Can anybody suggest some ways to use Detectron2 library, as in this competition we need to submit notebooks with Internet off. I am unable to find any way to use those models or ensemble of those. If anyone knows please suggest, Thanks in Advance π
Can anybody suggest some ways to use Detectron2 library, as in this competition we need to submit notebooks with Internet off. I am unable to find any way to use those models or ensemble of those. If anyone knows please suggest, Thanks in Advance π
Hi, Vivek! I am not using Detectron2 but I am trying to use MMDetection without the Internet. In my setting, I want to train models in local environment, and upload the checkpoint for inference. So, the notebook will be no Internet connecting but with some extra self-upload datasets.
\n\nFirst, I followed this blog to install python dependences. In this step, I installed the whl files to my computer and uploaded these files as a new dataset. Second, I uploaded cocoapi and mmdetection libraries as same as the first step. Third, I add these datasets from step 1 and step 2 to the inference notebooks and complied for the libraries. Fourth, I also uploaded some scripts which help to integrated the Global Wheat Detection dataset in the MMDetection framework. Finally, I uploaded the pertained model.
There is something I should mention. First, I copy the cocoapi and mmdetection to /kaggle/working diretory for installing because the input directory is read-only, and I need to make some changes in the framework.\nSecond, there are still many file operations in order to make these step go will, but I think you can figure it out. Third, in this notebook, I should compile the libraries every time I want to get the final output, so it is time-consuming. Finally, I ran the inference correctly, and get the right output. But there is no commit button to competition, and I am still struggling!!!
This is my first time to kaggle competition, so if I got something wrong, PLEASE correct me as soon as possible.
", "rawMarkdown": "Hi, Vivek! I am not using Detectron2 but I am trying to use MMDetection without the Internet. In my setting, I want to train models in local environment, and upload the checkpoint for inference. So, the notebook will be no Internet connecting but with some extra self-upload datasets.\n\nFirst, I followed this [blog](https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/113195) to install python dependences. In this step, I installed the `whl` files to my computer and uploaded these files as a new dataset. Second, I uploaded `cocoapi` and `mmdetection` libraries as same as the first step. Third, I add these datasets from step 1 and step 2 to the inference notebooks and complied for the libraries. Fourth, I also uploaded some scripts which help to integrated the `Global Wheat Detection` dataset in the MMDetection framework. Finally, I uploaded the pertained model.\n\nThere is something I should mention. First, I copy the `cocoapi` and `mmdetection` to `/kaggle/working` diretory for installing because the input directory is read-only, and I need to make some changes in the framework.\nSecond, there are still many file operations in order to make these step go will, but I think you can figure it out. Third, in this notebook, I should compile the libraries every time I want to get the final output, so it is time-consuming. Finally, I ran the inference correctly, and get the right output. But there is no commit button to competition, and I am still struggling!!!\n\nThis is my first time to kaggle competition, so if I got something wrong, PLEASE correct me as soon as possible.", "votes": null }, { "id": "849795", "postDate": "05/16/2020 04:34:32", "content": "@superkevingit Thank you so much for this detailed answer π. I think that I must try this myself. If everything goes fine I will help you soon.
", "rawMarkdown": "superkevingit Thank you so much for this detailed answer π. I think that I must try this myself. If everything goes fine I will help you soon.", "votes": null }, { "id": "849876", "postDate": "05/16/2020 06:22:05", "content": "YesοΌ I think mmdtection is very useful here.
", "rawMarkdown": "YesοΌ I think mmdtection is very useful here.", "votes": null }, { "id": "850330", "postDate": "05/16/2020 14:24:08", "content": "@imvivek14 @xujingzhao Thanks and If I solve the problems I will let you know.π
", "rawMarkdown": "imvivek14 @xujingzhao Thanks and If I solve the problems I will let you know.π", "votes": null }, { "id": "851356", "postDate": "05/17/2020 15:02:02", "content": "@superkevingit I also use mmdetection. I want to ask how to compile mmdet in the notebook. I have followed your method to copy the mmdetection to /kaggle/working diretory, but it did not compile successfully. Thanks.
", "rawMarkdown": "superkevingit I also use mmdetection. I want to ask how to compile mmdet in the notebook. I have followed your method to copy the mmdetection to /kaggle/working diretory, but it did not compile successfully. Thanks.", "votes": null }, { "id": "851944", "postDate": "05/18/2020 03:47:51", "content": "@koosung Hi, Ko! I open source the code and dataset in this notebook and point out the problem I met in this discussion.
", "rawMarkdown": "koosung Hi, Ko! I open source the code and dataset in this [notebook](https://www.kaggle.com/superkevingit/faster-rcnn-with-mmdetection-without-internet) and point out the problem I met in this [discussion](https://www.kaggle.com/c/global-wheat-detection/discussion/152010).", "votes": null }, { "id": "854508", "postDate": "05/20/2020 04:17:30", "content": "@imvivek14 @xujingzhao After I added one line 'rm -rf mmdetection' to move the useless output, and just kept the 'submission.csv' file, I found the 'commit to competition' button finally! But I only got 0.4790 on the leaderboard, maybe something wrong with my model.π¬
", "rawMarkdown": "imvivek14 @xujingzhao After I added one line 'rm -rf mmdetection' to move the useless output, and just kept the 'submission.csv' file, I found the 'commit to competition' button finally! But I only got 0.4790 on the leaderboard, maybe something wrong with my model.π¬", "votes": null }, { "id": "854812", "postDate": "05/20/2020 10:58:59", "content": "@superkevingit which model did you use ? FasterRcnn ?
", "rawMarkdown": "superkevingit which model did you use ? FasterRcnn ?", "votes": null }, { "id": "855005", "postDate": "05/20/2020 13:58:28", "content": "@xujingzhao Yes! I used pretrained model faster_rcnn_r50_fpn_1x_coco and use the Global Wheat Data to fine-tune for 50 epoch. Maybe the hyperparameters are not suitable.
Hi, Vivek! I am not using Detectron2 but I am trying to use MMDetection without the Internet. In my setting, I want to train models in local environment, and upload the checkpoint for inference. So, the notebook will be no Internet connecting but with some extra self-upload datasets.
\n\nFirst, I followed this blog to install python dependences. In this step, I installed the whl files to my computer and uploaded these files as a new dataset. Second, I uploaded cocoapi and mmdetection libraries as same as the first step. Third, I add these datasets from step 1 and step 2 to the inference notebooks and complied for the libraries. Fourth, I also uploaded some scripts which help to integrated the Global Wheat Detection dataset in the MMDetection framework. Finally, I uploaded the pertained model.
There is something I should mention. First, I copy the cocoapi and mmdetection to /kaggle/working diretory for installing because the input directory is read-only, and I need to make some changes in the framework.\nSecond, there are still many file operations in order to make these step go will, but I think you can figure it out. Third, in this notebook, I should compile the libraries every time I want to get the final output, so it is time-consuming. Finally, I ran the inference correctly, and get the right output. But there is no commit button to competition, and I am still struggling!!!
This is my first time to kaggle competition, so if I got something wrong, PLEASE correct me as soon as possible.
", "votes": null, "replies": [ { "id": 849795, "author_name": "imvivek14", "author_url": "", "post_date": "05/16/2020 04:34:32", "content": "@superkevingit Thank you so much for this detailed answer π. I think that I must try this myself. If everything goes fine I will help you soon.
", "votes": null, "replies": [] }, { "id": 849876, "author_name": "xujingzhao", "author_url": "", "post_date": "05/16/2020 06:22:05", "content": "YesοΌ I think mmdtection is very useful here.
", "votes": null, "replies": [] }, { "id": 850330, "author_name": "superkevingit", "author_url": "", "post_date": "05/16/2020 14:24:08", "content": "@imvivek14 @xujingzhao Thanks and If I solve the problems I will let you know.π
", "votes": null, "replies": [] }, { "id": 851356, "author_name": "koosung", "author_url": "", "post_date": "05/17/2020 15:02:02", "content": "@superkevingit I also use mmdetection. I want to ask how to compile mmdet in the notebook. I have followed your method to copy the mmdetection to /kaggle/working diretory, but it did not compile successfully. Thanks.
", "votes": null, "replies": [] }, { "id": 851944, "author_name": "superkevingit", "author_url": "", "post_date": "05/18/2020 03:47:51", "content": "@koosung Hi, Ko! I open source the code and dataset in this notebook and point out the problem I met in this discussion.
", "votes": null, "replies": [] }, { "id": 854508, "author_name": "superkevingit", "author_url": "", "post_date": "05/20/2020 04:17:30", "content": "@imvivek14 @xujingzhao After I added one line 'rm -rf mmdetection' to move the useless output, and just kept the 'submission.csv' file, I found the 'commit to competition' button finally! But I only got 0.4790 on the leaderboard, maybe something wrong with my model.π¬
", "votes": null, "replies": [] }, { "id": 854812, "author_name": "xujingzhao", "author_url": "", "post_date": "05/20/2020 10:58:59", "content": "@superkevingit which model did you use ? FasterRcnn ?
", "votes": null, "replies": [] }, { "id": 855005, "author_name": "superkevingit", "author_url": "", "post_date": "05/20/2020 13:58:28", "content": "@xujingzhao Yes! I used pretrained model faster_rcnn_r50_fpn_1x_coco and use the Global Wheat Data to fine-tune for 50 epoch. Maybe the hyperparameters are not suitable.