Thank you so much for explanation. If only based on detr GitHub, it is very difficult to apply to my own datasets. YOLOv5 v8 are more user friendly because they provide more details how to use it. Without your explanation, it is extremely difficult to use detr.
We have applied this DeTR method, following exactly the same in this video. It works perfectly fine for single-class Face detection detection & it also worked well for our own single-class object detection using Transformer. But when we extend this DeTR code using this tutorial, it didn't work well, I sought the tutorial helps for defining num_class (as num_classes+1) in detr.py file. But even then it didn't work well. If someone could guide us how to apply DeTR for multi-class object detection or if could please refer some DeTR tutorial for multi-class object detection. Thanx!
I'm getting an AssertionError when I run main.py. The output looked like this: File "F:\DETR\util\box_ops.py", line 51, in generalized_box_iou assert (boxes1[:, 2:] >= boxes1[:, :2]).all() AssertionError
I also have a lot of negative images training data, images without labels. In yolo, I can ignore it. In detr, is the same procedure. In coco label file, I only include the images with labels. But in training, I still feed all images into detr. Is my understanding correct.
i trained model with my own dataset. I’m using the DETR model for inference, but the probas values for the predicted bounding boxes are extremely low.The predicted probas values are so low that none of the bounding boxes are selected (they all fall below the threshold). Could this be due to inadequate model training? What adjustments would you recommend to improve the probas values, either in the model itself or during training? Even when I lower the threshold significantly, I still can’t get meaningful results.
1. You just need to add number of categories that you want to handle when creating the dataset 2. Increase the number of class value in detr.py (16:00)
@@deepreader8087 We have applied this DeTR method, following exactly the same in this video. It works perfectly fine for single-class Face detection detection & it also worked well for our own single-class object detection using Transformer. But when we extend this DeTR code using this tutorial, it didn't work well, I sought the tutorial helps for defining num_class (as num_classes+1) in detr.py file. But even then it didn't work well. If someone could guide us how to apply DeTR for multi-class object detection or if could please refer some DeTR tutorial for multi-class object detection. Thanx!
I have a question, in the DeTR code, the criterion is set to train mode, but there are no mention to its parameters in the optimizer. Why set it to train and does the loss even have learnable parameters? Thanks
Thank you for your awesome tutorial ! I tried on my own but I have a problem, during the learning everything is fine. But during the testing process I have a problem, my " output " from the model is giving me multiple "'pred_boxes" and "'pred_logits". Do you have any idea where it come from ?
@@sangthanhnguyen1459 Perhaps you could train under smaller image resolution with multi-scale training(already in the DETR github) and harder random crop augmentation? And during test time inference you could try to fit a larger (say 5000 x 5000) image..
python3 main.py --dataset_file face --data_path ../dataset/ --output_dir output File "main.py", line 204 checkpoint_paths.append(output_dir / f'checkpoint{epoch:04}.pth') ^ Respted Sir! thanks for uploading this valuable video.I tried too many times and have a problem here.Can you please look it.
This is working absolutely perfect on my own dataset. Thank you soo much. 💕
Thank you so much for explanation. If only based on detr GitHub, it is very difficult to apply to my own datasets. YOLOv5 v8 are more user friendly because they provide more details how to use it. Without your explanation, it is extremely difficult to use detr.
Thanks a lot! Please make similar video for other custom datasets
Really appreciate it :)) hope to experience your tutorials continuously for other famous models ~~
Thanks a lot. It's will be turning point for my level up.
Thanks for the Tutorial, Keep on to this type of coding tutorials!
Excellent work ! You helped me a lot. Thank you for your effort !!
Thanks for this, as a beginner this helps a lot.
We have applied this DeTR method, following exactly the same in this video. It works perfectly fine for single-class Face detection detection & it also worked well for our own single-class object detection using Transformer. But when we extend this DeTR code using this tutorial, it didn't work well, I sought the tutorial helps for defining num_class (as num_classes+1) in detr.py file. But even then it didn't work well. If someone could guide us how to apply DeTR for multi-class object detection or if could please refer some DeTR tutorial for multi-class object detection. Thanx!
Can you help with multiple classes ? I am working on plant_leaf_diseases dataset.
what is your labelling method?
i am also searching for that, what is the labelling method @Tony Shin???
How is it possible to save (text file) proba and label for each bounding box of the test phase ?
I'm getting an AssertionError when I run main.py. The output looked like this:
File "F:\DETR\util\box_ops.py", line 51, in generalized_box_iou
assert (boxes1[:, 2:] >= boxes1[:, :2]).all()
AssertionError
Do you use num queries is equal to 100?
Thanks a lot, very helpful for me.
I also have a lot of negative images training data, images without labels. In yolo, I can ignore it. In detr, is the same procedure. In coco label file, I only include the images with labels. But in training, I still feed all images into detr. Is my understanding correct.
Hey, It is not showing my test ig. how to visualize test image
i trained model with my own dataset. I’m using the DETR model for inference, but the probas values for the predicted bounding boxes are extremely low.The predicted probas values are so low that none of the bounding boxes are selected (they all fall below the threshold). Could this be due to inadequate model training? What adjustments would you recommend to improve the probas values, either in the model itself or during training? Even when I lower the threshold significantly, I still can’t get meaningful results.
how to train for multiple classes
1. You just need to add number of categories that you want to handle when creating the dataset
2. Increase the number of class value in detr.py (16:00)
@@deepreader8087 We have applied this DeTR method, following exactly the same in this video. It works perfectly fine for single-class Face detection detection & it also worked well for our own single-class object detection using Transformer. But when we extend this DeTR code using this tutorial, it didn't work well, I sought the tutorial helps for defining num_class (as num_classes+1) in detr.py file. But even then it didn't work well. If someone could guide us how to apply DeTR for multi-class object detection or if could please refer some DeTR tutorial for multi-class object detection. Thanx!
Could you do a video why delete these parameters, it will work.
I have a question, in the DeTR code, the criterion is set to train mode, but there are no mention to its parameters in the optimizer. Why set it to train and does the loss even have learnable parameters?
Thanks
Thanks for the tutorial. but how could one evaluate the performance? like, mAp/Ap, Recall?
jonathan-hui.medium.com/map-mean-average-precision-for-object-detection-45c121a31173
There's a good blogpost about mAP score :)
Kindly Guide me How to Train For Multiple Classes
Thanks a lot. very educative. Is it possible to do another video for End-to-End Instance Segmentation with Transformers (ISTR) in the near future?
Are the used images raw or paired with a segmentation file?
They're all raw images :)
@@deepreader8087 I mean are there any annotations included in the training? :)
Thank you :)
Hey great video! I was wondering if you can override PIL's image load somehow or even better, replace repo's dataloaders with your own.
Late reply but some GIS packages allow for a lot more easily opening raster formats.
your video helps me a lot 👍
Thank you for your awesome tutorial ! I tried on my own but I have a problem, during the learning everything is fine. But during the testing process I have a problem, my " output " from the model is giving me multiple "'pred_boxes" and "'pred_logits". Do you have any idea where it come from ?
I appreciate your sharing
Thank for great video :D
This model can be trained with large image, about 5000 pixels x 5000 pixels ?
Using 5000 x 5000 image for training will cause out of memory error.
I suggest using a smaller resolution :)
@@deepreader8087 Yeah I also think so, but datas which I am having in that size. Any recommend for this problem? Many thanks :D
@@sangthanhnguyen1459 Perhaps you could train under smaller image resolution with multi-scale training(already in the DETR github) and harder random crop augmentation?
And during test time inference you could try to fit a larger (say 5000 x 5000) image..
Thank you for your hard work.
python3 main.py --dataset_file face --data_path ../dataset/ --output_dir output
File "main.py", line 204
checkpoint_paths.append(output_dir / f'checkpoint{epoch:04}.pth')
^
Respted Sir! thanks for uploading this valuable video.I tried too many times and have a problem here.Can you please look it.
Is it possible for you to give me the full error log?It'd be nice if you open this as an issue in the github repo I shared
@@deepreader8087 Sorry for late reply i solved this problem
Thank you
this if fkg video 🤔
extremely poor audio....very disappointing