Hi all, I have converted the labels to the YOLO format(class x y width height) so feel free to use it. It doesn't contain labels for the images without any annotations.\nHere is an implementation that I was testing out: https://github.com/eriklindernoren/PyTorch-YOLOv3\nSadly, I couldn't get very good results(only tried yolov3 tiny due to my local machine's restrictions) but I hope it can help others!\nEDIT:\nCreation kernel is here: https://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels
", "messages": [ { "id": "585085", "postDate": "07/26/2019 23:30:35", "content": "Hi all, I have converted the labels to the YOLO format(class x y width height) so feel free to use it. It doesn't contain labels for the images without any annotations.\nHere is an implementation that I was testing out: https://github.com/eriklindernoren/PyTorch-YOLOv3\nSadly, I couldn't get very good results(only tried yolov3 tiny due to my local machine's restrictions) but I hope it can help others!\nEDIT:\nCreation kernel is here: https://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels
", "rawMarkdown": "Hi all, I have converted the labels to the YOLO format(class x y width height) so feel free to use it. It doesn't contain labels for the images without any annotations.\nHere is an implementation that I was testing out: https://github.com/eriklindernoren/PyTorch-YOLOv3\nSadly, I couldn't get very good results(only tried yolov3 tiny due to my local machine's restrictions) but I hope it can help others!\nEDIT:\nCreation kernel is here: https://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels", "votes": null }, { "id": "585438", "postDate": "07/27/2019 13:19:17", "content": "Nice! I've got a converter up for converting to COCO format as well that I'll be sharing once I've verified that everything works the way I'd like
", "rawMarkdown": "Nice! I've got a converter up for converting to COCO format as well that I'll be sharing once I've verified that everything works the way I'd like", "votes": null }, { "id": "586374", "postDate": "07/29/2019 04:41:00", "content": "I tried your yolo_train_labels with my classes_names and failed to run yolo training. it seems my target classes_names did not match your labels . : ( Can you share your classes_name file?
", "rawMarkdown": "I tried your yolo_train_labels with my classes_names and failed to run yolo training. it seems my target classes_names did not match your labels . : ( Can you share your classes_name file?", "votes": null }, { "id": "586923", "postDate": "07/29/2019 22:13:07", "content": "Ive attached my names file(its called obj.names)\nEdit, what implementation are you using? I could try it out on my end as well.
", "rawMarkdown": "Ive attached my names file(its called obj.names)\nEdit, what implementation are you using? I could try it out on my end as well.", "votes": null }, { "id": "588028", "postDate": "07/30/2019 03:33:03", "content": "I've just created a kernel that does it here:\nhttps://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels
", "rawMarkdown": "I've just created a kernel that does it here:\nhttps://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels", "votes": null }, { "id": "588131", "postDate": "07/30/2019 07:42:44", "content": "Thanks. I use this pytorch yolov3: https://github.com/eriklindernoren/PyTorch-YOLOv3. I tried both the tiny yolo setup and yolo setup and got the same cuda error:
\nFile \"E:\\kaggle\\kuzushiji\\PyTorch-YOLOv3-master\\utils\\utils.py\", line 317, in build_targets\n class_mask[b, best_n, gj, gi] = (pred_cls[b, best_n, gj, gi].argmax(-1) == target_labels).float()\nRuntimeError: cuda runtime error (59) : device-side assert triggered at C:/w/1/s/tmp_conda_3.6_041836/conda/conda-bld/pytorch_1556684464974/work/aten/src\\THC/THCTensorMathCompareT.cuh:69
--------------update-----------\nI solved this bug. Because I do not use the create_custom_model.sh to create my yolov3-custom.cfg, I used the default yolov3.cfg. Hence the classes was wrong and some conv filters were wrong.
\n\nI modified the train.py and add some codes of nvidia apex , now I can use yolov3 with batch_size=6 for training.
", "rawMarkdown": "Thanks. I use this pytorch yolov3: https://github.com/eriklindernoren/PyTorch-YOLOv3. I tried both the tiny yolo setup and yolo setup and got the same cuda error:Nice, what gpu are you using to be able to get batch_size of 6?
", "rawMarkdown": "Nice, what gpu are you using to be able to get batch_size of 6?", "votes": null }, { "id": "592622", "postDate": "08/05/2019 15:14:04", "content": "It is 1080TI. I trained 20 epochs with yolov3, but the result was very bad. The train loss was over 300. I tried to detect some trained pictures, the detected bboxes were very bad. : ( Maybe I should try other models like CTPN or PixelLink .
", "rawMarkdown": "It is 1080TI. I trained 20 epochs with yolov3, but the result was very bad. The train loss was over 300. I tried to detect some trained pictures, the detected bboxes were very bad. : ( Maybe I should try other models like CTPN or [PixelLink](https://arxiv.org/abs/1801.01315) .", "votes": null } ], "comments": [ { "id": 585438, "author_name": "squidinator", "author_url": "", "post_date": "07/27/2019 13:19:17", "content": "Nice! I've got a converter up for converting to COCO format as well that I'll be sharing once I've verified that everything works the way I'd like
", "votes": null, "replies": [] }, { "id": 586374, "author_name": "qinhui1999", "author_url": "", "post_date": "07/29/2019 04:41:00", "content": "I tried your yolo_train_labels with my classes_names and failed to run yolo training. it seems my target classes_names did not match your labels . : ( Can you share your classes_name file?
", "votes": null, "replies": [ { "id": 586923, "author_name": "sidhanthholalkere", "author_url": "", "post_date": "07/29/2019 22:13:07", "content": "Ive attached my names file(its called obj.names)\nEdit, what implementation are you using? I could try it out on my end as well.
", "votes": null, "replies": [] }, { "id": 588028, "author_name": "sidhanthholalkere", "author_url": "", "post_date": "07/30/2019 03:33:03", "content": "I've just created a kernel that does it here:\nhttps://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels
", "votes": null, "replies": [] }, { "id": 588131, "author_name": "qinhui1999", "author_url": "", "post_date": "07/30/2019 07:42:44", "content": "Thanks. I use this pytorch yolov3: https://github.com/eriklindernoren/PyTorch-YOLOv3. I tried both the tiny yolo setup and yolo setup and got the same cuda error:
\nFile \"E:\\kaggle\\kuzushiji\\PyTorch-YOLOv3-master\\utils\\utils.py\", line 317, in build_targets\n class_mask[b, best_n, gj, gi] = (pred_cls[b, best_n, gj, gi].argmax(-1) == target_labels).float()\nRuntimeError: cuda runtime error (59) : device-side assert triggered at C:/w/1/s/tmp_conda_3.6_041836/conda/conda-bld/pytorch_1556684464974/work/aten/src\\THC/THCTensorMathCompareT.cuh:69
--------------update-----------\nI solved this bug. Because I do not use the create_custom_model.sh to create my yolov3-custom.cfg, I used the default yolov3.cfg. Hence the classes was wrong and some conv filters were wrong.
\n\nI modified the train.py and add some codes of nvidia apex , now I can use yolov3 with batch_size=6 for training.
", "votes": null, "replies": [] }, { "id": 592151, "author_name": "sidhanthholalkere", "author_url": "", "post_date": "08/04/2019 22:03:38", "content": "Nice, what gpu are you using to be able to get batch_size of 6?
", "votes": null, "replies": [] }, { "id": 592622, "author_name": "qinhui1999", "author_url": "", "post_date": "08/05/2019 15:14:04", "content": "It is 1080TI. I trained 20 epochs with yolov3, but the result was very bad. The train loss was over 300. I tried to detect some trained pictures, the detected bboxes were very bad. : ( Maybe I should try other models like CTPN or PixelLink .
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "585085": "Hi all, I have converted the labels to the YOLO format(class x y width height) so feel free to use it. It doesn't contain labels for the images without any annotations.\nHere is an implementation that I was testing out: https://github.com/eriklindernoren/PyTorch-YOLOv3\nSadly, I couldn't get very good results(only tried yolov3 tiny due to my local machine's restrictions) but I hope it can help others!\nEDIT:\nCreation kernel is here: https://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels", "585438": "Nice! I've got a converter up for converting to COCO format as well that I'll be sharing once I've verified that everything works the way I'd like", "586374": "I tried your yolo_train_labels with my classes_names and failed to run yolo training. it seems my target classes_names did not match your labels . : ( Can you share your classes_name file?", "586923": "Ive attached my names file(its called obj.names)\nEdit, what implementation are you using? I could try it out on my end as well.", "588028": "I've just created a kernel that does it here:\nhttps://www.kaggle.com/sidhanthholalkere/how-to-create-yolo-labels", "588131": "Thanks. I use this pytorch yolov3: https://github.com/eriklindernoren/PyTorch-YOLOv3. I tried both the tiny yolo setup and yolo setup and got the same cuda error: