During the training of the model, I tried to identify all irrelevant images that shouldn't be present in the dataset. These dataset ids have been categorized into 2 parts. The first category is irrelevant images (all images are cassava fruits not leaves) and the second one is noisy images (all images are having more background area rather than leaves). However, removing these datasets has been decreased the accuracy of the model with the same model but I believe these are not relevant for feeding the model to just make them learn wrong images.
\nirrelevant image ids = [274726002, 9224019, 159654644, 199112616, 226533928, 262902341, 269713568, 274726002, 384390206, 390601409, 421035788, 457405364, 600736721, 580111608,
\n 616718743, 695438825, 723564013, 826231979, 847847826, 927165736, 1004389140,
\n 1008244905, 1338159402, 1339403533, 1359893940, 1366430957, 1689510013, 9224019,
\n 4269208386, 4239074071, 3810809174, 3652033201, 3609350672, 3609986814,
\n 3477169212, 3435954655, 3425850136, 3251960666, 3252232501, 3199643560,
\n 3126296051, 3040241097, 2981404650, 2925605732, 2839068946, 2698282165,
\n 2604713994, 2415837573, 2382642453, 2321669192, 2320471703, 2278166989,
\n 2276509518, 2262263316, 2182500020, 2139839273, 2084868828, 1848686439,
\n 1689510013, 1359893940]
noisy image ids = [ 410880003, 411955232, 501215014, 549854027, 554488826,
\n 724195836, 744383303, 888983519, 1096438409, 1130568730,
\n 1709404074, 1770746162, 4280523848, 3530560257, 3421208425,
\n 3321193739, 3086663390, 3045134829, 1862072615]
Notes: I tried to upload some sample images but seems Kaggle is not allowing me to do so. Showing the message We have disabled uploading forum attachments for the time being. Please use an alternative host for your file, and link to it from your forum post.
", "messages": [ { "id": 1186462, "postDate": "2021-02-04T20:41:42.137Z", "content": "During the training of the model, I tried to identify all irrelevant images that shouldn't be present in the dataset. These dataset ids have been categorized into 2 parts. The first category is irrelevant images (all images are cassava fruits not leaves) and the second one is noisy images (all images are having more background area rather than leaves). However, removing these datasets has been decreased the accuracy of the model with the same model but I believe these are not relevant for feeding the model to just make them learn wrong images.
\nirrelevant image ids = [274726002, 9224019, 159654644, 199112616, 226533928, 262902341, 269713568, 274726002, 384390206, 390601409, 421035788, 457405364, 600736721, 580111608,
\n 616718743, 695438825, 723564013, 826231979, 847847826, 927165736, 1004389140,
\n 1008244905, 1338159402, 1339403533, 1359893940, 1366430957, 1689510013, 9224019,
\n 4269208386, 4239074071, 3810809174, 3652033201, 3609350672, 3609986814,
\n 3477169212, 3435954655, 3425850136, 3251960666, 3252232501, 3199643560,
\n 3126296051, 3040241097, 2981404650, 2925605732, 2839068946, 2698282165,
\n 2604713994, 2415837573, 2382642453, 2321669192, 2320471703, 2278166989,
\n 2276509518, 2262263316, 2182500020, 2139839273, 2084868828, 1848686439,
\n 1689510013, 1359893940]
noisy image ids = [ 410880003, 411955232, 501215014, 549854027, 554488826,
\n 724195836, 744383303, 888983519, 1096438409, 1130568730,
\n 1709404074, 1770746162, 4280523848, 3530560257, 3421208425,
\n 3321193739, 3086663390, 3045134829, 1862072615]
Notes: I tried to upload some sample images but seems Kaggle is not allowing me to do so. Showing the message We have disabled uploading forum attachments for the time being. Please use an alternative host for your file, and link to it from your forum post.
", "rawMarkdown": "During the training of the model, I tried to identify all irrelevant images that shouldn't be present in the dataset. These dataset ids have been categorized into 2 parts. The first category is irrelevant images (all images are cassava fruits not leaves) and the second one is noisy images (all images are having more background area rather than leaves). However, removing these datasets has been decreased the accuracy of the model with the same model but I believe these are not relevant for feeding the model to just make them learn wrong images. \n\n***irrelevant image ids*** = [274726002, 9224019, 159654644, 199112616, 226533928, 262902341, 269713568, 274726002, 384390206, 390601409, 421035788, 457405364, 600736721, 580111608,\n 616718743, 695438825, 723564013, 826231979, 847847826, 927165736, 1004389140, \n 1008244905, 1338159402, 1339403533, 1359893940, 1366430957, 1689510013, 9224019,\n 4269208386, 4239074071, 3810809174, 3652033201, 3609350672, 3609986814, \n 3477169212, 3435954655, 3425850136, 3251960666, 3252232501, 3199643560, \n 3126296051, 3040241097, 2981404650, 2925605732, 2839068946, 2698282165,\n 2604713994, 2415837573, 2382642453, 2321669192, 2320471703, 2278166989,\n 2276509518, 2262263316, 2182500020, 2139839273, 2084868828, 1848686439,\n 1689510013, 1359893940]\n\n***noisy image ids*** = [ 410880003, 411955232, 501215014, 549854027, 554488826, \n 724195836, 744383303, 888983519, 1096438409, 1130568730, \n 1709404074, 1770746162, 4280523848, 3530560257, 3421208425, \n 3321193739, 3086663390, 3045134829, 1862072615]\n\n**Notes**: I tried to upload some sample images but seems Kaggle is not allowing me to do so. Showing the message *We have disabled uploading forum attachments for the time being. Please use an alternative host for your file, and link to it from your forum post.*", "votes": 28 }, { "id": 1186556, "postDate": "2021-02-04T22:22:25.863Z", "content": "Why are there so many root images in the dataset? I understand that some leaf images are wrong labeled but i don't get why someone would put the root images in the data?
", "rawMarkdown": "Why are there so many root images in the dataset? I understand that some leaf images are wrong labeled but i don't get why someone would put the root images in the data?", "votes": 3, "replies": [ { "id": 1186572, "postDate": "2021-02-04T22:44:03.560Z", "content": "There was what is now a very old USA television show called \"Have Gun - Will Travel\". It was a cowboy TV series about a professional gun fighter.
\n\"Have camera - will take photo\" (or selfie) is perhaps a modern version. The train and test set for 2019 did not contain roots (at least as found by my root model).
\nThis years competition appears aimed at helping weed out the noise that will occur when you place a camera in the hands of a novice. There are a number of strange images in the 2020 test set - my favorite is a picture taken of a picture of a leaf.
", "rawMarkdown": "There was what is now a very old USA television show called \"Have Gun - Will Travel\". It was a cowboy TV series about a professional gun fighter.\n\n\"Have camera - will take photo\" (or selfie) is perhaps a modern version. The train and test set for 2019 did not contain roots (at least as found by my root model). \n\nThis years competition appears aimed at helping weed out the noise that will occur when you place a camera in the hands of a novice. There are a number of strange images in the 2020 test set - my favorite is a picture taken of a picture of a leaf. " }, { "id": 1186986, "postDate": "2021-02-05T06:38:26.213Z", "content": "I think Cassava roots can also indicate what disease the leaves are suffering from.
", "rawMarkdown": "I think Cassava roots can also indicate what disease the leaves are suffering from." } ] }, { "id": 1187923, "postDate": "2021-02-05T18:55:13.113Z", "content": "By eyeballing the 'wrong prediction' of my model, I find that there may be much more irrelevant, wrong focused and error label.
", "rawMarkdown": "By eyeballing the 'wrong prediction' of my model, I find that there may be much more irrelevant, wrong focused and error label.", "votes": 1, "replies": [ { "id": 1187980, "postDate": "2021-02-05T20:35:44.550Z", "content": "That's true.
", "rawMarkdown": "That's true." } ] }, { "id": 1186554, "postDate": "2021-02-04T22:19:50.397Z", "content": "The images of the roots (fruit) number closer to at least 96, so I think you missed a bunch. Not sure of the total since I worked with all 43K images that are found in the 2019 and 2020 data. (9 in the 2019 extra images) and stopped my search after the 4th version of my roots model.
\nIt's not a real difficult task to build a model to identify roots. It took me 4 iterations of a model to get the 96 identified after I started with around 25 detected manually.
\nThey are almost all class 1 (all but 8).
", "rawMarkdown": "The images of the roots (fruit) number closer to at least 96, so I think you missed a bunch. Not sure of the total since I worked with all 43K images that are found in the 2019 and 2020 data. (9 in the 2019 extra images) and stopped my search after the 4th version of my roots model.\n\nIt's not a real difficult task to build a model to identify roots. It took me 4 iterations of a model to get the 96 identified after I started with around 25 detected manually.\n\nThey are almost all class 1 (all but 8).", "votes": 1, "replies": [ { "id": 1186558, "postDate": "2021-02-04T22:23:40.263Z", "content": "@pcjimmmy i found this root dataset, if you need some more imeages for your root classifier: https://data.mendeley.com/datasets/gvp7vshvnh/1
", "rawMarkdown": "@pcjimmmy i found this root dataset, if you need some more imeages for your root classifier: https://data.mendeley.com/datasets/gvp7vshvnh/1", "votes": 1 }, { "id": 1186562, "postDate": "2021-02-04T22:34:07.600Z", "content": "Thanks - what I really need is the Python skill set to successfully use my root classifier in my leaves model :)
\nI can successfully generate a tf model with two pre-trained models, but can't successfully build one with my root model combined with a pre-trained. I can't get the weights loaded and frozen.
", "rawMarkdown": "Thanks - what I really need is the Python skill set to successfully use my root classifier in my leaves model :) \n\nI can successfully generate a tf model with two pre-trained models, but can't successfully build one with my root model combined with a pre-trained. I can't get the weights loaded and frozen.\n\n\n" }, { "id": 1186570, "postDate": "2021-02-04T22:42:19.587Z", "content": "I only filtered the records for this competition's dataset. There could be possibilities of missing other root images but my intension was also to find high noise images apart from root one and I ended up filtering it manually. However, I think suggestions made by you and others sounds promising.
", "rawMarkdown": "I only filtered the records for this competition's dataset. There could be possibilities of missing other root images but my intension was also to find high noise images apart from root one and I ended up filtering it manually. However, I think suggestions made by you and others sounds promising." }, { "id": 1186583, "postDate": "2021-02-04T22:55:58.323Z", "content": "In addition to not having the skill set to add a roots model to my leaf classifier I abandoned the effort after noting that my current best leaves model was already correctly identifying the class for all but one or two of the root pictures.
\nIf I get to a 95% model than I might return to the effort of adding roots. But chasing after noise that the model already can detect is a task for those with more talent than I.
\nHaving worked on Kaggle competitions for several years now I continue to get surprised by how powerful these tools can be - I was once again reminded - look to identify the noise that your model is missing - when I look at the images that my model is missing the noise is not as obvious as pictures of roots.
", "rawMarkdown": "In addition to not having the skill set to add a roots model to my leaf classifier I abandoned the effort after noting that my current best leaves model was already correctly identifying the class for all but one or two of the root pictures. \n\nIf I get to a 95% model than I might return to the effort of adding roots. But chasing after noise that the model already can detect is a task for those with more talent than I. \n \nHaving worked on Kaggle competitions for several years now I continue to get surprised by how powerful these tools can be - I was once again reminded - look to identify the noise that your model is missing - when I look at the images that my model is missing the noise is not as obvious as pictures of roots." }, { "id": 1192889, "postDate": "2021-02-09T10:46:09.657Z", "content": "Could you share the label of 2019 extra images? Thanks.
", "rawMarkdown": "Could you share the label of 2019 extra images? Thanks." }, { "id": 1193591, "postDate": "2021-02-09T18:22:38.267Z", "content": "You can get the data here
", "rawMarkdown": "You can get the data [here ](https://www.kaggle.com/c/cassava-disease/data)" } ] }, { "id": 1186505, "postDate": "2021-02-04T21:20:02.053Z", "content": "Interesting work.
\nMay I ask how you found these irrelevant images? manually or by algorithm?
", "rawMarkdown": "Interesting work. \n\nMay I ask how you found these irrelevant images? manually or by algorithm? ", "votes": 1, "replies": [ { "id": 1186535, "postDate": "2021-02-04T21:59:39.697Z", "content": "Thanks. It is manual work because the clustering and cosine similarity did not yield a good result for the noisy labels. This is one notebook where cosine similarity is used to get similar images.
", "rawMarkdown": "Thanks. It is manual work because the clustering and cosine similarity did not yield a good result for the noisy labels. This is one [notebook ](https://www.kaggle.com/aliabdin1/find-similar-images-with-cosine-similarity)where cosine similarity is used to get similar images.", "votes": 1 }, { "id": 1186888, "postDate": "2021-02-05T05:23:14.137Z", "content": "@saurabh2mishra was curious to know if you got your CV better and hence the LB upon removing the images.
\n@pcjimmmy any possibility for these diseases to affect the roots also by any chance ?
CV score improved but LB decreased by -0.002
", "rawMarkdown": "CV score improved but LB decreased by -0.002" }, { "id": 1187961, "postDate": "2021-02-05T19:47:40.960Z", "content": "\nYes - the root images are mostly class 1 - so that disease type is likely the only one of the three that has a distinct impact on the root.
", "rawMarkdown": "[Jaldeep](urhttps://www.kaggle.com/jaideepvalanil)\n\nYes - the root images are mostly class 1 - so that disease type is likely the only one of the three that has a distinct impact on the root. " }, { "id": 1193991, "postDate": "2021-02-10T02:44:55.277Z", "content": "Are there labels avail for extra images?
", "rawMarkdown": "Are there labels avail for extra images?" } ] }, { "id": 1191593, "postDate": "2021-02-08T14:57:57.863Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1187908, "postDate": "2021-02-05T18:50:00.867Z", "content": "Very nice compilation, thanks! 👍
", "rawMarkdown": "Very nice compilation, thanks! 👍", "votes": 1 }, { "id": 1186845, "postDate": "2021-02-05T04:45:15.360Z", "content": "Excellent work! Thanks for sharing!
", "rawMarkdown": "Excellent work! Thanks for sharing!", "votes": 1 }, { "id": 1186794, "postDate": "2021-02-05T03:32:32.610Z", "content": "Thank you for sharing
", "rawMarkdown": "Thank you for sharing", "votes": 1 } ], "comments": [ { "id": 1186556, "author_name": "Alexander Riedel", "author_url": "", "post_date": "2021-02-04T22:22:25.863000", "content": "Why are there so many root images in the dataset? I understand that some leaf images are wrong labeled but i don't get why someone would put the root images in the data?
", "votes": 3, "replies": [ { "id": 1186572, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2021-02-04T22:44:03.560000", "content": "There was what is now a very old USA television show called \"Have Gun - Will Travel\". It was a cowboy TV series about a professional gun fighter.
\n\"Have camera - will take photo\" (or selfie) is perhaps a modern version. The train and test set for 2019 did not contain roots (at least as found by my root model).
\nThis years competition appears aimed at helping weed out the noise that will occur when you place a camera in the hands of a novice. There are a number of strange images in the 2020 test set - my favorite is a picture taken of a picture of a leaf.
", "votes": 0, "replies": [] }, { "id": 1186986, "author_name": "ayu055", "author_url": "", "post_date": "2021-02-05T06:38:26.213000", "content": "I think Cassava roots can also indicate what disease the leaves are suffering from.
", "votes": 0, "replies": [] } ] }, { "id": 1187923, "author_name": "sheep", "author_url": "", "post_date": "2021-02-05T18:55:13.113000", "content": "By eyeballing the 'wrong prediction' of my model, I find that there may be much more irrelevant, wrong focused and error label.
", "votes": 1, "replies": [ { "id": 1187980, "author_name": "Saurabh Mishra", "author_url": "", "post_date": "2021-02-05T20:35:44.550000", "content": "That's true.
", "votes": 0, "replies": [] } ] }, { "id": 1186554, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2021-02-04T22:19:50.397000", "content": "The images of the roots (fruit) number closer to at least 96, so I think you missed a bunch. Not sure of the total since I worked with all 43K images that are found in the 2019 and 2020 data. (9 in the 2019 extra images) and stopped my search after the 4th version of my roots model.
\nIt's not a real difficult task to build a model to identify roots. It took me 4 iterations of a model to get the 96 identified after I started with around 25 detected manually.
\nThey are almost all class 1 (all but 8).
", "votes": 1, "replies": [ { "id": 1186558, "author_name": "Alexander Riedel", "author_url": "", "post_date": "2021-02-04T22:23:40.263000", "content": "@pcjimmmy i found this root dataset, if you need some more imeages for your root classifier: https://data.mendeley.com/datasets/gvp7vshvnh/1
", "votes": 1, "replies": [] }, { "id": 1186562, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2021-02-04T22:34:07.600000", "content": "Thanks - what I really need is the Python skill set to successfully use my root classifier in my leaves model :)
\nI can successfully generate a tf model with two pre-trained models, but can't successfully build one with my root model combined with a pre-trained. I can't get the weights loaded and frozen.
", "votes": 0, "replies": [] }, { "id": 1186570, "author_name": "Saurabh Mishra", "author_url": "", "post_date": "2021-02-04T22:42:19.587000", "content": "I only filtered the records for this competition's dataset. There could be possibilities of missing other root images but my intension was also to find high noise images apart from root one and I ended up filtering it manually. However, I think suggestions made by you and others sounds promising.
", "votes": 0, "replies": [] }, { "id": 1186583, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2021-02-04T22:55:58.323000", "content": "In addition to not having the skill set to add a roots model to my leaf classifier I abandoned the effort after noting that my current best leaves model was already correctly identifying the class for all but one or two of the root pictures.
\nIf I get to a 95% model than I might return to the effort of adding roots. But chasing after noise that the model already can detect is a task for those with more talent than I.
\nHaving worked on Kaggle competitions for several years now I continue to get surprised by how powerful these tools can be - I was once again reminded - look to identify the noise that your model is missing - when I look at the images that my model is missing the noise is not as obvious as pictures of roots.
", "votes": 0, "replies": [] }, { "id": 1192889, "author_name": "komakiyyy", "author_url": "", "post_date": "2021-02-09T10:46:09.657000", "content": "Could you share the label of 2019 extra images? Thanks.
", "votes": 0, "replies": [] }, { "id": 1193591, "author_name": "Saurabh Mishra", "author_url": "", "post_date": "2021-02-09T18:22:38.267000", "content": "You can get the data here
", "votes": 0, "replies": [] } ] }, { "id": 1186505, "author_name": "LeoF", "author_url": "", "post_date": "2021-02-04T21:20:02.053000", "content": "Interesting work.
\nMay I ask how you found these irrelevant images? manually or by algorithm?
", "votes": 1, "replies": [ { "id": 1186535, "author_name": "Saurabh Mishra", "author_url": "", "post_date": "2021-02-04T21:59:39.697000", "content": "Thanks. It is manual work because the clustering and cosine similarity did not yield a good result for the noisy labels. This is one notebook where cosine similarity is used to get similar images.
", "votes": 1, "replies": [] }, { "id": 1186888, "author_name": "Jaideep", "author_url": "", "post_date": "2021-02-05T05:23:14.137000", "content": "@saurabh2mishra was curious to know if you got your CV better and hence the LB upon removing the images.
\n@pcjimmmy any possibility for these diseases to affect the roots also by any chance ?
CV score improved but LB decreased by -0.002
", "votes": 0, "replies": [] }, { "id": 1187961, "author_name": "PC Jimmmy", "author_url": "", "post_date": "2021-02-05T19:47:40.960000", "content": "\nYes - the root images are mostly class 1 - so that disease type is likely the only one of the three that has a distinct impact on the root.
", "votes": 0, "replies": [] }, { "id": 1193991, "author_name": "Jaideep", "author_url": "", "post_date": "2021-02-10T02:44:55.277000", "content": "Are there labels avail for extra images?
", "votes": 0, "replies": [] } ] }, { "id": 1191593, "author_name": "", "author_url": "", "post_date": "2021-02-08T14:57:57.863000", "content": "", "votes": 0, "replies": [] }, { "id": 1187908, "author_name": "Sinan Calisir", "author_url": "", "post_date": "2021-02-05T18:50:00.867000", "content": "Very nice compilation, thanks! 👍
", "votes": 1, "replies": [] }, { "id": 1186845, "author_name": "zlannn", "author_url": "", "post_date": "2021-02-05T04:45:15.360000", "content": "Excellent work! Thanks for sharing!
", "votes": 1, "replies": [] }, { "id": 1186794, "author_name": "cswwp", "author_url": "", "post_date": "2021-02-05T03:32:32.610000", "content": "Thank you for sharing
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1186462": "During the training of the model, I tried to identify all irrelevant images that shouldn't be present in the dataset. These dataset ids have been categorized into 2 parts. The first category is irrelevant images (all images are cassava fruits not leaves) and the second one is noisy images (all images are having more background area rather than leaves). However, removing these datasets has been decreased the accuracy of the model with the same model but I believe these are not relevant for feeding the model to just make them learn wrong images. \n\n***irrelevant image ids*** = [274726002, 9224019, 159654644, 199112616, 226533928, 262902341, 269713568, 274726002, 384390206, 390601409, 421035788, 457405364, 600736721, 580111608,\n 616718743, 695438825, 723564013, 826231979, 847847826, 927165736, 1004389140, \n 1008244905, 1338159402, 1339403533, 1359893940, 1366430957, 1689510013, 9224019,\n 4269208386, 4239074071, 3810809174, 3652033201, 3609350672, 3609986814, \n 3477169212, 3435954655, 3425850136, 3251960666, 3252232501, 3199643560, \n 3126296051, 3040241097, 2981404650, 2925605732, 2839068946, 2698282165,\n 2604713994, 2415837573, 2382642453, 2321669192, 2320471703, 2278166989,\n 2276509518, 2262263316, 2182500020, 2139839273, 2084868828, 1848686439,\n 1689510013, 1359893940]\n\n***noisy image ids*** = [ 410880003, 411955232, 501215014, 549854027, 554488826, \n 724195836, 744383303, 888983519, 1096438409, 1130568730, \n 1709404074, 1770746162, 4280523848, 3530560257, 3421208425, \n 3321193739, 3086663390, 3045134829, 1862072615]\n\n**Notes**: I tried to upload some sample images but seems Kaggle is not allowing me to do so. Showing the message *We have disabled uploading forum attachments for the time being. Please use an alternative host for your file, and link to it from your forum post.*", "1186556": "Why are there so many root images in the dataset? I understand that some leaf images are wrong labeled but i don't get why someone would put the root images in the data?", "1187923": "By eyeballing the 'wrong prediction' of my model, I find that there may be much more irrelevant, wrong focused and error label.", "1186554": "The images of the roots (fruit) number closer to at least 96, so I think you missed a bunch. Not sure of the total since I worked with all 43K images that are found in the 2019 and 2020 data. (9 in the 2019 extra images) and stopped my search after the 4th version of my roots model.\n\nIt's not a real difficult task to build a model to identify roots. It took me 4 iterations of a model to get the 96 identified after I started with around 25 detected manually.\n\nThey are almost all class 1 (all but 8).", "1186505": "Interesting work. \n\nMay I ask how you found these irrelevant images? manually or by algorithm? ", "1191593": "", "1187908": "Very nice compilation, thanks! 👍", "1186845": "Excellent work! Thanks for sharing!", "1186794": "Thank you for sharing" } }