![\"\"](\"https://raw.githubusercontent.com/ZFTurbo/Keras-Mask-RCNN-for-Open-Images-2019-Instance-Segmentation/master/img/mask_rcnn_prediction_example.jpg\")
I made a release of my Mask-RCNN code for Instance Segmentation task:
\n\nhttps://github.com/ZFTurbo/Keras-Mask-RCNN-for-Open-Images-2019-Instance-Segmentation
\n\nIt wasn’t used in our best submit because I finished training too late. But it gives good result on LB, so I decided to release it.
\n\n
Repository contains:
\n\nI was able to get 0.4670 Public LB and 0.4311 Private LB using ensemble of these 3 models.
\n\nThere are 3 graphics with training process for each model. I changed parameters of training over some epochs (Freeze Backbone, ReduceLR, AdamAccumulate, different image sampling strategy). Fast growth of score are in places where I switched to uniform classes sampling. Also to speed up the ResNet101 training, I copied weights for decoder from trained Mask-RCNN with ResNet152 backbone. So ResNet101 got good score for much lower number of epochs.
\n\n![\"\"](\"https://storage.googleapis.com/kaggle-forum-message-attachments/661345/14368/ResNet50.png\")
![\"\"](\"https://storage.googleapis.com/kaggle-forum-message-attachments/661345/14369/ResNet101.png\")
![\"\"](\"https://storage.googleapis.com/kaggle-forum-message-attachments/661345/14370/ResNet152.png\")
I made a release of my Mask-RCNN code for Instance Segmentation task:
\n\nhttps://github.com/ZFTurbo/Keras-Mask-RCNN-for-Open-Images-2019-Instance-Segmentation
\n\nIt wasn’t used in our best submit because I finished training too late. But it gives good result on LB, so I decided to release it.
\n\n
Repository contains:
\n\nI was able to get 0.4670 Public LB and 0.4311 Private LB using ensemble of these 3 models.
\n\nThere are 3 graphics with training process for each model. I changed parameters of training over some epochs (Freeze Backbone, ReduceLR, AdamAccumulate, different image sampling strategy). Fast growth of score are in places where I switched to uniform classes sampling. Also to speed up the ResNet101 training, I copied weights for decoder from trained Mask-RCNN with ResNet152 backbone. So ResNet101 got good score for much lower number of epochs.
\n\n
@zfturbo , just wanted to know , what are the training size for the model (github).
", "rawMarkdown": "@zfturbo , just wanted to know , what are the training size for the model ([github](https://github.com/ZFTurbo/Keras-Mask-RCNN-for-Open-Images-2019-Instance-Segmentation)). ", "replies": [ { "id": 1069371, "postDate": "2020-11-04T11:13:28.707Z", "content": "What do you mean by \"training size\"?
", "rawMarkdown": "What do you mean by \"training size\"?" }, { "id": 1069393, "postDate": "2020-11-04T11:53:04.803Z", "content": "on how many data you have trained
", "rawMarkdown": "on how many data you have trained" }, { "id": 1069399, "postDate": "2020-11-04T12:05:07.713Z", "content": "All data from Open Images Dataset which were available.
", "rawMarkdown": "All data from Open Images Dataset which were available." } ] }, { "id": 1069200, "postDate": "2020-11-04T07:37:54.567Z", "content": "@zfturbo could you suggest version of tensorflow and keras for running pretained models?
", "rawMarkdown": "@zfturbo could you suggest version of tensorflow and keras for running pretained models?", "replies": [ { "id": 1069274, "postDate": "2020-11-04T09:24:29.190Z", "content": "It must work with [keras 2.3 + tensorflow 1.15]
", "rawMarkdown": "It must work with [keras 2.3 + tensorflow 1.15]" } ] }, { "id": 663360, "postDate": "2019-11-01T21:35:59.623Z", "content": "good!
", "rawMarkdown": "good!" }, { "id": 661834, "postDate": "2019-10-30T18:50:15.440Z", "content": "great work!!!
", "rawMarkdown": "great work!!!" }, { "id": 665185, "postDate": "2019-11-04T17:54:03.243Z", "rawMarkdown": "", "isDeleted": true }, { "id": 663692, "postDate": "2019-11-02T14:39:40.570Z", "content": "Thanks @zfturbo !!
", "rawMarkdown": "Thanks @zfturbo !!" }, { "id": 662934, "postDate": "2019-11-01T08:35:46.620Z", "content": "thanks!
", "rawMarkdown": "thanks!" }, { "id": 662283, "postDate": "2019-10-31T11:25:29.427Z", "content": "Awesome, thanks a lot!
", "rawMarkdown": "Awesome, thanks a lot! " } ], "comments": [ { "id": 1069368, "author_name": "Randheer kumar", "author_url": "", "post_date": "2020-11-04T11:08:29.220000", "content": "@zfturbo , just wanted to know , what are the training size for the model (github).
", "votes": 0, "replies": [ { "id": 1069371, "author_name": "ZFTurbo", "author_url": "", "post_date": "2020-11-04T11:13:28.707000", "content": "What do you mean by \"training size\"?
", "votes": 0, "replies": [] }, { "id": 1069393, "author_name": "Randheer kumar", "author_url": "", "post_date": "2020-11-04T11:53:04.803000", "content": "on how many data you have trained
", "votes": 0, "replies": [] }, { "id": 1069399, "author_name": "ZFTurbo", "author_url": "", "post_date": "2020-11-04T12:05:07.713000", "content": "All data from Open Images Dataset which were available.
", "votes": 0, "replies": [] } ] }, { "id": 1069200, "author_name": "Randheer kumar", "author_url": "", "post_date": "2020-11-04T07:37:54.567000", "content": "@zfturbo could you suggest version of tensorflow and keras for running pretained models?
", "votes": 0, "replies": [ { "id": 1069274, "author_name": "ZFTurbo", "author_url": "", "post_date": "2020-11-04T09:24:29.190000", "content": "It must work with [keras 2.3 + tensorflow 1.15]
", "votes": 0, "replies": [] } ] }, { "id": 663360, "author_name": "Leandro Amoras", "author_url": "", "post_date": "2019-11-01T21:35:59.623000", "content": "good!
", "votes": 0, "replies": [] }, { "id": 661834, "author_name": "Krishna Katyal", "author_url": "", "post_date": "2019-10-30T18:50:15.440000", "content": "great work!!!
", "votes": 0, "replies": [] }, { "id": 665185, "author_name": "", "author_url": "", "post_date": "2019-11-04T17:54:03.243000", "content": "", "votes": 0, "replies": [] }, { "id": 663692, "author_name": "ShankarKurni", "author_url": "", "post_date": "2019-11-02T14:39:40.570000", "content": "Thanks @zfturbo !!
", "votes": 0, "replies": [] }, { "id": 662934, "author_name": "欧阳逸云", "author_url": "", "post_date": "2019-11-01T08:35:46.620000", "content": "thanks!
", "votes": 0, "replies": [] }, { "id": 662283, "author_name": "Daniel Haller", "author_url": "", "post_date": "2019-10-31T11:25:29.427000", "content": "Awesome, thanks a lot!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "661345": "I made a release of my Mask-RCNN code for Instance Segmentation task:\n\nhttps://github.com/ZFTurbo/Keras-Mask-RCNN-for-Open-Images-2019-Instance-Segmentation\n\nIt wasn’t used in our best submit because I finished training too late. But it gives good result on LB, so I decided to release it. \n\n\n\nRepository contains:\n\n- Pre-trained Mask R-CNN models (ResNet50, ResNet101 and ResNet152 backbones)\n- Example code to get predictions with these models for any set of images\n- Code to train (continue training) model based on Keras Mask R-CNN and OID dataset\n\nI was able to get **0.4670** Public LB and **0.4311** Private LB using ensemble of these 3 models.\n\nThere are 3 graphics with training process for each model. I changed parameters of training over some epochs (Freeze Backbone, ReduceLR, AdamAccumulate, different image sampling strategy). Fast growth of score are in places where I switched to uniform classes sampling. Also to speed up the ResNet101 training, I copied weights for decoder from trained Mask-RCNN with ResNet152 backbone. So ResNet101 got good score for much lower number of epochs.\n\n\n\n\n\n", "1069368": "@zfturbo , just wanted to know , what are the training size for the model ([github](https://github.com/ZFTurbo/Keras-Mask-RCNN-for-Open-Images-2019-Instance-Segmentation)). ", "1069200": "@zfturbo could you suggest version of tensorflow and keras for running pretained models?", "663360": "good!", "661834": "great work!!!", "665185": "", "663692": "Thanks @zfturbo !!", "662934": "thanks!", "662283": "Awesome, thanks a lot! " } }