![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2621907%2F1f06ed4836eee42f0414c1891a159882%2Fapproach.JPG?generation=1595011724912690&alt=media\")
W -> weights\nI tried only 256 and 384. you can try with other image sizes
", "messages": [ { "id": 933521, "postDate": "2020-07-17T18:40:34.167Z", "content": "
W -> weights\nI tried only 256 and 384. you can try with other image sizes
", "rawMarkdown": "\n\n\nW -> weights\nI tried only 256 and 384. you can try with other image sizes", "votes": 13 }, { "id": 933689, "postDate": "2020-07-17T23:05:58.183Z", "content": "In my opinion, it should be better for you and for us to share this when the competition ends.
", "rawMarkdown": "In my opinion, it should be better for you and for us to share this when the competition ends.", "votes": 3 }, { "id": 954670, "postDate": "2020-08-02T01:47:04.797Z", "content": "Great job Ashok
", "rawMarkdown": "Great job Ashok", "votes": 1, "replies": [ { "id": 954741, "postDate": "2020-08-02T03:55:03.643Z", "content": "Thanks Chris
", "rawMarkdown": "Thanks Chris", "votes": 1 } ] }, { "id": 933532, "postDate": "2020-07-17T18:53:33.803Z", "content": "you can't explain why you choose such w1,w2 weights and why those should work in private lb,so you should never put trust on such kind of models for private lb,these are overfitting,nothing else, public lb is a beautiful illusion :)
", "rawMarkdown": "you can't explain why you choose such w1,w2 weights and why those should work in private lb,so you should never put trust on such kind of models for private lb,these are overfitting,nothing else, public lb is a beautiful illusion :)", "votes": 1, "replies": [ { "id": 933550, "postDate": "2020-07-17T19:13:47.263Z", "content": "Weights are random but not pure random. i gave higher weights randomly to only image-model.\nBecause metadata is adding some noise to the model both in training and LB
\n\nMy mean score is 94.9\nMy random weighted score is 95.2
\n\ni dont think it is overfitting
", "rawMarkdown": "Weights are random but not pure random. i gave higher weights randomly to only image-model.\nBecause metadata is adding some noise to the model both in training and LB\n\nMy mean score is 94.9\nMy random weighted score is 95.2\n\ni dont think it is overfitting" }, { "id": 933862, "postDate": "2020-07-18T04:53:47.450Z", "content": "are weights are on the basis of CV? (combining these models increases your LB with an increase in CV )
", "rawMarkdown": "are weights are on the basis of CV? (combining these models increases your LB with an increase in CV )" }, { "id": 934281, "postDate": "2020-07-18T11:00:37.127Z", "content": "@coolcoder22 Yes. weights i used is totally based on CV. These weights increased my CV and LB too.\n@jacekpoplawski I think. It is not over fitting because my cv*weights directly preportional to submission*weights. if my CV increases my LB increases.
", "rawMarkdown": "@coolcoder22 Yes. weights i used is totally based on CV. These weights increased my CV and LB too.\n@jacekpoplawski I think. It is not over fitting because my cv*weights directly preportional to submission*weights. if my CV increases my LB increases. " } ] }, { "id": 933796, "postDate": "2020-07-18T03:02:06.587Z", "content": "What is your mean CV AUC? For Image only model and Image+Tabular model?
", "rawMarkdown": "What is your mean CV AUC? For Image only model and Image+Tabular model? ", "votes": -1, "replies": [ { "id": 934286, "postDate": "2020-07-18T11:01:45.910Z", "content": "You can get the info in above image
", "rawMarkdown": "You can get the info in above image", "votes": -1 } ] }, { "id": 954318, "postDate": "2020-08-01T16:18:00.207Z", "content": "CV or LB score?
", "rawMarkdown": "CV or LB score?", "replies": [ { "id": 954342, "postDate": "2020-08-01T16:40:01.720Z", "content": "@projdev Mean score - 94.9 and LB - 95.2
", "rawMarkdown": "@projdev Mean score - 94.9 and LB - 95.2" }, { "id": 954347, "postDate": "2020-08-01T16:42:43.787Z", "content": "that's a very good score. I'm having a hard time to achieve 0.94 CV
", "rawMarkdown": "that's a very good score. I'm having a hard time to achieve 0.94 CV" }, { "id": 954360, "postDate": "2020-08-01T17:02:27.003Z", "content": "Things i tried.
\n\nHope this helps you
", "rawMarkdown": "Things i tried.\n\n- Tried external data provided by chris\n- Trained my model on 256 to 768 image sizes (5 folds)\n- Tried EfficientNet B3 to B7 -- with image and image+metadata\n- focal loss gave me better result than binary cross entropy loss\n- layer and learning rate tweaking \n- ensembling is the key \n- finally pseudo labeling\n\nHope this helps you ", "votes": 2 } ] }, { "id": 934084, "postDate": "2020-07-18T08:14:04.047Z", "content": "Do you use any external data ?
", "rawMarkdown": "Do you use any external data ?", "replies": [ { "id": 934283, "postDate": "2020-07-18T11:01:12.460Z", "content": "Yes i used
", "rawMarkdown": "Yes i used " }, { "id": 952518, "postDate": "2020-07-31T02:20:29.220Z", "content": "Could I ask which external data did you use？ or use some kaggle dataset created by someone else？
", "rawMarkdown": "Could I ask which external data did you use？ or use some kaggle dataset created by someone else？" }, { "id": 954210, "postDate": "2020-08-01T13:52:42.043Z", "content": "@seefun I was referring to the same external data provided by chris
", "rawMarkdown": "@seefun I was referring to the same external data provided by chris " } ] }, { "id": 933675, "postDate": "2020-07-17T22:19:25.607Z", "content": "Can you explain Ensemble? You train model with 6 weights?
", "rawMarkdown": "Can you explain Ensemble? You train model with 6 weights?", "replies": [ { "id": 933682, "postDate": "2020-07-17T22:39:33.143Z", "content": "I trained my model with B7 and B6. i got 4 submission files from 4 different models. \nI assigned weights to different submission files. which is purely based on training and LB score.\nIn my case model with only image data gave me good training and LB score. So, i assigned more weight to submission files which are from image model(only image data).
\n\n```\nexample:
\n\n(image_metadata_submission)*0.8+(only_images_submission)*1.2\n```\nHope this is clear
", "rawMarkdown": "I trained my model with B7 and B6. i got 4 submission files from 4 different models. \nI assigned weights to different submission files. which is purely based on training and LB score.\nIn my case model with only image data gave me good training and LB score. So, i assigned more weight to submission files which are from image model(only image data). \n\n```\nexample:\n\n(image_metadata_submission)*0.8+(only_images_submission)*1.2\n```\nHope this is clear", "votes": 3 }, { "id": 933687, "postDate": "2020-07-17T22:52:15.303Z", "content": "do you mean you don't train the ensemble model but you assigned weights manually?
", "rawMarkdown": "do you mean you don't train the ensemble model but you assigned weights manually?" }, { "id": 933696, "postDate": "2020-07-17T23:18:25.343Z", "content": "yes.
", "rawMarkdown": "yes.", "votes": -1 }, { "id": 933707, "postDate": "2020-07-17T23:54:11.180Z", "content": "and why do you think it is not overfitting?
", "rawMarkdown": "and why do you think it is not overfitting?" } ] }, { "id": 933885, "postDate": "2020-07-18T05:18:41.373Z", "rawMarkdown": "", "votes": -2, "isDeleted": true, "replies": [ { "id": 934301, "postDate": "2020-07-18T11:20:26.590Z", "content": "Hi, I am not posting any solution. My approach will not change the structure of LB score. Because i have not revealed my model structure, my hyper parameters, my folding logic and mostly my submission files.My approach is not something new. everybody can figure it out by following discussions properly.
", "rawMarkdown": "Hi, I am not posting any solution. My approach will not change the structure of LB score. Because i have not revealed my model structure, my hyper parameters, my folding logic and mostly my submission files.My approach is not something new. everybody can figure it out by following discussions properly.", "votes": 4 }, { "id": 934470, "postDate": "2020-07-18T13:17:37.897Z", "content": "I believe it is far more better than sharing the notebook with ~0.95 score on LB, at least someone will try to understand and implement this if I have to consider your statement.
\n\nI am in this competition from day one and I have seen hundreds of people submitting the submission files shared here.
\n\nAlso, he stated that he got 0.93+ with 256x256 only, while I could not get this until now(although its my first competition), so there is a lot more there which he has not shared.
\n\nAgain, thanks @ashoksrinivas It is really of great help! :)
", "rawMarkdown": "I believe it is far more better than sharing the notebook with ~0.95 score on LB, at least someone will try to understand and implement this if I have to consider your statement. \n\nI am in this competition from day one and I have seen hundreds of people submitting the submission files shared here.\n\nAlso, he stated that he got 0.93+ with 256x256 only, while I could not get this until now(although its my first competition), so there is a lot more there which he has not shared.\n\nAgain, thanks @ashoksrinivas It is really of great help! :)", "votes": 2 }, { "id": 934715, "postDate": "2020-07-18T17:05:04.253Z", "content": "Thanks buddy :)\nKeep trying. Dont give up
", "rawMarkdown": "Thanks buddy :)\nKeep trying. Dont give up" } ] }, { "id": 933749, "postDate": "2020-07-18T01:42:19.773Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 933689, "author_name": "Seifeddine Fezzani", "author_url": "", "post_date": "2020-07-17T23:05:58.183000", "content": "In my opinion, it should be better for you and for us to share this when the competition ends.
", "votes": 3, "replies": [] }, { "id": 954670, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-08-02T01:47:04.797000", "content": "Great job Ashok
", "votes": 1, "replies": [ { "id": 954741, "author_name": "ashok", "author_url": "", "post_date": "2020-08-02T03:55:03.643000", "content": "Thanks Chris
", "votes": 1, "replies": [] } ] }, { "id": 933532, "author_name": "Mobassir", "author_url": "", "post_date": "2020-07-17T18:53:33.803000", "content": "you can't explain why you choose such w1,w2 weights and why those should work in private lb,so you should never put trust on such kind of models for private lb,these are overfitting,nothing else, public lb is a beautiful illusion :)
", "votes": 1, "replies": [ { "id": 933550, "author_name": "ashok", "author_url": "", "post_date": "2020-07-17T19:13:47.263000", "content": "Weights are random but not pure random. i gave higher weights randomly to only image-model.\nBecause metadata is adding some noise to the model both in training and LB
\n\nMy mean score is 94.9\nMy random weighted score is 95.2
\n\ni dont think it is overfitting
", "votes": 0, "replies": [] }, { "id": 933862, "author_name": "Hitesh Gorana", "author_url": "", "post_date": "2020-07-18T04:53:47.450000", "content": "are weights are on the basis of CV? (combining these models increases your LB with an increase in CV )
", "votes": 0, "replies": [] }, { "id": 934281, "author_name": "ashok", "author_url": "", "post_date": "2020-07-18T11:00:37.127000", "content": "@coolcoder22 Yes. weights i used is totally based on CV. These weights increased my CV and LB too.\n@jacekpoplawski I think. It is not over fitting because my cv*weights directly preportional to submission*weights. if my CV increases my LB increases.
", "votes": 0, "replies": [] } ] }, { "id": 933796, "author_name": "Abhishek Bhat", "author_url": "", "post_date": "2020-07-18T03:02:06.587000", "content": "What is your mean CV AUC? For Image only model and Image+Tabular model?
", "votes": -1, "replies": [ { "id": 934286, "author_name": "ashok", "author_url": "", "post_date": "2020-07-18T11:01:45.910000", "content": "You can get the info in above image
", "votes": -1, "replies": [] } ] }, { "id": 954318, "author_name": "FGPC", "author_url": "", "post_date": "2020-08-01T16:18:00.207000", "content": "CV or LB score?
", "votes": 0, "replies": [ { "id": 954342, "author_name": "ashok", "author_url": "", "post_date": "2020-08-01T16:40:01.720000", "content": "@projdev Mean score - 94.9 and LB - 95.2
", "votes": 0, "replies": [] }, { "id": 954347, "author_name": "FGPC", "author_url": "", "post_date": "2020-08-01T16:42:43.787000", "content": "that's a very good score. I'm having a hard time to achieve 0.94 CV
", "votes": 0, "replies": [] }, { "id": 954360, "author_name": "ashok", "author_url": "", "post_date": "2020-08-01T17:02:27.003000", "content": "Things i tried.
\n\nHope this helps you
", "votes": 2, "replies": [] } ] }, { "id": 934084, "author_name": "Volodymyr", "author_url": "", "post_date": "2020-07-18T08:14:04.047000", "content": "Do you use any external data ?
", "votes": 0, "replies": [ { "id": 934283, "author_name": "ashok", "author_url": "", "post_date": "2020-07-18T11:01:12.460000", "content": "Yes i used
", "votes": 0, "replies": [] }, { "id": 952518, "author_name": "seefun", "author_url": "", "post_date": "2020-07-31T02:20:29.220000", "content": "Could I ask which external data did you use？ or use some kaggle dataset created by someone else？
", "votes": 0, "replies": [] }, { "id": 954210, "author_name": "ashok", "author_url": "", "post_date": "2020-08-01T13:52:42.043000", "content": "@seefun I was referring to the same external data provided by chris
", "votes": 0, "replies": [] } ] }, { "id": 933675, "author_name": "Jacek Poplawski", "author_url": "", "post_date": "2020-07-17T22:19:25.607000", "content": "Can you explain Ensemble? You train model with 6 weights?
", "votes": 0, "replies": [ { "id": 933682, "author_name": "ashok", "author_url": "", "post_date": "2020-07-17T22:39:33.143000", "content": "I trained my model with B7 and B6. i got 4 submission files from 4 different models. \nI assigned weights to different submission files. which is purely based on training and LB score.\nIn my case model with only image data gave me good training and LB score. So, i assigned more weight to submission files which are from image model(only image data).
\n\n```\nexample:
\n\n(image_metadata_submission)*0.8+(only_images_submission)*1.2\n```\nHope this is clear
", "votes": 3, "replies": [] }, { "id": 933687, "author_name": "Jacek Poplawski", "author_url": "", "post_date": "2020-07-17T22:52:15.303000", "content": "do you mean you don't train the ensemble model but you assigned weights manually?
", "votes": 0, "replies": [] }, { "id": 933696, "author_name": "ashok", "author_url": "", "post_date": "2020-07-17T23:18:25.343000", "content": "yes.
", "votes": -1, "replies": [] }, { "id": 933707, "author_name": "Jacek Poplawski", "author_url": "", "post_date": "2020-07-17T23:54:11.180000", "content": "and why do you think it is not overfitting?
", "votes": 0, "replies": [] } ] }, { "id": 933885, "author_name": "", "author_url": "", "post_date": "2020-07-18T05:18:41.373000", "content": "", "votes": -2, "replies": [ { "id": 934301, "author_name": "ashok", "author_url": "", "post_date": "2020-07-18T11:20:26.590000", "content": "Hi, I am not posting any solution. My approach will not change the structure of LB score. Because i have not revealed my model structure, my hyper parameters, my folding logic and mostly my submission files.My approach is not something new. everybody can figure it out by following discussions properly.
", "votes": 4, "replies": [] }, { "id": 934470, "author_name": "Gajendra Saraswat", "author_url": "", "post_date": "2020-07-18T13:17:37.897000", "content": "I believe it is far more better than sharing the notebook with ~0.95 score on LB, at least someone will try to understand and implement this if I have to consider your statement.
\n\nI am in this competition from day one and I have seen hundreds of people submitting the submission files shared here.
\n\nAlso, he stated that he got 0.93+ with 256x256 only, while I could not get this until now(although its my first competition), so there is a lot more there which he has not shared.
\n\nAgain, thanks @ashoksrinivas It is really of great help! :)
", "votes": 2, "replies": [] }, { "id": 934715, "author_name": "ashok", "author_url": "", "post_date": "2020-07-18T17:05:04.253000", "content": "Thanks buddy :)\nKeep trying. Dont give up
", "votes": 0, "replies": [] } ] }, { "id": 933749, "author_name": "", "author_url": "", "post_date": "2020-07-18T01:42:19.773000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "933521": "\n\n\nW -> weights\nI tried only 256 and 384. you can try with other image sizes", "933689": "In my opinion, it should be better for you and for us to share this when the competition ends.", "954670": "Great job Ashok", "933532": "you can't explain why you choose such w1,w2 weights and why those should work in private lb,so you should never put trust on such kind of models for private lb,these are overfitting,nothing else, public lb is a beautiful illusion :)", "933796": "What is your mean CV AUC? For Image only model and Image+Tabular model? ", "954318": "CV or LB score?", "934084": "Do you use any external data ?", "933675": "Can you explain Ensemble? You train model with 6 weights?", "933885": "", "933749": "" } }