2' from 'object_detection.protos' (C:\Anaconda3\lib\site-packages\object_detection-0.1-py3.7.egg\object_detection\protos\__init__.py) i have this error is there any solution to clear this out

Hi Edje, Thankyou so much for such a beautiful tutorial. I am using Tensorflow 2.0 and was successful in creating the record file, but when trying this "python train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/faster_rcnn_inception_v2_pets.config" I am encountering with an issue "File "train.py", line 49, in from tensorflow.contrib import framework as contrib_framework ImportError: No module named 'tensorflow.contrib'" Also tried to downgrade the Tensorflow version, but still not working. Any feedback would be greatly appreciated.

I am sorry to ask, is there any way to find out and extract the feature extraction part of this algorithm?

i also dont have the object detection tutorial file in the directory you say and python setup.py build says the file cannot be found. because its not in there. Is this tutorial so old or what is happening here!