It has been wonderful journey throught the competition. After working tirelessely i am happy i stand at some good place in the competition ,with small regret like many may be having i couldnt select the one submission that stood at 50th rank unofficially.\nHere is the solution\nWhat worked most for 90th solution\n1) Model : Efficientnet s b5 ,b4 ensembling of 0.55 and 0.45\n2) Augmentations: Zoom ,Contrast,brightness(fast ai inbuilt with changed probability to .85 from .75 default ),most important rotations from 120 to -120\n3 ) Pre-processing - Simple Bens cropping with threshold of 9 \n4) Pre-trained weights- Inherited from public kernel of drHb thanks to him..
\n\n5) Image size- 256 to 300,304\nFor 50th solution unofficially ,same as above just some difference in the model and ensembling.\n1) Model- Mix of B5,B6 (0.55,0.45) done in last hours of the day,may be if had come to this from beginning could have got even better rank.\nB6 was unexplored by many,it could have worked better i feel.
\n\n*What dint work *\n1) Data: External data from Irda ,Old competition- may be dint try much
\n\n2) Image processing : High pass and low pass filters from open cv library bilateral2d for noise reduction ,filter2d high pass.This visually produced very good images but unsure as to why it dint deliver best,may be some one who tried it and got better results can share proper way and parameters of using them..
\n\n3) All conventional models InceptionResnet,Xception,dense201,dense101,resnet series till 101,Seresnet
\n\nUnexplored\n1) Circle crop with above set of augmentations,with default augs i could achieve the score of only 79.6 on public leader board\n2) Trying More CV loop with image size456 for model b5..With just one CV 5 fold loop i was intuting ti could have worked more if it was trained in sequence from 300,380,456. In my case i jumped from 300 to 456 so it was insufficiently trained but still with just one CV loop run it could achieve pvt score of 90+
\n\nPlease do share your more ideas to help me improve for the next competition..
", "messages": [ { "id": 621104, "postDate": "2019-09-08T07:54:12.970Z", "content": "It has been wonderful journey throught the competition. After working tirelessely i am happy i stand at some good place in the competition ,with small regret like many may be having i couldnt select the one submission that stood at 50th rank unofficially.\nHere is the solution\nWhat worked most for 90th solution\n1) Model : Efficientnet s b5 ,b4 ensembling of 0.55 and 0.45\n2) Augmentations: Zoom ,Contrast,brightness(fast ai inbuilt with changed probability to .85 from .75 default ),most important rotations from 120 to -120\n3 ) Pre-processing - Simple Bens cropping with threshold of 9 \n4) Pre-trained weights- Inherited from public kernel of drHb thanks to him..
\n\n5) Image size- 256 to 300,304\nFor 50th solution unofficially ,same as above just some difference in the model and ensembling.\n1) Model- Mix of B5,B6 (0.55,0.45) done in last hours of the day,may be if had come to this from beginning could have got even better rank.\nB6 was unexplored by many,it could have worked better i feel.
\n\n*What dint work *\n1) Data: External data from Irda ,Old competition- may be dint try much
\n\n2) Image processing : High pass and low pass filters from open cv library bilateral2d for noise reduction ,filter2d high pass.This visually produced very good images but unsure as to why it dint deliver best,may be some one who tried it and got better results can share proper way and parameters of using them..
\n\n3) All conventional models InceptionResnet,Xception,dense201,dense101,resnet series till 101,Seresnet
\n\nUnexplored\n1) Circle crop with above set of augmentations,with default augs i could achieve the score of only 79.6 on public leader board\n2) Trying More CV loop with image size456 for model b5..With just one CV 5 fold loop i was intuting ti could have worked more if it was trained in sequence from 300,380,456. In my case i jumped from 300 to 456 so it was insufficiently trained but still with just one CV loop run it could achieve pvt score of 90+
\n\nPlease do share your more ideas to help me improve for the next competition..
", "rawMarkdown": "It has been wonderful journey throught the competition. After working tirelessely i am happy i stand at some good place in the competition ,with small regret like many may be having i couldnt select the one submission that stood at 50th rank unofficially.\nHere is the solution\n**What worked most for 90th solution**\n1) Model : Efficientnet s b5 ,b4 ensembling of 0.55 and 0.45\n2) Augmentations: Zoom ,Contrast,brightness(fast ai inbuilt with changed probability to .85 from .75 default ),most important rotations from 120 to -120\n3 ) Pre-processing - Simple Bens cropping with threshold of 9 \n4) Pre-trained weights- Inherited from public kernel of drHb thanks to him..\n\n5) Image size- 256 to 300,304\n**For 50th solution unofficially** ,same as above just some difference in the model and ensembling.\n1) Model- Mix of B5,B6 (0.55,0.45) done in last hours of the day,may be if had come to this from beginning could have got even better rank.\nB6 was unexplored by many,it could have worked better i feel.\n\n**What dint work **\n1) Data: External data from Irda ,Old competition- may be dint try much\n\n2) Image processing : High pass and low pass filters from open cv library bilateral2d for noise reduction ,filter2d high pass.This visually produced very good images but unsure as to why it dint deliver best,may be some one who tried it and got better results can share proper way and parameters of using them..\n\n3) All conventional models InceptionResnet,Xception,dense201,dense101,resnet series till 101,Seresnet\n\n**Unexplored**\n1) Circle crop with above set of augmentations,with default augs i could achieve the score of only 79.6 on public leader board\n2) Trying More CV loop with image size456 for model b5..With just one CV 5 fold loop i was intuting ti could have worked more if it was trained in sequence from 300,380,456. In my case i jumped from 300 to 456 so it was insufficiently trained but still with just one CV loop run it could achieve pvt score of 90+\n\nPlease do share your more ideas to help me improve for the next competition..", "votes": 1 }, { "id": 621950, "postDate": "2019-09-09T05:46:20.080Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 621950, "author_name": "", "author_url": "", "post_date": "2019-09-09T05:46:20.080000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "621104": "It has been wonderful journey throught the competition. After working tirelessely i am happy i stand at some good place in the competition ,with small regret like many may be having i couldnt select the one submission that stood at 50th rank unofficially.\nHere is the solution\n**What worked most for 90th solution**\n1) Model : Efficientnet s b5 ,b4 ensembling of 0.55 and 0.45\n2) Augmentations: Zoom ,Contrast,brightness(fast ai inbuilt with changed probability to .85 from .75 default ),most important rotations from 120 to -120\n3 ) Pre-processing - Simple Bens cropping with threshold of 9 \n4) Pre-trained weights- Inherited from public kernel of drHb thanks to him..\n\n5) Image size- 256 to 300,304\n**For 50th solution unofficially** ,same as above just some difference in the model and ensembling.\n1) Model- Mix of B5,B6 (0.55,0.45) done in last hours of the day,may be if had come to this from beginning could have got even better rank.\nB6 was unexplored by many,it could have worked better i feel.\n\n**What dint work **\n1) Data: External data from Irda ,Old competition- may be dint try much\n\n2) Image processing : High pass and low pass filters from open cv library bilateral2d for noise reduction ,filter2d high pass.This visually produced very good images but unsure as to why it dint deliver best,may be some one who tried it and got better results can share proper way and parameters of using them..\n\n3) All conventional models InceptionResnet,Xception,dense201,dense101,resnet series till 101,Seresnet\n\n**Unexplored**\n1) Circle crop with above set of augmentations,with default augs i could achieve the score of only 79.6 on public leader board\n2) Trying More CV loop with image size456 for model b5..With just one CV 5 fold loop i was intuting ti could have worked more if it was trained in sequence from 300,380,456. In my case i jumped from 300 to 456 so it was insufficiently trained but still with just one CV loop run it could achieve pvt score of 90+\n\nPlease do share your more ideas to help me improve for the next competition..", "621950": "" } }