At first, I want to say thanks for all of my teammates :)
\n\nmodel1 takes 3 texts as input, and has 3berts in it\nBert1.\ninput is question_title + question_body and predict only columns that are relevant to question (first 21 of columns)\nBert2.\ninput is question_title + answer and predict only columns that are relevant to answer\n(last 9 of columns)\nBert 3.\ninput is question_body + answer and predict all columns\nand have 1 linear layer which input is concat(bert1_out, bert2_out, bert3_out)\nand calculate bce for\nloss1 for bert1 prediction & columns[:21],\nloss2 for bert2 prediction & columns[-9:],\nloss3 for bert3 prediction & all columns,\nloss4 for last linear layer prediction & all columns\n&backward each loss.
\n\nmodel2 is just xlnet version of it.
\n\n-> Spanish -> English argumentation
\n\nFlexible rating module for encoding 2 texts\n(ex. if len(text1) < max len1 & Len(text2) > max len2, max len2 = max len2 + (max_len1 - len(text1)))
\n\nBaba shared the idea here โ\nhttps://www.kaggle.com/c/google-quest-challenge/discussion/129831
\n\nThis competition was really fun, thanks all and let's compete at other competition again !!!
", "comments": [ { "id": 742253, "author_name": "Aditya Soni", "author_url": "", "post_date": "2020-02-11T07:27:11.430000", "content": "3 Bert's in one model with shared weights? Like,\n\nself.bert_1 = BertModel(...)\nself.bert_2 = BertModel(...)\nself.bert_3 = BertModel(...)\n
Thanks!
", "votes": 5, "replies": [ { "id": 742256, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T07:29:11.543000", "content": "Yes, right
", "votes": 3, "replies": [] }, { "id": 744726, "author_name": "Massoud Hosseinali", "author_url": "", "post_date": "2020-02-13T04:28:23.090000", "content": "Are you sure if you assign 3 BertModel(...)s like that they will share the same weights?
Good question,\nWe tried sharing bert weights model too,\nbut this non-sharing model worked better in this case
", "votes": 1, "replies": [] }, { "id": 744813, "author_name": "Massoud Hosseinali", "author_url": "", "post_date": "2020-02-13T06:56:50.210000", "content": "Thank you for clarification.
", "votes": 0, "replies": [] } ] }, { "id": 745613, "author_name": "Seam Chung", "author_url": "", "post_date": "2020-02-14T01:59:46.403000", "content": "Congrats @ryunosukeishizaki and team!!
", "votes": 1, "replies": [ { "id": 745715, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-14T05:30:37.707000", "content": "Thanks!!
", "votes": 0, "replies": [] } ] }, { "id": 744554, "author_name": "corochann", "author_url": "", "post_date": "2020-02-12T23:37:21.420000", "content": "Congratulations and thanks for sharing summary.\nAugmentation by changing language is interesting! What kind of language translation is used? by google translation or somthing?
", "votes": 1, "replies": [ { "id": 744581, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-13T00:12:42.327000", "content": "Thank you !!!\nYou can both pip install google trans/textblob,\nthis time I used TextBlob!
", "votes": 0, "replies": [] } ] }, { "id": 742843, "author_name": "Aditya Soni", "author_url": "", "post_date": "2020-02-11T15:31:06.213000", "content": "Hey, I have another question, after carefully reading for 3rd time, I see you said \"backward each loss\", so I presume, you had 4 output heads and \"loss_[1-4]\" for each and you guys did .backward() on each or was it by summing them up and a backward? Isn't both the same ๐ค?\nThanks!
\n\nnvm, i got the answer! Now i understand the solution (hopefully) !! (By Baba Firas)
\n\n>your solution is so similar to ours, The 2 differences is that we used 3 transformers instead of 2 + we backpropagate in every transformer using 4 losses separately ( loss1(answerstarget, output from bertanswers) , loss2(questiontargets, output from bertquestion), loss3(targets, output from bert3 which combines answer+question), loss4 is final loss after concatenating all logits from different transofmers))
", "votes": 1, "replies": [] }, { "id": 742658, "author_name": "Gerwyn", "author_url": "", "post_date": "2020-02-11T12:59:10.127000", "content": "\n\nCongrats on your results :)
\n\nMay I ask how are you doing the translations? Is there a package that is able to translate efficiently that you are using?
", "votes": 1, "replies": [ { "id": 743247, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T23:30:24.013000", "content": "Thanks,\nIt's TextBlob\nYou can pip install it
", "votes": 1, "replies": [] } ] }, { "id": 742577, "author_name": "OmarSalman", "author_url": "", "post_date": "2020-02-11T11:44:41.273000", "content": "Congratulations! Superb solution.
", "votes": 1, "replies": [ { "id": 742602, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T12:06:22.667000", "content": "Thanks :)
", "votes": 0, "replies": [] } ] }, { "id": 742519, "author_name": "Serigne ", "author_url": "", "post_date": "2020-02-11T10:38:42.027000", "content": "Congrats @ryunosukeishizaki
\n\nMy bert trained on question and answer separately is pretty similar to yours :
\n\nquestiontitle + questionbody as input and predict on questions related targets (21 columns)
questiontitle + questionbody + answer as input and predict on answers related targets (9 columns)
Cool to know inverse translation augmentation was helpful for you. My teammate @xiaojiu1414 gave it a try too (Spanish -> English)
", "votes": 1, "replies": [ { "id": 742606, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T12:07:50.980000", "content": "Thanks!\nthis time wasn't but let's meet at next competition!๐
", "votes": 0, "replies": [] }, { "id": 742840, "author_name": "Serigne ", "author_url": "", "post_date": "2020-02-11T15:24:07.627000", "content": "Yeah ! Would be great !!
", "votes": 1, "replies": [] } ] }, { "id": 742373, "author_name": "", "author_url": "", "post_date": "2020-02-11T09:10:47.303000", "content": "Congratulations! Seems augmentation worked quite well for you. Frankly I was sort of afraid to augment the data myself thinking I will spoil the results only. Thank you for sharing your approach with us! :3
", "votes": 1, "replies": [ { "id": 742429, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T09:38:09.223000", "content": "Thanks :)\nYes, then I did augmentation in 20% when dataload to make it works
", "votes": 0, "replies": [] } ] }, { "id": 742366, "author_name": "0xFunky", "author_url": "", "post_date": "2020-02-11T09:03:16.210000", "content": "Thanks for your sharing!\nI have tired this kind of model structure, can I ask what score of LB without post process you got?\nMany thanks!
", "votes": 1, "replies": [ { "id": 742418, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T09:35:37.487000", "content": "I don't remember the score without pp precisely,\nbut it was like LB0.405 -> 0.445
", "votes": 1, "replies": [] } ] }, { "id": 742331, "author_name": "Gursewak Dhiman", "author_url": "", "post_date": "2020-02-11T08:37:12.340000", "content": "May I ask which one was better xlnet or bert because my strategy was closely related to yours but score wasn't :p\nI mean you did augmentation also and post processing as well.\nhttps://www.kaggle.com/gskdhiman/inference-kernel-for-bert-memory-intensive\nIf you can give any suggestions, it will be helpful\nAnd congratulations for the gold.
", "votes": 1, "replies": [ { "id": 742338, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T08:42:57.147000", "content": "bert was better,\nwe had better score after tuning lr
", "votes": 2, "replies": [] }, { "id": 742350, "author_name": "Gursewak Dhiman", "author_url": "", "post_date": "2020-02-11T08:52:18.503000", "content": "Thanks for your quick reply
", "votes": 1, "replies": [] } ] }, { "id": 742325, "author_name": "Zhuang Jia", "author_url": "", "post_date": "2020-02-11T08:28:49.970000", "content": "Congratulations! Thanks for your sharing! I still have a question, how did you backward each loss? Can you share us your model definition and training strategy. Thanks very much!
the code is not only mine,\nthen if it's possible after talking with teammates I can
", "votes": 2, "replies": [] }, { "id": 742428, "author_name": "Zhuang Jia", "author_url": "", "post_date": "2020-02-11T09:37:43.503000", "content": "Thanks very much!
", "votes": 1, "replies": [] } ] }, { "id": 742299, "author_name": "Ailurophile", "author_url": "", "post_date": "2020-02-11T08:15:18.030000", "content": "Congratulations @ryunosukeishizaki and team!\nThanks for Sharing your Approach!!
", "votes": 1, "replies": [ { "id": 742340, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T08:45:38.513000", "content": "Thanks :)
", "votes": 1, "replies": [] } ] }, { "id": 742282, "author_name": "Mourad", "author_url": "", "post_date": "2020-02-11T07:53:03.167000", "content": "Congratulations @ryunosukeishizaki and team!
\n\nMay I ask how did you go about augmentation?
\n\nAnd how much of an effect did it have on your score?
", "votes": 1, "replies": [ { "id": 742283, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T07:57:09.540000", "content": "I referred this โ\nhttps://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/48038\nIt was just about +0.004-5 on LB
", "votes": 1, "replies": [] } ] }, { "id": 742272, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-02-11T07:43:45.213000", "content": "Congrats @ryunosukeishizaki and team.
", "votes": 1, "replies": [ { "id": 742273, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T07:44:45.577000", "content": "Thank you!!
", "votes": 1, "replies": [] } ] }, { "id": 742257, "author_name": "Ryunosuke Ishizaki", "author_url": "", "post_date": "2020-02-11T07:29:53.837000", "content": "*Validations are both MultiStratifiedKFold
", "votes": 2, "replies": [] }, { "id": 866356, "author_name": "ๅๅฐ็็งปๅจๅๅ ก", "author_url": "", "post_date": "2020-05-29T10:21:16.320000", "content": "Hey, mate ,sorry for my question at this time.\nHow do you make emsemble in model1 ? There should be 3 models , 4 output here. Do you just make a average of them ?
", "votes": 0, "replies": [] } ], "comment_count": 35, "messages": [ { "id": 742244, "postDate": "2020-02-11T07:20:18.787Z", "content": "At first, I want to say thanks for all of my teammates :)
\n\nmodel1 takes 3 texts as input, and has 3berts in it\nBert1.\ninput is question_title + question_body and predict only columns that are relevant to question (first 21 of columns)\nBert2.\ninput is question_title + answer and predict only columns that are relevant to answer\n(last 9 of columns)\nBert 3.\ninput is question_body + answer and predict all columns\nand have 1 linear layer which input is concat(bert1_out, bert2_out, bert3_out)\nand calculate bce for\nloss1 for bert1 prediction & columns[:21],\nloss2 for bert2 prediction & columns[-9:],\nloss3 for bert3 prediction & all columns,\nloss4 for last linear layer prediction & all columns\n&backward each loss.
\n\nmodel2 is just xlnet version of it.
\n\n-> Spanish -> English argumentation
\n\nFlexible rating module for encoding 2 texts\n(ex. if len(text1) < max len1 & Len(text2) > max len2, max len2 = max len2 + (max_len1 - len(text1)))
\n\nBaba shared the idea here โ\nhttps://www.kaggle.com/c/google-quest-challenge/discussion/129831
\n\nThis competition was really fun, thanks all and let's compete at other competition again !!!
", "rawMarkdown": "At first, I want to say thanks for all of my teammates :)\n\n## Our final submission is an ensemble of 2 models\n\n### About Model\n\nmodel1 takes 3 texts as input, and has 3berts in it\nBert1.\ninput is question_title + question_body and predict only columns that are relevant to question (first 21 of columns)\nBert2.\ninput is question_title + answer and predict only columns that are relevant to answer\n(last 9 of columns)\nBert 3.\ninput is question_body + answer and predict all columns\nand have 1 linear layer which input is concat(bert1_out, bert2_out, bert3_out)\nand calculate bce for\nloss1 for bert1 prediction & columns[:21],\nloss2 for bert2 prediction & columns[-9:],\nloss3 for bert3 prediction & all columns,\nloss4 for last linear layer prediction & all columns\n&backward each loss.\n\nmodel2 is just xlnet version of it.\n\n### About Training\n\n-> Spanish -> English argumentation\n\nFlexible rating module for encoding 2 texts\n(ex. if len(text1) < max len1 & Len(text2) > max len2, max len2 = max len2 + (max_len1 - len(text1)))\n\n### About Post Processing\n\nBaba shared the idea here โ\nhttps://www.kaggle.com/c/google-quest-challenge/discussion/129831\n\nThis competition was really fun, thanks all and let's compete at other competition again !!!", "votes": 65 }, { "id": 742253, "postDate": "2020-02-11T07:27:11.430Z", "content": "3 Bert's in one model with shared weights? Like,\n\nself.bert_1 = BertModel(...)\nself.bert_2 = BertModel(...)\nself.bert_3 = BertModel(...)\n
Thanks!
", "rawMarkdown": "3 Bert's in one model with shared weights? Like,\n```\nself.bert_1 = BertModel(...)\nself.bert_2 = BertModel(...)\nself.bert_3 = BertModel(...)\n```\n\nThanks!", "votes": 5, "replies": [ { "id": 742256, "postDate": "2020-02-11T07:29:11.543Z", "content": "Yes, right
", "rawMarkdown": "Yes, right", "votes": 3 }, { "id": 744726, "postDate": "2020-02-13T04:28:23.090Z", "content": "Are you sure if you assign 3 BertModel(...)s like that they will share the same weights?
Good question,\nWe tried sharing bert weights model too,\nbut this non-sharing model worked better in this case
", "rawMarkdown": "Good question,\nWe tried sharing bert weights model too,\nbut this non-sharing model worked better in this case", "votes": 1 }, { "id": 744813, "postDate": "2020-02-13T06:56:50.210Z", "content": "Thank you for clarification.
", "rawMarkdown": "Thank you for clarification." } ] }, { "id": 745613, "postDate": "2020-02-14T01:59:46.403Z", "content": "Congrats @ryunosukeishizaki and team!!
", "rawMarkdown": "Congrats @ryunosukeishizaki and team!!", "votes": 1, "replies": [ { "id": 745715, "postDate": "2020-02-14T05:30:37.707Z", "content": "Thanks!!
", "rawMarkdown": "Thanks!!" } ] }, { "id": 744554, "postDate": "2020-02-12T23:37:21.420Z", "content": "Congratulations and thanks for sharing summary.\nAugmentation by changing language is interesting! What kind of language translation is used? by google translation or somthing?
", "rawMarkdown": "Congratulations and thanks for sharing summary.\nAugmentation by changing language is interesting! What kind of language translation is used? by google translation or somthing?", "votes": 1, "replies": [ { "id": 744581, "postDate": "2020-02-13T00:12:42.327Z", "content": "Thank you !!!\nYou can both pip install google trans/textblob,\nthis time I used TextBlob!
", "rawMarkdown": "Thank you !!!\nYou can both pip install google trans/textblob,\nthis time I used TextBlob!" } ] }, { "id": 742843, "postDate": "2020-02-11T15:31:06.213Z", "content": "Hey, I have another question, after carefully reading for 3rd time, I see you said \"backward each loss\", so I presume, you had 4 output heads and \"loss_[1-4]\" for each and you guys did .backward() on each or was it by summing them up and a backward? Isn't both the same ๐ค?\nThanks!
\n\nnvm, i got the answer! Now i understand the solution (hopefully) !! (By Baba Firas)
\n\n>your solution is so similar to ours, The 2 differences is that we used 3 transformers instead of 2 + we backpropagate in every transformer using 4 losses separately ( loss1(answerstarget, output from bertanswers) , loss2(questiontargets, output from bertquestion), loss3(targets, output from bert3 which combines answer+question), loss4 is final loss after concatenating all logits from different transofmers))
", "rawMarkdown": "Hey, I have another question, after carefully reading for 3rd time, I see you said \"backward each loss\", so I presume, you had 4 output heads and \"loss_[1-4]\" for each and you guys did .backward() on each or was it by summing them up and a backward? Isn't both the same ๐ค?\nThanks!\n\nnvm, i got the answer! Now i understand the solution (hopefully) !! (By Baba Firas)\n\n>your solution is so similar to ours, The 2 differences is that we used 3 transformers instead of 2 + we backpropagate in every transformer using 4 losses separately ( loss1(answerstarget, output from bertanswers) , loss2(questiontargets, output from bertquestion), loss3(targets, output from bert3 which combines answer+question), loss4 is final loss after concatenating all logits from different transofmers))", "votes": 1 }, { "id": 742658, "postDate": "2020-02-11T12:59:10.127Z", "content": "\n\nCongrats on your results :)
\n\nMay I ask how are you doing the translations? Is there a package that is able to translate efficiently that you are using?
", "rawMarkdown": "Hi @ryunosukeishizaki \n\nCongrats on your results :)\n\nMay I ask how are you doing the translations? Is there a package that is able to translate efficiently that you are using?", "votes": 1, "replies": [ { "id": 743247, "postDate": "2020-02-11T23:30:24.013Z", "content": "Thanks,\nIt's TextBlob\nYou can pip install it
", "rawMarkdown": "Thanks,\nIt's TextBlob\nYou can pip install it", "votes": 1 } ] }, { "id": 742577, "postDate": "2020-02-11T11:44:41.273Z", "content": "Congratulations! Superb solution.
", "rawMarkdown": "Congratulations! Superb solution.", "votes": 1, "replies": [ { "id": 742602, "postDate": "2020-02-11T12:06:22.667Z", "content": "Thanks :)
", "rawMarkdown": "Thanks :)" } ] }, { "id": 742519, "postDate": "2020-02-11T10:38:42.027Z", "content": "Congrats @ryunosukeishizaki
\n\nMy bert trained on question and answer separately is pretty similar to yours :
\n\nquestiontitle + questionbody as input and predict on questions related targets (21 columns)
questiontitle + questionbody + answer as input and predict on answers related targets (9 columns)
Cool to know inverse translation augmentation was helpful for you. My teammate @xiaojiu1414 gave it a try too (Spanish -> English)
", "rawMarkdown": "Congrats @ryunosukeishizaki \n\nMy bert trained on question and answer separately is pretty similar to yours : \n\n` questiontitle` + `questionbody ` as input and predict on questions related targets (21 columns)\n\n` questiontitle` + `questionbody ` + `answer` as input and predict on answers related targets (9 columns)\n\nCool to know inverse translation augmentation was helpful for you. My teammate @xiaojiu1414 gave it a try too (Spanish -> English) \n\n", "votes": 1, "replies": [ { "id": 742606, "postDate": "2020-02-11T12:07:50.980Z", "content": "Thanks!\nthis time wasn't but let's meet at next competition!๐
", "rawMarkdown": "Thanks!\nthis time wasn't but let's meet at next competition!๐" }, { "id": 742840, "postDate": "2020-02-11T15:24:07.627Z", "content": "Yeah ! Would be great !!
", "rawMarkdown": "Yeah ! Would be great !! \n", "votes": 1 } ] }, { "id": 742373, "postDate": "2020-02-11T09:10:47.303Z", "content": "Congratulations! Seems augmentation worked quite well for you. Frankly I was sort of afraid to augment the data myself thinking I will spoil the results only. Thank you for sharing your approach with us! :3
", "rawMarkdown": "Congratulations! Seems augmentation worked quite well for you. Frankly I was sort of afraid to augment the data myself thinking I will spoil the results only. Thank you for sharing your approach with us! :3", "votes": 1, "replies": [ { "id": 742429, "postDate": "2020-02-11T09:38:09.223Z", "content": "Thanks :)\nYes, then I did augmentation in 20% when dataload to make it works
", "rawMarkdown": "Thanks :)\nYes, then I did augmentation in 20% when dataload to make it works" } ] }, { "id": 742366, "postDate": "2020-02-11T09:03:16.210Z", "content": "Thanks for your sharing!\nI have tired this kind of model structure, can I ask what score of LB without post process you got?\nMany thanks!
", "rawMarkdown": "Thanks for your sharing!\nI have tired this kind of model structure, can I ask what score of LB without post process you got?\nMany thanks!", "votes": 1, "replies": [ { "id": 742418, "postDate": "2020-02-11T09:35:37.487Z", "content": "I don't remember the score without pp precisely,\nbut it was like LB0.405 -> 0.445
", "rawMarkdown": "I don't remember the score without pp precisely,\nbut it was like LB0.405 -> 0.445", "votes": 1 } ] }, { "id": 742331, "postDate": "2020-02-11T08:37:12.340Z", "content": "May I ask which one was better xlnet or bert because my strategy was closely related to yours but score wasn't :p\nI mean you did augmentation also and post processing as well.\nhttps://www.kaggle.com/gskdhiman/inference-kernel-for-bert-memory-intensive\nIf you can give any suggestions, it will be helpful\nAnd congratulations for the gold.
", "rawMarkdown": "May I ask which one was better xlnet or bert because my strategy was closely related to yours but score wasn't :p\nI mean you did augmentation also and post processing as well.\nhttps://www.kaggle.com/gskdhiman/inference-kernel-for-bert-memory-intensive\nIf you can give any suggestions, it will be helpful\nAnd congratulations for the gold.", "votes": 1, "replies": [ { "id": 742338, "postDate": "2020-02-11T08:42:57.147Z", "content": "bert was better,\nwe had better score after tuning lr
", "rawMarkdown": "bert was better,\nwe had better score after tuning lr", "votes": 2 }, { "id": 742350, "postDate": "2020-02-11T08:52:18.503Z", "content": "Thanks for your quick reply
", "rawMarkdown": "Thanks for your quick reply ", "votes": 1 } ] }, { "id": 742325, "postDate": "2020-02-11T08:28:49.970Z", "content": "Congratulations! Thanks for your sharing! I still have a question, how did you backward each loss? Can you share us your model definition and training strategy. Thanks very much!
the code is not only mine,\nthen if it's possible after talking with teammates I can
", "rawMarkdown": "the code is not only mine,\nthen if it's possible after talking with teammates I can", "votes": 2 }, { "id": 742428, "postDate": "2020-02-11T09:37:43.503Z", "content": "Thanks very much!
", "rawMarkdown": "Thanks very much!", "votes": 1 } ] }, { "id": 742299, "postDate": "2020-02-11T08:15:18.030Z", "content": "Congratulations @ryunosukeishizaki and team!\nThanks for Sharing your Approach!!
", "rawMarkdown": "Congratulations @ryunosukeishizaki and team!\nThanks for Sharing your Approach!!", "votes": 1, "replies": [ { "id": 742340, "postDate": "2020-02-11T08:45:38.513Z", "content": "Thanks :)
", "rawMarkdown": "Thanks :)", "votes": 1 } ] }, { "id": 742282, "postDate": "2020-02-11T07:53:03.167Z", "content": "Congratulations @ryunosukeishizaki and team!
\n\nMay I ask how did you go about augmentation?
\n\nAnd how much of an effect did it have on your score?
", "rawMarkdown": "Congratulations @ryunosukeishizaki and team! \n\nMay I ask how did you go about augmentation? \n\nAnd how much of an effect did it have on your score? ", "votes": 1, "replies": [ { "id": 742283, "postDate": "2020-02-11T07:57:09.540Z", "content": "I referred this โ\nhttps://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/48038\nIt was just about +0.004-5 on LB
", "rawMarkdown": "I referred this โ\nhttps://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/discussion/48038\nIt was just about +0.004-5 on LB", "votes": 1 } ] }, { "id": 742272, "postDate": "2020-02-11T07:43:45.213Z", "content": "Congrats @ryunosukeishizaki and team.
", "rawMarkdown": "Congrats @ryunosukeishizaki and team. ", "votes": 1, "replies": [ { "id": 742273, "postDate": "2020-02-11T07:44:45.577Z", "content": "Thank you!!
", "rawMarkdown": "Thank you!!", "votes": 1 } ] }, { "id": 742257, "postDate": "2020-02-11T07:29:53.837Z", "content": "*Validations are both MultiStratifiedKFold
", "rawMarkdown": "*Validations are both MultiStratifiedKFold", "votes": 2 }, { "id": 866356, "postDate": "2020-05-29T10:21:16.320Z", "content": "Hey, mate ,sorry for my question at this time.\nHow do you make emsemble in model1 ? There should be 3 models , 4 output here. Do you just make a average of them ?
", "rawMarkdown": "Hey, mate ,sorry for my question at this time.\nHow do you make emsemble in model1 ? There should be 3 models , 4 output here. Do you just make a average of them ?" } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T19:07:52Z", "competition_id": 7968, "competition_slug": "google-quest-challenge", "team_id": 4064302, "solution_url": "/competitions/google-quest-challenge/writeups/babaconda-jigsaw-winners-4th-place-solution-overvi" }