How to Train a Custom Faster R-CNN Model with Facebook AI's Detectron2 | Use Your Own Dataset

[@Baddogbling321]

HI, the colab notebooks are very outdated to the point nothing works. PLEASE UPDATE THE COLAB NOTEBOOKS ROBOFLOW! Thanks.

[@Mee-yp5in]

A very fine platform for preparing dataset and training model, recommended!!

[@athlete645]

When the model was trained, shouldn't have the cfg.MODEL.ROID_HEADS.NUM_CLASSES been changed? The comment on that line says "your number of classes +1". Thank you for the great explanation! This was exactly the video I needed.

[@shoebbashar4385]

Please update the colab notebook to tensorflow latest version.

[@cheerio4742]

error when installing torch and torchvision in their version "Could not find a version that satisfies the requirement torch==1.5"

[@GaJolGuru]

Hi Roboflow Thanks for this great video, can you perahps help me out here. I used the workflow and updated both torch and torchvision as these old packages are not available anymore. However detectron2 seems not to be found anywhere, any ideas? Thanks

[@__-dr2if]

the code uses tensorflow_gpu 1.5 but now google colab supports 2.8.0 or above and so not able t download API

[@alhasanalshaebi]

How does one choose an optimal number of epochs?especially in this case of Dataset.

[@Fatima-kj9ws]

Great, very well explained tutorial. Thank you very much

[@Lifewithballi]

i have question are you free ???

[@Paperwhite786]

This is a Roboflow ad

[@omerkaya5669]

I have several questions? How can I plot PR curve and ROC curve? I also want to reach the image of confusion matrix. Where can I find the F score value and mAP values?

[@Roboflow]

Unfortunately, those metrics don't come out of the box with Detectron2. However, you can calculate them using other packages, for example: github.com/SkalskiP/onemetric

[@omerkaya5669]

@@Roboflow Thank you for answer. Roboflow is the best learning page

[@omerkaya5669]

@@Roboflow confusion_matrix = ConfusionMatrix.from_detections( true_batches=true_batches, detection_batches=detection_batches, num_classes=1 ) I run this code. but i take this error. NameError: name 'ConfusionMatrix' is not defined

[@tanzimmostafa]

Hello, can someone please tell me how to find out the mAP metric. The notebook only shows the AP for each object class.

[@drswapnil1234]

Hi. I need to detect circles in a custom model. I have annonated the training images using labelme. It gives 2 no pairs in the json. Can you pls help in how to modify the code for circle detection. Its giving an error as not polygon with current code. Thanks

[@akshayakki5692]

Sir, is it necessary to annotate test images also or why we should have annotated labels for test images those images are for only testing right.

[@danaali1710]

Hi, thank you please answer my question. if my project consists of detecting one class of object , should i annotate my images and create bounding box.

[@anirudhputsala8276]

Hi, After generating the dataset, the train folder neither show any preview of the images nor has a folder of it's name in the download. Is there any limit of images I could train without pro version?

[@minsoojeong199]

Hello, thank you for the good lecture. I'd like to know how many photos I can upload to roboflow.

[@rohanrane7835]

Hi, it is showing me import error when executing "from detectron2 import model_zoo". How do I fix this?

[@melvin1919]

When i run , the class names are not showed up? any reason why?

[@_ABDULGHANI]

Hello there and thank you, if I want to train for longe time is this method support continue from the last checkpoint.

[@pavan3078]

anyone knows what is happening here ERROR: Could not find a version that satisfies the requirement torch==1.5 (from versions: 1.11.0, 1.12.0, 1.12.1, 1.13.0, 1.13.1, 2.0.0, 2.0.1, 2.1.0) ERROR: No matching distribution found for torch==1.5

[@g666fan]

Yeah same. Did you already fix it?

[@vigneshbalaji6826]

How to use multi-gpu in detectron2 for custom data training

[@tutsy98]

THANK U ROBOFLOWWWWW

[@zubairsk1624]

can any one please tell me how to train it on my local GPU instead of colab notebook

[@altamontholmes8719]

Hi, how can I visualize the feature maps? Btw thanks for the good and brief tutorial!

[@himomhey3437]

Good Tutorial! How can I deploy the model.pth in a raspberry pi 4

[@JOHN-vb5bh]

Can anyone give suggestions or instructions on integrating this code with detectron2 official code. I had done from detectron2.engine import MyTrainer from detectron2.engine import LossEvalHook and trainer = MyTrainer(cfg) in my training scripts But getting an error TypeError Traceback (most recent call last) in 21 22 os.makedirs(cfg.OUTPUT_DIR, exist_ok=True) ---> 23 trainer = MyTrainer(cfg) 24 trainer.resume_or_load(resume=False) 25 trainer.train() TypeError: 'module' object is not callable Please help me to solve this problem

[@neelkanthbhavnagarwala6001]

SizeMismatchError when trying to train a model on custom dataset

[@hailuu625]

Is there anyone got error KeyError: 'instances'

[@janeng91]

Did you trained this model? How long for the training time?

[@neelkanthbhavnagarwala6001]

sizeMismatchError when trying to train a model using custom dataset..

[@darkamax1]

Hello, Thank you for the amazingly well explained tutorial. I wrote a code and am using detectron2 to identify objects. It is working great with my in built webcam but whenever I try and use my webcam, it displays only the first frame and then says ..."ASSERT: "false" in file qasciikey.cpp, line 501" My Hardware config is as follows Ubuntu 18.04 Nvidia GeForce GTX 1650Ti Card Torch Version '1.8.1+cu102' Cuda 10.2 Python 3.6.9 Any help will be much appreciated. Thanks in advance.

[@darkamax1]

Also my code is unable to read any key inputs when it is running, when attempting to quit the code with if ord=='q': cv2.destroyAllWindows() it either hangs or dumps the code with error Segmentation fault, core dumped Thanks

[@DiepHuongTran]

hello sir. cfg.MODEL.ROI_HEADS.NUM_CLASSES = 4 # your number of classes + 1. If I train 6 classes, set it to 7 right?

[@vigneshbalaji6826]

How to use multi-gpu in detectron2 for custom data training