I am so surprised that temperature sharpening can give me a 0.0034 boost over mean ensemble after choosing different post_process threshold,I wonder is that normal for this competition?
I am so surprised that temperature sharpening can give me a 0.0034 boost over mean ensemble after choosing different post_process threshold,I wonder is that normal for this competition?
0.03 or 0.003. 0.03 is ANAZING boost!!!!\nBut my perspective is this may cause shape up ?
", "rawMarkdown": "0.03 or 0.003. 0.03 is ANAZING boost!!!!\nBut my perspective is this may cause shape up ?", "votes": null }, { "id": "673603", "postDate": "11/15/2019 07:57:46", "content": "yup,I applied four models ensembling scoring at [0.6567,0.6561,0.6532,0.6539],and used hengck23 method explained here,and scored about 0.611,then I applied Mobassair 's classifier here which has a little bug in his Datagenerator ,he's not using the full test data cuz the test data is not divisable of 32 batch_size,and then boom, I got this huge boost, like 0.0042 after temperature ensemble.Not playing hero,but I'd like to share my post_process parameter like 0.665 and 14000. And now I'm so worried about this huge shakeup would come to me
so i think the key is classification right? if you try mean ensemble with classification, I think your score will boost also.\nPS. your four model is good enough to reach 0.665
", "rawMarkdown": "so i think the key is classification right? if you try mean ensemble with classification, I think your score will boost also.\nPS. your four model is good enough to reach 0.665", "votes": null }, { "id": "673651", "postDate": "11/15/2019 09:31:49", "content": "ThX,Yeah,I'm giving a try on ensembling two classifier like efficientnetb2 and efficientnetb4,so thanks for your advice.比心
", "rawMarkdown": "ThX,Yeah,I'm giving a try on ensembling two classifier like efficientnetb2 and efficientnetb4,so thanks for your advice.比心", "votes": null }, { "id": "673724", "postDate": "11/15/2019 11:59:35", "content": "Hey @sj626591833 nice score, indeed it was a huge boost, I'm also curious to know your score without temperature sharpening, to see if it actually helped, I have also tried it on this competition but I got a low score.
", "rawMarkdown": "Hey @sj626591833 nice score, indeed it was a huge boost, I'm also curious to know your score without temperature sharpening, to see if it actually helped, I have also tried it on this competition but I got a low score.", "votes": null }, { "id": "673759", "postDate": "11/15/2019 13:05:27", "content": "ThX for your reply @dimitreoliveira yeah, I think there is some luck in there,my former score is 0.6619 with mean ensemble and post_processing and classification like everything indeed(haven't tryied K_Fold yet cause short of GPUs),if time is permitted, I would like to try K_Fold on Colab next.
", "rawMarkdown": "ThX for your reply @dimitreoliveira yeah, I think there is some luck in there,my former score is 0.6619 with mean ensemble and post_processing and classification like everything indeed(haven't tryied K_Fold yet cause short of GPUs),if time is permitted, I would like to try K_Fold on Colab next.", "votes": null }, { "id": "673921", "postDate": "11/15/2019 16:58:50", "content": "Nice, I think I'll give temperature sharpening another try. But just one more thing, after applying temperature, do you use the same thresholds, or do you tune it again?
", "rawMarkdown": "Nice, I think I'll give temperature sharpening another try. But just one more thing, after applying temperature, do you use the same thresholds, or do you tune it again?", "votes": null }, { "id": "674118", "postDate": "11/15/2019 22:57:10", "content": "I think you need to.tune it again
", "rawMarkdown": "I think you need to.tune it again", "votes": null } ], "comments": [ { "id": 673564, "author_name": "jiahecao", "author_url": "", "post_date": "11/15/2019 06:48:48", "content": "0.03 or 0.003. 0.03 is ANAZING boost!!!!\nBut my perspective is this may cause shape up ?
", "votes": null, "replies": [ { "id": 673603, "author_name": "sj626591833", "author_url": "", "post_date": "11/15/2019 07:57:46", "content": "yup,I applied four models ensembling scoring at [0.6567,0.6561,0.6532,0.6539],and used hengck23 method explained here,and scored about 0.611,then I applied Mobassair 's classifier here which has a little bug in his Datagenerator ,he's not using the full test data cuz the test data is not divisable of 32 batch_size,and then boom, I got this huge boost, like 0.0042 after temperature ensemble.Not playing hero,but I'd like to share my post_process parameter like 0.665 and 14000. And now I'm so worried about this huge shakeup would come to me
so i think the key is classification right? if you try mean ensemble with classification, I think your score will boost also.\nPS. your four model is good enough to reach 0.665
", "votes": null, "replies": [] }, { "id": 673651, "author_name": "sj626591833", "author_url": "", "post_date": "11/15/2019 09:31:49", "content": "ThX,Yeah,I'm giving a try on ensembling two classifier like efficientnetb2 and efficientnetb4,so thanks for your advice.比心
", "votes": null, "replies": [] }, { "id": 673724, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/15/2019 11:59:35", "content": "Hey @sj626591833 nice score, indeed it was a huge boost, I'm also curious to know your score without temperature sharpening, to see if it actually helped, I have also tried it on this competition but I got a low score.
", "votes": null, "replies": [] }, { "id": 673759, "author_name": "sj626591833", "author_url": "", "post_date": "11/15/2019 13:05:27", "content": "ThX for your reply @dimitreoliveira yeah, I think there is some luck in there,my former score is 0.6619 with mean ensemble and post_processing and classification like everything indeed(haven't tryied K_Fold yet cause short of GPUs),if time is permitted, I would like to try K_Fold on Colab next.
", "votes": null, "replies": [] }, { "id": 673921, "author_name": "dimitreoliveira", "author_url": "", "post_date": "11/15/2019 16:58:50", "content": "Nice, I think I'll give temperature sharpening another try. But just one more thing, after applying temperature, do you use the same thresholds, or do you tune it again?
", "votes": null, "replies": [] }, { "id": 674118, "author_name": "sj626591833", "author_url": "", "post_date": "11/15/2019 22:57:10", "content": "I think you need to.tune it again
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "673560": "I am so surprised that temperature sharpening can give me a 0.0034 boost over mean ensemble after choosing different `post_process threshold`,I wonder is that normal for this competition?", "673564": "0.03 or 0.003. 0.03 is ANAZING boost!!!!\nBut my perspective is this may cause shape up ?", "673603": "yup,I applied four models ensembling scoring at `[0.6567,0.6561,0.6532,0.6539]`,and used hengck23 method explained [here](https://www.kaggle.com/c/severstal-steel-defect-detection/discussion/107716#latest-643059),and scored about 0.611,then I applied Mobassair 's classifier [here](https://www.kaggle.com/mobassir/keras-efficientnetb2-for-classifying-cloud) which has a little bug in his Datagenerator ,he's not using the full test data cuz the test data is not divisable of 32 batch_size,and then boom, I got this huge boost, like 0.0042 after temperature ensemble.Not playing hero,but I'd like to share my post_process parameter like 0.665 and 14000. And now I'm so worried about this huge shakeup would come to me", "673644": "so i think the key is classification right? if you try mean ensemble with classification, I think your score will boost also.\nPS. your four model is good enough to reach 0.665", "673651": "ThX,Yeah,I'm giving a try on ensembling two classifier like efficientnetb2 and efficientnetb4,so thanks for your advice.比心", "673724": "Hey @sj626591833 nice score, indeed it was a huge boost, I'm also curious to know your score without temperature sharpening, to see if it actually helped, I have also tried it on this competition but I got a low score.", "673759": "ThX for your reply @dimitreoliveira yeah, I think there is some luck in there,my former score is 0.6619 with mean ensemble and post_processing and classification like everything indeed(haven't tryied K_Fold yet cause short of GPUs),if time is permitted, I would like to try K_Fold on Colab next.", "673921": "Nice, I think I'll give temperature sharpening another try. But just one more thing, after applying temperature, do you use the same thresholds, or do you tune it again?", "674118": "I think you need to.tune it again" }, "source": "meta" }