![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2F088f0cba66bbed17902bfbf24d5bb56c%2FP1.png?generation=1589855029741157&alt=media\")
| ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2Fa70e038daa665e018dc5f3cd537761d2%2FP2.png?generation=1589855042062692&alt=media\")
|\n| --- | --- |\n| ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2Fd110a0bb735ec13c3f154e5942c28a8a%2FP3.png?generation=1589855053186548&alt=media\")
| ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2Fc1353f7909284938982b55385c2e8cc8%2FP4.png?generation=1589855062668683&alt=media\")
|\n|![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2Fa9f26f3298de9d0edf91523f556c5ea5%2FP5.png?generation=1589855131489666&alt=media\")
|![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2Fab8c10a2ddb8edae4b6a1b2f6a06dd05%2FP6.png?generation=1589855148126106&alt=media\")
|\n|![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2F5d669a88bbcdaaf862454933c387cc73%2FP7.png?generation=1589855165469084&alt=media\")
|![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2F2a098122c88f93aafa883f856cadcfab%2FP8.png?generation=1589855196668046&alt=media\")
|\n|![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2F36fa313a7fb1a053be0c832b4d8d210c%2FP9.png?generation=1589855233731502&alt=media\")
| ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F651363%2F9af3010755d49847cc13e10f9580194c%2FP10.png?generation=1589855250304359&alt=media\")
|These are the masks produced by our model and their corresponding in the correct dimension. The more yellower the color the higher the label class is (This is just done for displaying here and not how we are submitting it).\nThe images are P1 through P10 row wise with P1 and P2 being the first row. The predictions for them and the corresponding items of clothing are given below the images:-
\n\n| | |\n| --- | --- |\n| | |\n|||\n|| |\n|| |
P1- [31 31 10 33]\nP2- [33]\nP3- [31 31 23 23 23 28 6 29 1 4 23]\nP4- [31 14]\nP5- [23 10 23]\nP6- [ 6 23 23 31 31 10 33 22 10 23 1]\nP7- [23 23 23 31 23 33 31 23 23 1 23]\nP8- [31 31 23 6 29 33 31 23 23]\nP9- [23 23 6 32 1 32 33]\nP10-[23 33 10 10]
\n\nLegend:-\n1-top,t-shirt,sweatshirt, 4-jacket, 6-pants, 8-skirt\n10-dress, 14-hat, 23-shoe\n28-collar, 29-lapel, 31-sleeve\n32-pocket, 33-neckline, 35-zipper
\n\nThese are the attr ids predictions:\nP1 - ['31: 295', '31: 295', '10: 115,295,317', '33: 295']\nP2 - ['33: 295']\nP3- ['31: 295', '31: 295', '23: ', '23: ', '23: ', '28: 295', '6: 115,295,317', '29: 295', '1: 295', '4: 295', '23: ']\nP4- ['31: 295', '14: ']\nP5 - ['23: ', '10: 115,295,317', '23: ']\nP6 - ['6: 295', '23: ', '23: ', '31: 295', '31: 295', '10: 295', '33: 295', '22: ', '10: 115,295', '23: ', '1: 295']\nP7 - ['23: ', '23: ', '23: ', '31: 295', '23: ', '33: 295', '31: 295', '23: ', '23: ', '1: 295', '23: ']\nP8 - ['31: 295', '31: 295', '23: ', '6: 295', '29: 295', '33: 295', '31: 295', '23: ', '23: ']\nP9 - ['23: ', '23: ', '6: 115,295', '32: 295', '1: 295', '32: 295', '33: 295']\nP10 - ['31: 295', '1: 295', '32: 295', '32: 295', '33: 295', '8: 295', '33: 295', '33: 295', '1: 295', '31: 295', '23: ', '35: 295']\nP11 - ['23: ', '33: 295', '10: 295', '10: 115,295,317']
\n\nHope someone can help me understand why this gets a 0 score?These masks are not perfect but I mean shouldn't at least one of these images (and there are total 3200 such images) get a non zero score?\nI anyone can figure out looking at these results what's wrong please let me know! Really don't know what is wrong
\n\nNote: Already tried all the suggestions made in the previous 2 discussion posts. Thanks to everyone who helped!\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/150915\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/151974
\n\n@makeitworkjml @polosin @loushun @guiguzhixing @shraddhaamohan
\n\nSubmission csv file for P1 to P9 also attached.
", "messages": [ { "id": "853182", "postDate": "05/19/2020 02:34:37", "content": "These are the masks produced by our model and their corresponding in the correct dimension. The more yellower the color the higher the label class is (This is just done for displaying here and not how we are submitting it).\nThe images are P1 through P10 row wise with P1 and P2 being the first row. The predictions for them and the corresponding items of clothing are given below the images:-
\n\n| | |\n| --- | --- |\n| | |\n|||\n|| |\n|| |
P1- [31 31 10 33]\nP2- [33]\nP3- [31 31 23 23 23 28 6 29 1 4 23]\nP4- [31 14]\nP5- [23 10 23]\nP6- [ 6 23 23 31 31 10 33 22 10 23 1]\nP7- [23 23 23 31 23 33 31 23 23 1 23]\nP8- [31 31 23 6 29 33 31 23 23]\nP9- [23 23 6 32 1 32 33]\nP10-[23 33 10 10]
\n\nLegend:-\n1-top,t-shirt,sweatshirt, 4-jacket, 6-pants, 8-skirt\n10-dress, 14-hat, 23-shoe\n28-collar, 29-lapel, 31-sleeve\n32-pocket, 33-neckline, 35-zipper
\n\nThese are the attr ids predictions:\nP1 - ['31: 295', '31: 295', '10: 115,295,317', '33: 295']\nP2 - ['33: 295']\nP3- ['31: 295', '31: 295', '23: ', '23: ', '23: ', '28: 295', '6: 115,295,317', '29: 295', '1: 295', '4: 295', '23: ']\nP4- ['31: 295', '14: ']\nP5 - ['23: ', '10: 115,295,317', '23: ']\nP6 - ['6: 295', '23: ', '23: ', '31: 295', '31: 295', '10: 295', '33: 295', '22: ', '10: 115,295', '23: ', '1: 295']\nP7 - ['23: ', '23: ', '23: ', '31: 295', '23: ', '33: 295', '31: 295', '23: ', '23: ', '1: 295', '23: ']\nP8 - ['31: 295', '31: 295', '23: ', '6: 295', '29: 295', '33: 295', '31: 295', '23: ', '23: ']\nP9 - ['23: ', '23: ', '6: 115,295', '32: 295', '1: 295', '32: 295', '33: 295']\nP10 - ['31: 295', '1: 295', '32: 295', '32: 295', '33: 295', '8: 295', '33: 295', '33: 295', '1: 295', '31: 295', '23: ', '35: 295']\nP11 - ['23: ', '33: 295', '10: 295', '10: 115,295,317']
\n\nHope someone can help me understand why this gets a 0 score?These masks are not perfect but I mean shouldn't at least one of these images (and there are total 3200 such images) get a non zero score?\nI anyone can figure out looking at these results what's wrong please let me know! Really don't know what is wrong
\n\nNote: Already tried all the suggestions made in the previous 2 discussion posts. Thanks to everyone who helped!\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/150915\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/151974
\n\n@makeitworkjml @polosin @loushun @guiguzhixing @shraddhaamohan
\n\nSubmission csv file for P1 to P9 also attached.
", "rawMarkdown": "These are the masks produced by our model and their corresponding in the correct dimension. The more yellower the color the higher the label class is (This is just done for displaying here and not how we are submitting it).\nThe images are P1 through P10 row wise with P1 and P2 being the first row. The predictions for them and the corresponding items of clothing are given below the images:- \n\n| | |\n| --- | --- |\n| | |\n|||\n|| |\n|| |\n\n\n### **Predictions: **\nP1- [31 31 10 33]\nP2- [33]\nP3- [31 31 23 23 23 28 6 29 1 4 23]\nP4- [31 14]\nP5- [23 10 23]\nP6- [ 6 23 23 31 31 10 33 22 10 23 1]\nP7- [23 23 23 31 23 33 31 23 23 1 23]\nP8- [31 31 23 6 29 33 31 23 23]\nP9- [23 23 6 32 1 32 33]\nP10-[23 33 10 10]\n\nLegend:-\n1-top,t-shirt,sweatshirt, 4-jacket, 6-pants, 8-skirt\n10-dress, 14-hat, 23-shoe\n28-collar, 29-lapel, 31-sleeve\n32-pocket, 33-neckline, 35-zipper\n\nThese are the attr ids predictions:\nP1 - ['31: 295', '31: 295', '10: 115,295,317', '33: 295']\nP2 - ['33: 295']\nP3- ['31: 295', '31: 295', '23: ', '23: ', '23: ', '28: 295', '6: 115,295,317', '29: 295', '1: 295', '4: 295', '23: ']\nP4- ['31: 295', '14: ']\nP5 - ['23: ', '10: 115,295,317', '23: ']\nP6 - ['6: 295', '23: ', '23: ', '31: 295', '31: 295', '10: 295', '33: 295', '22: ', '10: 115,295', '23: ', '1: 295']\nP7 - ['23: ', '23: ', '23: ', '31: 295', '23: ', '33: 295', '31: 295', '23: ', '23: ', '1: 295', '23: ']\nP8 - ['31: 295', '31: 295', '23: ', '6: 295', '29: 295', '33: 295', '31: 295', '23: ', '23: ']\nP9 - ['23: ', '23: ', '6: 115,295', '32: 295', '1: 295', '32: 295', '33: 295']\nP10 - ['31: 295', '1: 295', '32: 295', '32: 295', '33: 295', '8: 295', '33: 295', '33: 295', '1: 295', '31: 295', '23: ', '35: 295']\nP11 - ['23: ', '33: 295', '10: 295', '10: 115,295,317']\n\n\n**Hope someone can help me understand why this gets a 0 score?These masks are not perfect but I mean shouldn't at least one of these images (and there are total 3200 such images) get a non zero score?\nI anyone can figure out looking at these results what's wrong please let me know! Really don't know what is wrong**\n\nNote: Already tried all the suggestions made in the previous 2 discussion posts. Thanks to everyone who helped!\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/150915\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/151974\n\n@makeitworkjml @polosin @loushun @guiguzhixing @shraddhaamohan \n\nSubmission csv file for P1 to P9 also attached.", "votes": null }, { "id": "853204", "postDate": "05/19/2020 03:04:22", "content": "How many rows are usually in your submission file?
", "rawMarkdown": "How many rows are usually in your submission file?", "votes": null }, { "id": "853316", "postDate": "05/19/2020 04:55:45", "content": "My last submission had 19,906 rows. I'm guessing 20,000 approx for each. Why do you ask? @interneuron
", "rawMarkdown": "My last submission had 19,906 rows. I'm guessing 20,000 approx for each. Why do you ask? @interneuron", "votes": null }, { "id": "853329", "postDate": "05/19/2020 05:12:37", "content": "I was thinking you may have been entering the classids in an unexpected way but that seems not to be the case.
\n\nThe advice given in your other threads is accurate, I looked at them carefully because I was also repeatedly getting 0 with sensible looking masks.
\n\nThe functions for encoding, decoding, and resizing rle given in the public notebooks are sufficient for generating a valid submission. In my case I had done something weird with the class indicies, and it was only after I compared some masks of a given class index with its identity in the json file I saw they were not what I expected, for example I looked at the class id for sleeve but saw a mask for a shirt.
", "rawMarkdown": "I was thinking you may have been entering the classids in an unexpected way but that seems not to be the case. \n\nThe advice given in your other threads is accurate, I looked at them carefully because I was also repeatedly getting 0 with sensible looking masks. \n\nThe functions for encoding, decoding, and resizing rle given in the public notebooks are sufficient for generating a valid submission. In my case I had done something weird with the class indicies, and it was only after I compared some masks of a given class index with its identity in the json file I saw they were not what I expected, for example I looked at the class id for sleeve but saw a mask for a shirt.", "votes": null }, { "id": "853368", "postDate": "05/19/2020 06:13:09", "content": "these images are from test set? Have you tried inferencing on the training set? It would be helpful to compare images from training set where you have ground-truth to compare to, so you can rule out any pre-processing and post-processing issues.
\n\nAnd how about your attributes predictions?
", "rawMarkdown": "these images are from test set? Have you tried inferencing on the training set? It would be helpful to compare images from training set where you have ground-truth to compare to, so you can rule out any pre-processing and post-processing issues.\n\nAnd how about your attributes predictions?", "votes": null }, { "id": "853401", "postDate": "05/19/2020 06:44:10", "content": "For the post-processing of the predictions, we have presented exampled code and requirements (e.g. 1). convert to EncodedPixels, 2). resize images, 3). get attributeIds, 4). the need to remove overlapped masks for the same category). Did you check those processes?
\n\nAnd classId 0-indexing is also a good point from previous 2 posts, since coco-based detection systems tend to use 0 for background. \nHow about attribute ids? I assumed that you used continuous ids for the attributes during training process. Have you remap the AttributeIds back to the original ids?
", "rawMarkdown": "For the post-processing of the predictions, we have presented exampled code and requirements (e.g. 1). convert to EncodedPixels, 2). resize images, 3). get attributeIds, 4). the need to remove overlapped masks for the same category). Did you check those processes? \n\nAnd classId 0-indexing is also a good point from previous 2 posts, since coco-based detection systems tend to use 0 for background. \nHow about attribute ids? I assumed that you used continuous ids for the attributes during training process. Have you remap the AttributeIds back to the original ids?", "votes": null }, { "id": "853457", "postDate": "05/19/2020 07:48:34", "content": "You did not deduce the attribute id.
", "rawMarkdown": "You did not deduce the attribute id.", "votes": null }, { "id": "853818", "postDate": "05/19/2020 14:06:33", "content": "It is possible to get a score above 0.0000 with a blank attribute id column.
", "rawMarkdown": "It is possible to get a score above 0.0000 with a blank attribute id column.", "votes": null }, { "id": "854028", "postDate": "05/19/2020 17:24:11", "content": "Hmm, I checked that the class id's are correct. For example if you look at the first image the class ids are [31 31 10 33] - which is [sleeve, sleeve, dress and neckline] which if you look at the image is correct and that is case for the other images too. \nCan you give the example of your mistake in classIds and how you corrected it? Maybe i am making the same experiment. \n@interneuron
", "rawMarkdown": "Hmm, I checked that the class id's are correct. For example if you look at the first image the class ids are [31 31 10 33] - which is [sleeve, sleeve, dress and neckline] which if you look at the image is correct and that is case for the other images too. \nCan you give the example of your mistake in classIds and how you corrected it? Maybe i am making the same experiment. \n@interneuron", "votes": null }, { "id": "854037", "postDate": "05/19/2020 17:37:09", "content": "Just added the attribute ids to the post! @interneuron @hongym7 \nYa so some of the entries as you see above are blank but not all of them.
", "rawMarkdown": "Just added the attribute ids to the post! @interneuron @hongym7 \nYa so some of the entries as you see above are blank but not all of them.", "votes": null }, { "id": "854041", "postDate": "05/19/2020 17:40:58", "content": "If you are referring to these code snippets https://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/overview/evaluation, then yes I am doing is as mentioned here.
Yes i checked, the class indexing predictions start from 0 to 45 and not 1 to 46
Yes i am remapping attributeIds back to the original ids. I just shared the attribute predictions in the post to.\n@makeitworkjml
Added attribute predictions to the original post.
\n\nI don't get the point of running again on training images. I mean the encoding and decoding works well on testing images and i would expect it to work the same way with the same rle functions and same model.\n@makeitworkjml
", "rawMarkdown": "Added attribute predictions to the original post. \n\nI don't get the point of running again on training images. I mean the encoding and decoding works well on testing images and i would expect it to work the same way with the same rle functions and same model.\n@makeitworkjml", "votes": null }, { "id": "854073", "postDate": "05/19/2020 18:22:55", "content": "I had left a classid-1 somewhere in the code by mistake where my inference function was already accounting for the index change between model predictions and expected class id.
", "rawMarkdown": "I had left a classid-1 somewhere in the code by mistake where my inference function was already accounting for the index change between model predictions and expected class id.", "votes": null }, { "id": "854084", "postDate": "05/19/2020 18:40:50", "content": "@mihirmavalankar Can you also generate a submission CSV file with these 10 records and post the content of the file here?
", "rawMarkdown": "mihirmavalankar Can you also generate a submission CSV file with these 10 records and post the content of the file here?", "votes": null }, { "id": "854350", "postDate": "05/20/2020 01:26:23", "content": "Hey here is the csv file for results for images P1 to P9. P10 was taked from somewhere in the middle of test data. Thanks! Maybe I should have shared this sooner. Also put it in the original post on top.\n@polosin
", "rawMarkdown": "Hey here is the csv file for results for images P1 to P9. P10 was taked from somewhere in the middle of test data. Thanks! Maybe I should have shared this sooner. Also put it in the original post on top.\n@polosin", "votes": null }, { "id": "854374", "postDate": "05/20/2020 02:00:37", "content": "This is how your segmentation actually looks like: ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F853978%2F43c5093eacbe7c469eaf7ad7f5080328%2Fsegmentation.png?generation=1589938600700077&alt=media\")
When I changed the order for the np.reshape() method to the default one, it worked:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F853978%2Ffc4a20e04c5a07c8029219c993b63a72%2Fsegmentation2.png?generation=1589939297375243&alt=media\")
So this means you're using a wrong order when encoding masks. Take the rle_encode method from this notebook.
OMG! Finally! It was one line of code\nFrom: masks = masks.reshape((n,h*w))\nTo: masks = masks.reshape((n,h*w), order='F')
\n\nThanks so much for this @polosin! We owe you one!
", "rawMarkdown": "OMG! Finally! It was one line of code\nFrom: masks = masks.reshape((n,h*w))\nTo: masks = masks.reshape((n,h*w), order='F')\n\nThanks so much for this @polosin! We owe you one!", "votes": null } ], "comments": [ { "id": 853204, "author_name": "interneuron", "author_url": "", "post_date": "05/19/2020 03:04:22", "content": "How many rows are usually in your submission file?
", "votes": null, "replies": [ { "id": 853316, "author_name": "mihirmavalankar", "author_url": "", "post_date": "05/19/2020 04:55:45", "content": "My last submission had 19,906 rows. I'm guessing 20,000 approx for each. Why do you ask? @interneuron
", "votes": null, "replies": [] }, { "id": 853329, "author_name": "interneuron", "author_url": "", "post_date": "05/19/2020 05:12:37", "content": "I was thinking you may have been entering the classids in an unexpected way but that seems not to be the case.
\n\nThe advice given in your other threads is accurate, I looked at them carefully because I was also repeatedly getting 0 with sensible looking masks.
\n\nThe functions for encoding, decoding, and resizing rle given in the public notebooks are sufficient for generating a valid submission. In my case I had done something weird with the class indicies, and it was only after I compared some masks of a given class index with its identity in the json file I saw they were not what I expected, for example I looked at the class id for sleeve but saw a mask for a shirt.
", "votes": null, "replies": [] }, { "id": 854028, "author_name": "mihirmavalankar", "author_url": "", "post_date": "05/19/2020 17:24:11", "content": "Hmm, I checked that the class id's are correct. For example if you look at the first image the class ids are [31 31 10 33] - which is [sleeve, sleeve, dress and neckline] which if you look at the image is correct and that is case for the other images too. \nCan you give the example of your mistake in classIds and how you corrected it? Maybe i am making the same experiment. \n@interneuron
", "votes": null, "replies": [] }, { "id": 854073, "author_name": "interneuron", "author_url": "", "post_date": "05/19/2020 18:22:55", "content": "I had left a classid-1 somewhere in the code by mistake where my inference function was already accounting for the index change between model predictions and expected class id.
", "votes": null, "replies": [] } ] }, { "id": 853368, "author_name": "makeitworkjml", "author_url": "", "post_date": "05/19/2020 06:13:09", "content": "these images are from test set? Have you tried inferencing on the training set? It would be helpful to compare images from training set where you have ground-truth to compare to, so you can rule out any pre-processing and post-processing issues.
\n\nAnd how about your attributes predictions?
", "votes": null, "replies": [ { "id": 854042, "author_name": "mihirmavalankar", "author_url": "", "post_date": "05/19/2020 17:42:52", "content": "Added attribute predictions to the original post.
\n\nI don't get the point of running again on training images. I mean the encoding and decoding works well on testing images and i would expect it to work the same way with the same rle functions and same model.\n@makeitworkjml
", "votes": null, "replies": [] } ] }, { "id": 853401, "author_name": "makeitworkjml", "author_url": "", "post_date": "05/19/2020 06:44:10", "content": "For the post-processing of the predictions, we have presented exampled code and requirements (e.g. 1). convert to EncodedPixels, 2). resize images, 3). get attributeIds, 4). the need to remove overlapped masks for the same category). Did you check those processes?
\n\nAnd classId 0-indexing is also a good point from previous 2 posts, since coco-based detection systems tend to use 0 for background. \nHow about attribute ids? I assumed that you used continuous ids for the attributes during training process. Have you remap the AttributeIds back to the original ids?
", "votes": null, "replies": [ { "id": 854041, "author_name": "mihirmavalankar", "author_url": "", "post_date": "05/19/2020 17:40:58", "content": "If you are referring to these code snippets https://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/overview/evaluation, then yes I am doing is as mentioned here.
Yes i checked, the class indexing predictions start from 0 to 45 and not 1 to 46
Yes i am remapping attributeIds back to the original ids. I just shared the attribute predictions in the post to.\n@makeitworkjml
You did not deduce the attribute id.
", "votes": null, "replies": [ { "id": 853818, "author_name": "interneuron", "author_url": "", "post_date": "05/19/2020 14:06:33", "content": "It is possible to get a score above 0.0000 with a blank attribute id column.
", "votes": null, "replies": [] }, { "id": 854037, "author_name": "mihirmavalankar", "author_url": "", "post_date": "05/19/2020 17:37:09", "content": "Just added the attribute ids to the post! @interneuron @hongym7 \nYa so some of the entries as you see above are blank but not all of them.
", "votes": null, "replies": [] } ] }, { "id": 854084, "author_name": "polosin", "author_url": "", "post_date": "05/19/2020 18:40:50", "content": "@mihirmavalankar Can you also generate a submission CSV file with these 10 records and post the content of the file here?
", "votes": null, "replies": [ { "id": 854350, "author_name": "mihirmavalankar", "author_url": "", "post_date": "05/20/2020 01:26:23", "content": "Hey here is the csv file for results for images P1 to P9. P10 was taked from somewhere in the middle of test data. Thanks! Maybe I should have shared this sooner. Also put it in the original post on top.\n@polosin
", "votes": null, "replies": [] }, { "id": 854374, "author_name": "polosin", "author_url": "", "post_date": "05/20/2020 02:00:37", "content": "This is how your segmentation actually looks like:
When I changed the order for the np.reshape() method to the default one, it worked:\n
So this means you're using a wrong order when encoding masks. Take the rle_encode method from this notebook.
OMG! Finally! It was one line of code\nFrom: masks = masks.reshape((n,h*w))\nTo: masks = masks.reshape((n,h*w), order='F')
\n\nThanks so much for this @polosin! We owe you one!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "853182": "These are the masks produced by our model and their corresponding in the correct dimension. The more yellower the color the higher the label class is (This is just done for displaying here and not how we are submitting it).\nThe images are P1 through P10 row wise with P1 and P2 being the first row. The predictions for them and the corresponding items of clothing are given below the images:- \n\n| | |\n| --- | --- |\n| | |\n|||\n|| |\n|| |\n\n\n### **Predictions: **\nP1- [31 31 10 33]\nP2- [33]\nP3- [31 31 23 23 23 28 6 29 1 4 23]\nP4- [31 14]\nP5- [23 10 23]\nP6- [ 6 23 23 31 31 10 33 22 10 23 1]\nP7- [23 23 23 31 23 33 31 23 23 1 23]\nP8- [31 31 23 6 29 33 31 23 23]\nP9- [23 23 6 32 1 32 33]\nP10-[23 33 10 10]\n\nLegend:-\n1-top,t-shirt,sweatshirt, 4-jacket, 6-pants, 8-skirt\n10-dress, 14-hat, 23-shoe\n28-collar, 29-lapel, 31-sleeve\n32-pocket, 33-neckline, 35-zipper\n\nThese are the attr ids predictions:\nP1 - ['31: 295', '31: 295', '10: 115,295,317', '33: 295']\nP2 - ['33: 295']\nP3- ['31: 295', '31: 295', '23: ', '23: ', '23: ', '28: 295', '6: 115,295,317', '29: 295', '1: 295', '4: 295', '23: ']\nP4- ['31: 295', '14: ']\nP5 - ['23: ', '10: 115,295,317', '23: ']\nP6 - ['6: 295', '23: ', '23: ', '31: 295', '31: 295', '10: 295', '33: 295', '22: ', '10: 115,295', '23: ', '1: 295']\nP7 - ['23: ', '23: ', '23: ', '31: 295', '23: ', '33: 295', '31: 295', '23: ', '23: ', '1: 295', '23: ']\nP8 - ['31: 295', '31: 295', '23: ', '6: 295', '29: 295', '33: 295', '31: 295', '23: ', '23: ']\nP9 - ['23: ', '23: ', '6: 115,295', '32: 295', '1: 295', '32: 295', '33: 295']\nP10 - ['31: 295', '1: 295', '32: 295', '32: 295', '33: 295', '8: 295', '33: 295', '33: 295', '1: 295', '31: 295', '23: ', '35: 295']\nP11 - ['23: ', '33: 295', '10: 295', '10: 115,295,317']\n\n\n**Hope someone can help me understand why this gets a 0 score?These masks are not perfect but I mean shouldn't at least one of these images (and there are total 3200 such images) get a non zero score?\nI anyone can figure out looking at these results what's wrong please let me know! Really don't know what is wrong**\n\nNote: Already tried all the suggestions made in the previous 2 discussion posts. Thanks to everyone who helped!\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/150915\nhttps://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/discussion/151974\n\n@makeitworkjml @polosin @loushun @guiguzhixing @shraddhaamohan \n\nSubmission csv file for P1 to P9 also attached.", "853204": "How many rows are usually in your submission file?", "853316": "My last submission had 19,906 rows. I'm guessing 20,000 approx for each. Why do you ask? @interneuron", "853329": "I was thinking you may have been entering the classids in an unexpected way but that seems not to be the case. \n\nThe advice given in your other threads is accurate, I looked at them carefully because I was also repeatedly getting 0 with sensible looking masks. \n\nThe functions for encoding, decoding, and resizing rle given in the public notebooks are sufficient for generating a valid submission. In my case I had done something weird with the class indicies, and it was only after I compared some masks of a given class index with its identity in the json file I saw they were not what I expected, for example I looked at the class id for sleeve but saw a mask for a shirt.", "853368": "these images are from test set? Have you tried inferencing on the training set? It would be helpful to compare images from training set where you have ground-truth to compare to, so you can rule out any pre-processing and post-processing issues.\n\nAnd how about your attributes predictions?", "853401": "For the post-processing of the predictions, we have presented exampled code and requirements (e.g. 1). convert to EncodedPixels, 2). resize images, 3). get attributeIds, 4). the need to remove overlapped masks for the same category). Did you check those processes? \n\nAnd classId 0-indexing is also a good point from previous 2 posts, since coco-based detection systems tend to use 0 for background. \nHow about attribute ids? I assumed that you used continuous ids for the attributes during training process. Have you remap the AttributeIds back to the original ids?", "853457": "You did not deduce the attribute id.", "853818": "It is possible to get a score above 0.0000 with a blank attribute id column.", "854028": "Hmm, I checked that the class id's are correct. For example if you look at the first image the class ids are [31 31 10 33] - which is [sleeve, sleeve, dress and neckline] which if you look at the image is correct and that is case for the other images too. \nCan you give the example of your mistake in classIds and how you corrected it? Maybe i am making the same experiment. \n@interneuron", "854037": "Just added the attribute ids to the post! @interneuron @hongym7 \nYa so some of the entries as you see above are blank but not all of them.", "854041": "1. If you are referring to these code snippets https://www.kaggle.com/c/imaterialist-fashion-2020-fgvc7/overview/evaluation, then yes I am doing is as mentioned here.\n\n2. Yes i checked, the class indexing predictions start from 0 to 45 and not 1 to 46\n\n3. Yes i am remapping attributeIds back to the original ids. I just shared the attribute predictions in the post to.\n@makeitworkjml", "854042": "Added attribute predictions to the original post. \n\nI don't get the point of running again on training images. I mean the encoding and decoding works well on testing images and i would expect it to work the same way with the same rle functions and same model.\n@makeitworkjml", "854073": "I had left a classid-1 somewhere in the code by mistake where my inference function was already accounting for the index change between model predictions and expected class id.", "854084": "mihirmavalankar Can you also generate a submission CSV file with these 10 records and post the content of the file here?", "854350": "Hey here is the csv file for results for images P1 to P9. P10 was taked from somewhere in the middle of test data. Thanks! Maybe I should have shared this sooner. Also put it in the original post on top.\n@polosin", "854374": "This is how your segmentation actually looks like: \n\nWhen I changed the order for the `np.reshape()` method to the default one, it worked:\n\n\nSo this means you're using a wrong order when encoding masks. Take the `rle_encode` method from [this](https://www.kaggle.com/stainsby/fast-tested-rle-and-input-routines) notebook.", "854624": "OMG! Finally! It was one line of code\nFrom: masks = masks.reshape((n,h*w))\nTo: masks = masks.reshape((n,h*w), order='F')\n\nThanks so much for this @polosin! We owe you one!" }, "source": "meta" }