{ "id": 172416, "title": "[80th place solution😂] Source-separated Psuedo Labeling", "url": "/competitions/global-wheat-detection/discussion/172416", "author_name": "arutema47", "post_date": "2020-08-05T01:16:21.278000", "votes": 42, "comment_count": 14, "views": 0, "content": "

First of all, thanks for the host for holding this competition!

\n

And great thanks to our team members, @datahobbit, @ratthachat, @yashchoudhary, @nitindatta and it was a pleasure teaming up. Everyone played a great part in the solutions which @ratthachat elaborates! It was fun having discussions and everybody was a great analyst/coder. We did a great job and made it into the TOP-2% ;)

\n

This is our solution of our team Wheat Fields!

\n

(I will come back to update with the details.)

\n

Models

\n\n

We evaluated our CV based on data source separated CV with original weightings.

\n

Source separated psuedo labeling

\n

Idea

\n

We were scared that psuedo labeling with the entire test set prior will overfit to the UTokyo data which should be 70% of the test data. With this source-split, we can train model which overfit to UTokyo and other model to the rest, maximizing the PB.

\n

Step.1

\n

Predict on test set using Best CV Model and split the test set to two: TestSetB :UToyko which image has number of wheats <39 and TestSetC: number of wheats>=39. We save the labels for psuedo label as well

\n

Step2:

\n

train two models --> Model-B train with pseudo [TestSetB, traindata] and Model-C train with pseudo [TestSetC,traindata].

\n

Inference:

\n

Predict TestSetB with Model-B and TestSetC with Model-C.

\n

Codes!

\n

I uploaded my training codes here.
\nhttps://github.com/kentaroy47/kaggle-wheat-arutema47

", "messages": [ { "id": 958415, "postDate": "2020-08-05T01:16:21.277Z", "content": "

First of all, thanks for the host for holding this competition!

\n

And great thanks to our team members, @datahobbit, @ratthachat, @yashchoudhary, @nitindatta and it was a pleasure teaming up. Everyone played a great part in the solutions which @ratthachat elaborates! It was fun having discussions and everybody was a great analyst/coder. We did a great job and made it into the TOP-2% ;)

\n

This is our solution of our team Wheat Fields!

\n

(I will come back to update with the details.)

\n

Models

\n\n

We evaluated our CV based on data source separated CV with original weightings.

\n

Source separated psuedo labeling

\n

Idea

\n

We were scared that psuedo labeling with the entire test set prior will overfit to the UTokyo data which should be 70% of the test data. With this source-split, we can train model which overfit to UTokyo and other model to the rest, maximizing the PB.

\n

Step.1

\n

Predict on test set using Best CV Model and split the test set to two: TestSetB :UToyko which image has number of wheats <39 and TestSetC: number of wheats>=39. We save the labels for psuedo label as well

\n

Step2:

\n

train two models --> Model-B train with pseudo [TestSetB, traindata] and Model-C train with pseudo [TestSetC,traindata].

\n

Inference:

\n

Predict TestSetB with Model-B and TestSetC with Model-C.

\n

Codes!

\n

I uploaded my training codes here.
\nhttps://github.com/kentaroy47/kaggle-wheat-arutema47

", "rawMarkdown": "First of all, thanks for the host for holding this competition! \n\nAnd great thanks to our team members, @datahobbit, @ratthachat, @yashchoudhary, @nitindatta and it was a pleasure teaming up. Everyone played a great part in the solutions which @ratthachat elaborates! It was fun having discussions and everybody was a great analyst/coder. We did a great job and made it into the TOP-2% ;)\n\nThis is our solution of our team Wheat Fields!\n\n(I will come back to update with the details.)\n\n## Models\n* Model: EffdetD5 2 folds ensemble \n\nWe evaluated our CV based on data source separated CV with original weightings.\n\n## Source separated psuedo labeling\n### Idea\nWe were scared that psuedo labeling with the entire test set prior will overfit to the UTokyo data which should be 70% of the test data. With this source-split, we can train model which overfit to UTokyo and other model to the rest, maximizing the PB.\n\n### Step.1 \nPredict on test set using Best CV Model and split the test set to two: TestSetB :UToyko which image has `number of wheats <39` and TestSetC: `number of wheats>=39`. We save the labels for psuedo label as well\n\n### Step2: \ntrain two models --> Model-B train with pseudo [TestSetB, traindata] and Model-C train with pseudo [TestSetC,traindata].\n\n### Inference: \nPredict TestSetB with Model-B and TestSetC with Model-C.\n\n# Codes!\nI uploaded my training codes here.\nhttps://github.com/kentaroy47/kaggle-wheat-arutema47", "votes": 42 }, { "id": 958494, "postDate": "2020-08-05T02:48:45.167Z", "content": "

Thanks @kyoshioka47 (arutema47) again for his great job!! Also thanks to my other teammates @nitindatta @datahobbit @yashchoudhary for their dedicated efforts!

\n\n

In fact, our two selections are (1) EffDet as described by arutema here, and (2) FRCNN-ResNest which is able to reach 0.770+ Public LB at the end but not as good in Private LB (around 0.685). In FRCNN-ResNest we use the Pseudo+Ensemble technique which in case you guys feel interested in, you can see our published notebook here :

\n\n

https://www.kaggle.com/ratthachat/wheat-resnest-pseudo-ensemble-framework-for-77pub

", "rawMarkdown": "Thanks @kyoshioka47 (arutema47) again for his great job!! Also thanks to my other teammates @nitindatta @datahobbit @yashchoudhary for their dedicated efforts!\n\nIn fact, our two selections are (1) **EffDet** as described by arutema here, and (2) **FRCNN-ResNest** which is able to reach 0.770+ Public LB at the end but not as good in Private LB (around 0.685). In FRCNN-ResNest we use the Pseudo+Ensemble technique which in case you guys feel interested in, you can see our published notebook here : \n\nhttps://www.kaggle.com/ratthachat/wheat-resnest-pseudo-ensemble-framework-for-77pub", "votes": 4, "replies": [ { "id": 958502, "postDate": "2020-08-05T02:57:46.757Z", "content": "

Thanks for sharing!\nAnd Congrats @ratthachat !

", "rawMarkdown": "Thanks for sharing!\nAnd Congrats @ratthachat !", "votes": 1 }, { "id": 958505, "postDate": "2020-08-05T02:59:29.443Z", "content": "

Thanks Heroseo @piantic ! To answer your question above, both EffDet and ResNest use 1024 image size :)

", "rawMarkdown": "Thanks Heroseo @piantic ! To answer your question above, both EffDet and ResNest use 1024 image size :) ", "votes": 1 } ] }, { "id": 958476, "postDate": "2020-08-05T02:34:04.680Z", "content": "

Congrats @nguyenbadung!\nand Congrats @datahobbit @ratthachat @yashchoudhary @nitindatta!

\n\n

The idea seems simple but good!\nMay I ask that what is image_size for training?

\n\n

p.s. Maybe your team go up to a higher rank. :)

", "rawMarkdown": "Congrats @nguyenbadung!\nand Congrats @datahobbit @ratthachat @yashchoudhary @nitindatta!\n\nThe idea seems simple but good!\nMay I ask that what is image_size for training?\n\np.s. Maybe your team go up to a higher rank. :)\n", "votes": 4, "replies": [ { "id": 963379, "postDate": "2020-08-09T01:23:50.010Z", "content": "

Thanks for the warm words! Let's see after the YOLO subs are removed..

", "rawMarkdown": "Thanks for the warm words! Let's see after the YOLO subs are removed.." } ] }, { "id": 959297, "postDate": "2020-08-05T13:31:07.600Z", "content": "

Quite interesting

", "rawMarkdown": "Quite interesting", "votes": 1, "replies": [ { "id": 963382, "postDate": "2020-08-09T01:24:33.070Z", "content": "

Thank you!

", "rawMarkdown": "Thank you!" } ] }, { "id": 958540, "postDate": "2020-08-05T03:37:37.253Z", "content": "

wow, learn a new trick. Big congrats to you guys!

", "rawMarkdown": "wow, learn a new trick. Big congrats to you guys!", "votes": 1, "replies": [ { "id": 963381, "postDate": "2020-08-09T01:24:13.370Z", "content": "

Thank you!

", "rawMarkdown": "Thank you!" } ] }, { "id": 958522, "postDate": "2020-08-05T03:21:15.943Z", "content": "

Good results! Thanks for sharing your team solution @kyoshioka47 @ratthachat

", "rawMarkdown": "Good results! Thanks for sharing your team solution @kyoshioka47 @ratthachat ", "votes": 1, "replies": [ { "id": 963380, "postDate": "2020-08-09T01:24:04.863Z", "content": "

Big thanks to the whole team ;)

", "rawMarkdown": "Big thanks to the whole team ;)" } ] }, { "id": 962952, "postDate": "2020-08-08T15:09:12.520Z", "content": "

Thank you @kyoshioka47 for sharing your approach and congratulation... Soon gonna give it a hands-on practice...

", "rawMarkdown": "Thank you @kyoshioka47 for sharing your approach and congratulation... Soon gonna give it a hands-on practice... ", "replies": [ { "id": 963383, "postDate": "2020-08-09T01:25:03.770Z", "content": "

Thanks! The idea is simple but is quite complicated to implement.. ;)

", "rawMarkdown": "Thanks! The idea is simple but is quite complicated to implement.. ;)" } ] }, { "id": 958741, "postDate": "2020-08-05T06:06:13.810Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 958494, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2020-08-05T02:48:45.167000", "content": "

Thanks @kyoshioka47 (arutema47) again for his great job!! Also thanks to my other teammates @nitindatta @datahobbit @yashchoudhary for their dedicated efforts!

\n\n

In fact, our two selections are (1) EffDet as described by arutema here, and (2) FRCNN-ResNest which is able to reach 0.770+ Public LB at the end but not as good in Private LB (around 0.685). In FRCNN-ResNest we use the Pseudo+Ensemble technique which in case you guys feel interested in, you can see our published notebook here :

\n\n

https://www.kaggle.com/ratthachat/wheat-resnest-pseudo-ensemble-framework-for-77pub

", "votes": 4, "replies": [ { "id": 958502, "author_name": "Heroseo", "author_url": "", "post_date": "2020-08-05T02:57:46.757000", "content": "

Thanks for sharing!\nAnd Congrats @ratthachat !

", "votes": 1, "replies": [] }, { "id": 958505, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2020-08-05T02:59:29.443000", "content": "

Thanks Heroseo @piantic ! To answer your question above, both EffDet and ResNest use 1024 image size :)

", "votes": 1, "replies": [] } ] }, { "id": 958476, "author_name": "Heroseo", "author_url": "", "post_date": "2020-08-05T02:34:04.680000", "content": "

Congrats @nguyenbadung!\nand Congrats @datahobbit @ratthachat @yashchoudhary @nitindatta!

\n\n

The idea seems simple but good!\nMay I ask that what is image_size for training?

\n\n

p.s. Maybe your team go up to a higher rank. :)

", "votes": 4, "replies": [ { "id": 963379, "author_name": "arutema47", "author_url": "", "post_date": "2020-08-09T01:23:50.010000", "content": "

Thanks for the warm words! Let's see after the YOLO subs are removed..

", "votes": 0, "replies": [] } ] }, { "id": 959297, "author_name": "mayhatten", "author_url": "", "post_date": "2020-08-05T13:31:07.600000", "content": "

Quite interesting

", "votes": 1, "replies": [ { "id": 963382, "author_name": "arutema47", "author_url": "", "post_date": "2020-08-09T01:24:33.070000", "content": "

Thank you!

", "votes": 0, "replies": [] } ] }, { "id": 958540, "author_name": "", "author_url": "", "post_date": "2020-08-05T03:37:37.253000", "content": "

wow, learn a new trick. Big congrats to you guys!

", "votes": 1, "replies": [ { "id": 963381, "author_name": "arutema47", "author_url": "", "post_date": "2020-08-09T01:24:13.370000", "content": "

Thank you!

", "votes": 0, "replies": [] } ] }, { "id": 958522, "author_name": "KhanhVD", "author_url": "", "post_date": "2020-08-05T03:21:15.943000", "content": "

Good results! Thanks for sharing your team solution @kyoshioka47 @ratthachat

", "votes": 1, "replies": [ { "id": 963380, "author_name": "arutema47", "author_url": "", "post_date": "2020-08-09T01:24:04.863000", "content": "

Big thanks to the whole team ;)

", "votes": 0, "replies": [] } ] }, { "id": 962952, "author_name": "VinayVikram", "author_url": "", "post_date": "2020-08-08T15:09:12.520000", "content": "

Thank you @kyoshioka47 for sharing your approach and congratulation... Soon gonna give it a hands-on practice...

", "votes": 0, "replies": [ { "id": 963383, "author_name": "arutema47", "author_url": "", "post_date": "2020-08-09T01:25:03.770000", "content": "

Thanks! The idea is simple but is quite complicated to implement.. ;)

", "votes": 0, "replies": [] } ] }, { "id": 958741, "author_name": "", "author_url": "", "post_date": "2020-08-05T06:06:13.810000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "958415": "First of all, thanks for the host for holding this competition! \n\nAnd great thanks to our team members, @datahobbit, @ratthachat, @yashchoudhary, @nitindatta and it was a pleasure teaming up. Everyone played a great part in the solutions which @ratthachat elaborates! It was fun having discussions and everybody was a great analyst/coder. We did a great job and made it into the TOP-2% ;)\n\nThis is our solution of our team Wheat Fields!\n\n(I will come back to update with the details.)\n\n## Models\n* Model: EffdetD5 2 folds ensemble \n\nWe evaluated our CV based on data source separated CV with original weightings.\n\n## Source separated psuedo labeling\n### Idea\nWe were scared that psuedo labeling with the entire test set prior will overfit to the UTokyo data which should be 70% of the test data. With this source-split, we can train model which overfit to UTokyo and other model to the rest, maximizing the PB.\n\n### Step.1 \nPredict on test set using Best CV Model and split the test set to two: TestSetB :UToyko which image has `number of wheats <39` and TestSetC: `number of wheats>=39`. We save the labels for psuedo label as well\n\n### Step2: \ntrain two models --> Model-B train with pseudo [TestSetB, traindata] and Model-C train with pseudo [TestSetC,traindata].\n\n### Inference: \nPredict TestSetB with Model-B and TestSetC with Model-C.\n\n# Codes!\nI uploaded my training codes here.\nhttps://github.com/kentaroy47/kaggle-wheat-arutema47", "958494": "Thanks @kyoshioka47 (arutema47) again for his great job!! Also thanks to my other teammates @nitindatta @datahobbit @yashchoudhary for their dedicated efforts!\n\nIn fact, our two selections are (1) **EffDet** as described by arutema here, and (2) **FRCNN-ResNest** which is able to reach 0.770+ Public LB at the end but not as good in Private LB (around 0.685). In FRCNN-ResNest we use the Pseudo+Ensemble technique which in case you guys feel interested in, you can see our published notebook here : \n\nhttps://www.kaggle.com/ratthachat/wheat-resnest-pseudo-ensemble-framework-for-77pub", "958476": "Congrats @nguyenbadung!\nand Congrats @datahobbit @ratthachat @yashchoudhary @nitindatta!\n\nThe idea seems simple but good!\nMay I ask that what is image_size for training?\n\np.s. Maybe your team go up to a higher rank. :)\n", "959297": "Quite interesting", "958540": "wow, learn a new trick. Big congrats to you guys!", "958522": "Good results! Thanks for sharing your team solution @kyoshioka47 @ratthachat ", "962952": "Thank you @kyoshioka47 for sharing your approach and congratulation... Soon gonna give it a hands-on practice... ", "958741": "" } }