{ "id": 279996, "title": "Mask-RCNN | MMDetection Starter - LB: 0.270", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/279996", "author_name": "Awsaf", "post_date": "2021-10-19T21:03:56.917000", "votes": 66, "comment_count": 16, "views": 0, "content": "

\n

Published some starter notebooks using Mask-RCNN with MMDetection. I hope it helps.

\n

Keypoints:

\n\n

Notebooks:

\n

Mask-RCNN:

\n\n

UNet:

\n\n

WandB

\n

These notebooks come with wandb integration. Hence you can,

\n\n

Examples:

\n

\"results-12-1\"

\n

Predictions:

\n\n
\n

Update: 23/10/2021

\n

Train : v16

\n", "messages": [ { "id": 1550640, "postDate": "2021-10-19T21:03:56.917Z", "content": "

\n

Published some starter notebooks using Mask-RCNN with MMDetection. I hope it helps.

\n

Keypoints:

\n\n

Notebooks:

\n

Mask-RCNN:

\n\n

UNet:

\n\n

WandB

\n

These notebooks come with wandb integration. Hence you can,

\n\n

Examples:

\n

\"results-12-1\"

\n

Predictions:

\n\n
\n

Update: 23/10/2021

\n

Train : v16

\n", "rawMarkdown": "\n\nPublished some starter notebooks using **Mask-RCNN** with **MMDetection**. I hope it helps.\n\n## Keypoints:\n* Currently it get's `LB=0.270`. \n* Model isn't tuned hence one can improve the scores further.\n* It uses simple post-processing to remove the overlap between instances of the same image, check [here](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279995) for more details.\n* Model: Mask-RCNN\n* Backbone: ResNest50\n* Optimizer: SGD\n* Epochs: 12\n* Augmentation: RandomFlip, PhotometricDistortion(Brightness, Contrast, Hue, etc)\n* **Wandb** Integration\n\n## Notebooks:\n\n**Mask-RCNN**:\n* Train: [Sartorius: MMDetection [Train]](https://www.kaggle.com/awsaf49/sartorius-mmdetection-train)\n* Infer: [Sartorius: MMDetection [Infer]](https://www.kaggle.com/awsaf49/sartorius-mmdetection-infer)\n\n**UNet**:\n* Train: [[PyTorch] Sartorius: UNet Strikes Back [Train] πŸ”₯](https://www.kaggle.com/awsaf49/pytorch-sartorius-unet-strikes-back-train/)\n* Infer: [[PyTorch] Sartorius: UNet Strikes Back [Infer] πŸ”₯](https://www.kaggle.com/awsaf49/pytorch-sartorius-unet-strikes-back-infer/)\n\n\n## WandB\nThese notebooks come with **wandb** integration. Hence you can,\n* Track training real-time.\n* Do Exploratory Data Analysis.\n* Debug your Pipeline.\n* Analyze your result.\n\"index\"\n\n## Examples:\n\"results-12-1\"\n\n## Predictions:\n* Good:\n\"results-31-0\"\n\n* Bad:\n\"results-33-0\"\n\n-------------------------\n## Update: 23/10/2021\n**Train** : `v16`\n* Custom Config is added for Hyperparameter Tuning\n* `first cell missing` issue fixed\n* batch_size is increased from `2` to `8`\n", "votes": 65 }, { "id": 1552911, "postDate": "2021-10-21T18:25:39.257Z", "content": "

@awsaf49 Helpful as usual, a great example for novices.

", "rawMarkdown": "@awsaf49 Helpful as usual, a great example for novices.\n", "votes": 1 }, { "id": 1552429, "postDate": "2021-10-21T12:30:53.860Z", "content": "

This will be a nice starting point! Thank you for sharing

", "rawMarkdown": "This will be a nice starting point! Thank you for sharing", "votes": 1 }, { "id": 1552061, "postDate": "2021-10-21T06:02:18.060Z", "content": "

Great work Awsaf! Thanks for sharing!

", "rawMarkdown": "Great work Awsaf! Thanks for sharing!", "votes": 1, "replies": [ { "id": 1552073, "postDate": "2021-10-21T06:09:49.117Z", "content": "

Thanks Chris

", "rawMarkdown": "Thanks Chris", "votes": 1 } ] }, { "id": 1550687, "postDate": "2021-10-19T23:39:09.340Z", "content": "

That was quick! great job!

", "rawMarkdown": "That was quick! great job! ", "votes": 1, "replies": [ { "id": 1550825, "postDate": "2021-10-20T04:04:33.660Z", "content": "

thank you :)

", "rawMarkdown": "thank you :)", "votes": 1 } ] }, { "id": 1551400, "postDate": "2021-10-20T15:25:13.143Z", "content": "

Nice notebook (train+infer) for starter instance segmentation with mmdet @awsaf49

", "rawMarkdown": "Nice notebook (train+infer) for starter instance segmentation with mmdet @awsaf49 ", "votes": 2 }, { "id": 1551363, "postDate": "2021-10-20T14:54:47.233Z", "content": "

@awsaf49 Man, I will just mention the elephant in the room so there are no hard feelings: we spent our Tuesdays working along the same investigation lines and came up with similar work. I just saw all your work today, but I see that the publishing hours might make it look weird.

\n

I had worked all Tuesday to come up with my notebook, and I was just waiting for my submission quota to renew to publish it. And I can tell that you worked hard as well.

\n

I want to state this so it's clear that it was just a shitty coincidence and not something else so we are cool ;)

\n

Anyways, I like your work very much. And you got quite high LB scores! I hope I can get out of the 0.0-ish regime soon lol.

", "rawMarkdown": "@awsaf49 Man, I will just mention the elephant in the room so there are no hard feelings: we spent our Tuesdays working along the same investigation lines and came up with similar work. I just saw all your work today, but I see that the publishing hours might make it look weird. \n\nI had worked all Tuesday to come up with my notebook, and I was just waiting for my submission quota to renew to publish it. And I can tell that you worked hard as well.\n\nI want to state this so it's clear that it was just a shitty coincidence and not something else so we are cool ;)\n\nAnyways, I like your work very much. And you got quite high LB scores! I hope I can get out of the `0.0-ish` regime soon lol.\n", "votes": 2, "replies": [ { "id": 1551384, "postDate": "2021-10-20T15:09:16.600Z", "content": "

It's not that much of a coincedence actually, unet and mask-rcnn are common for segmentation hence most people will go for it. Even if we didn't publish it someone else would've done it …. . I wish you all the best :)

", "rawMarkdown": "It's not that much of a coincedence actually, `unet` and `mask-rcnn` are common for segmentation hence most people will go for it. Even if we didn't publish it someone else would've done it .... . I wish you all the best :)", "votes": 3 } ] }, { "id": 1559113, "postDate": "2021-10-26T18:24:09.873Z", "content": "

Thanks for sharing @awsaf49 , really useful for other folks participating here!

", "rawMarkdown": "Thanks for sharing @awsaf49 , really useful for other folks participating here!" }, { "id": 1558511, "postDate": "2021-10-26T09:44:25.320Z", "content": "

Hello, how do you deal with the overlapped annotation when you create mask in your code? The npz files, to be precise.

", "rawMarkdown": "Hello, how do you deal with the overlapped annotation when you create mask in your code? The npz files, to be precise.", "replies": [ { "id": 1558520, "postDate": "2021-10-26T09:51:45.053Z", "content": "

I'm using public dataset which uses following codes,

\n
for i, annot in enumerate(annotation): \n     img_mask = np.where(rle_decode(annot, (height, width))!=0, i, img_mask)\n
\n

So, I think it takes the first one.

", "rawMarkdown": "I'm using public dataset which uses following codes,\n
\nfor i, annot in enumerate(annotation): \n     img_mask = np.where(rle_decode(annot, (height, width))!=0, i, img_mask)\n
\nSo, I think it takes the first one.\n" }, { "id": 1558685, "postDate": "2021-10-26T12:42:19.107Z", "content": "

I got it! thank you for reply.

", "rawMarkdown": "I got it! thank you for reply. ", "votes": 1 } ] }, { "id": 1554009, "postDate": "2021-10-22T18:36:24.150Z", "content": "

No offense. Are frameworks like mmdetection allowed to be used in the competition?

", "rawMarkdown": "No offense. Are frameworks like mmdetection allowed to be used in the competition?" }, { "id": 1558422, "postDate": "2021-10-26T08:06:57.677Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1552600, "postDate": "2021-10-21T14:42:21.410Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 1552911, "author_name": "Muhammad Maaz", "author_url": "", "post_date": "2021-10-21T18:25:39.257000", "content": "

@awsaf49 Helpful as usual, a great example for novices.

", "votes": 1, "replies": [] }, { "id": 1552429, "author_name": "DanTe", "author_url": "", "post_date": "2021-10-21T12:30:53.860000", "content": "

This will be a nice starting point! Thank you for sharing

", "votes": 1, "replies": [] }, { "id": 1552061, "author_name": "Chris Deotte", "author_url": "", "post_date": "2021-10-21T06:02:18.060000", "content": "

Great work Awsaf! Thanks for sharing!

", "votes": 1, "replies": [ { "id": 1552073, "author_name": "Awsaf", "author_url": "", "post_date": "2021-10-21T06:09:49.117000", "content": "

Thanks Chris

", "votes": 1, "replies": [] } ] }, { "id": 1550687, "author_name": "DrHB", "author_url": "", "post_date": "2021-10-19T23:39:09.340000", "content": "

That was quick! great job!

", "votes": 1, "replies": [ { "id": 1550825, "author_name": "Awsaf", "author_url": "", "post_date": "2021-10-20T04:04:33.660000", "content": "

thank you :)

", "votes": 1, "replies": [] } ] }, { "id": 1551400, "author_name": "KhanhVD", "author_url": "", "post_date": "2021-10-20T15:25:13.143000", "content": "

Nice notebook (train+infer) for starter instance segmentation with mmdet @awsaf49

", "votes": 2, "replies": [] }, { "id": 1551363, "author_name": "JuliΓ‘n Peller (dataista0)", "author_url": "", "post_date": "2021-10-20T14:54:47.233000", "content": "

@awsaf49 Man, I will just mention the elephant in the room so there are no hard feelings: we spent our Tuesdays working along the same investigation lines and came up with similar work. I just saw all your work today, but I see that the publishing hours might make it look weird.

\n

I had worked all Tuesday to come up with my notebook, and I was just waiting for my submission quota to renew to publish it. And I can tell that you worked hard as well.

\n

I want to state this so it's clear that it was just a shitty coincidence and not something else so we are cool ;)

\n

Anyways, I like your work very much. And you got quite high LB scores! I hope I can get out of the 0.0-ish regime soon lol.

", "votes": 2, "replies": [ { "id": 1551384, "author_name": "Awsaf", "author_url": "", "post_date": "2021-10-20T15:09:16.600000", "content": "

It's not that much of a coincedence actually, unet and mask-rcnn are common for segmentation hence most people will go for it. Even if we didn't publish it someone else would've done it …. . I wish you all the best :)

", "votes": 3, "replies": [] } ] }, { "id": 1559113, "author_name": "Old Monk", "author_url": "", "post_date": "2021-10-26T18:24:09.873000", "content": "

Thanks for sharing @awsaf49 , really useful for other folks participating here!

", "votes": 0, "replies": [] }, { "id": 1558511, "author_name": "blueboy-97", "author_url": "", "post_date": "2021-10-26T09:44:25.320000", "content": "

Hello, how do you deal with the overlapped annotation when you create mask in your code? The npz files, to be precise.

", "votes": 0, "replies": [ { "id": 1558520, "author_name": "Awsaf", "author_url": "", "post_date": "2021-10-26T09:51:45.053000", "content": "

I'm using public dataset which uses following codes,

\n
for i, annot in enumerate(annotation): \n     img_mask = np.where(rle_decode(annot, (height, width))!=0, i, img_mask)\n
\n

So, I think it takes the first one.

", "votes": 0, "replies": [] }, { "id": 1558685, "author_name": "blueboy-97", "author_url": "", "post_date": "2021-10-26T12:42:19.107000", "content": "

I got it! thank you for reply.

", "votes": 1, "replies": [] } ] }, { "id": 1554009, "author_name": "Huajie Liang", "author_url": "", "post_date": "2021-10-22T18:36:24.150000", "content": "

No offense. Are frameworks like mmdetection allowed to be used in the competition?

", "votes": 0, "replies": [] }, { "id": 1558422, "author_name": "", "author_url": "", "post_date": "2021-10-26T08:06:57.677000", "content": "", "votes": 0, "replies": [] }, { "id": 1552600, "author_name": "", "author_url": "", "post_date": "2021-10-21T14:42:21.410000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1550640": "\n\nPublished some starter notebooks using **Mask-RCNN** with **MMDetection**. I hope it helps.\n\n## Keypoints:\n* Currently it get's `LB=0.270`. \n* Model isn't tuned hence one can improve the scores further.\n* It uses simple post-processing to remove the overlap between instances of the same image, check [here](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/279995) for more details.\n* Model: Mask-RCNN\n* Backbone: ResNest50\n* Optimizer: SGD\n* Epochs: 12\n* Augmentation: RandomFlip, PhotometricDistortion(Brightness, Contrast, Hue, etc)\n* **Wandb** Integration\n\n## Notebooks:\n\n**Mask-RCNN**:\n* Train: [Sartorius: MMDetection [Train]](https://www.kaggle.com/awsaf49/sartorius-mmdetection-train)\n* Infer: [Sartorius: MMDetection [Infer]](https://www.kaggle.com/awsaf49/sartorius-mmdetection-infer)\n\n**UNet**:\n* Train: [[PyTorch] Sartorius: UNet Strikes Back [Train] πŸ”₯](https://www.kaggle.com/awsaf49/pytorch-sartorius-unet-strikes-back-train/)\n* Infer: [[PyTorch] Sartorius: UNet Strikes Back [Infer] πŸ”₯](https://www.kaggle.com/awsaf49/pytorch-sartorius-unet-strikes-back-infer/)\n\n\n## WandB\nThese notebooks come with **wandb** integration. Hence you can,\n* Track training real-time.\n* Do Exploratory Data Analysis.\n* Debug your Pipeline.\n* Analyze your result.\n\"index\"\n\n## Examples:\n\"results-12-1\"\n\n## Predictions:\n* Good:\n\"results-31-0\"\n\n* Bad:\n\"results-33-0\"\n\n-------------------------\n## Update: 23/10/2021\n**Train** : `v16`\n* Custom Config is added for Hyperparameter Tuning\n* `first cell missing` issue fixed\n* batch_size is increased from `2` to `8`\n", "1552911": "@awsaf49 Helpful as usual, a great example for novices.\n", "1552429": "This will be a nice starting point! Thank you for sharing", "1552061": "Great work Awsaf! Thanks for sharing!", "1550687": "That was quick! great job! ", "1551400": "Nice notebook (train+infer) for starter instance segmentation with mmdet @awsaf49 ", "1551363": "@awsaf49 Man, I will just mention the elephant in the room so there are no hard feelings: we spent our Tuesdays working along the same investigation lines and came up with similar work. I just saw all your work today, but I see that the publishing hours might make it look weird. \n\nI had worked all Tuesday to come up with my notebook, and I was just waiting for my submission quota to renew to publish it. And I can tell that you worked hard as well.\n\nI want to state this so it's clear that it was just a shitty coincidence and not something else so we are cool ;)\n\nAnyways, I like your work very much. And you got quite high LB scores! I hope I can get out of the `0.0-ish` regime soon lol.\n", "1559113": "Thanks for sharing @awsaf49 , really useful for other folks participating here!", "1558511": "Hello, how do you deal with the overlapped annotation when you create mask in your code? The npz files, to be precise.", "1554009": "No offense. Are frameworks like mmdetection allowed to be used in the competition?", "1558422": "", "1552600": "" } }