{ "id": 299228, "title": "(Losing) my first bronze :\"(", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/299228", "author_name": "", "post_date": "2022-01-06T20:25:53.902148800Z", "votes": 18, "comment_count": 4, "views": 0, "content": "

Sartorius Cell Instance Segmentation was one of my first computer vision detection competitions which I participated in seriously. Almost 1 and a half months of efforts of training, trying out different settings and researching new ways to improve my CV, I got to a rank under the ~70s in the last few weeks. I even subscribed to Colab Pro+ in order to gain better compute to handle the large train times warranted by the competition.

\n

At this stage, being a bit active in the discussions, I was met with 4-5 team merge requests in a short duration. Being very excited to join in with a team and revel in my first \"data science collaboration\", I accepted one of my earliest requests who was an undergrad student at around ~150 rank at the time. Thinking that this person must obviously be a bit more knowledgeable than me I was hopeful to maintain my position. However soon after teaming I realized that this had not been a successful team-up at all.

\n

While we communicated a bit through Instagram, my teammate wasn't very enthusiastic about discussing/developing of code outside of what was available on public notebooks and most of their replies to me were 1 word comments like \"ok\" or asking elementary questions about code already there. However the real thing that confirmed my belief this as a not a good team up was the fact that they sent me a pdf encompassing some methods that a team in the top 50s ranks was using who had shared it since they were \"friends\" around the last 10 days.

\n

\"image\"

\n

Scared of what I should do now since this person could clearly be using some unoriginal or copied code as well (and since I had no options to disband the team now), I told them that we can \"try it out\" but we should focus on the current plan and framework I had made. After a bit of issues in implementing my pseudo-labels pipeline I was able to create a submission which reached 0.314 on the Public LB (and 0.323 on private) without much input from my teammate and mostly relying on public code/my variations of hyper-parameters.

\n

Then from around the week of 20th of December I got a bit busy with my college applications. There was almost 0 contact between us till 29th when I asked they about whether they could implement some of the data augmentation code I had asked them to look into. Just a few hours on the day of deadline, this person made 2 submissions (without corresponding with me or telling me that were going to make them and their 3rd submission in the team overall) which I would later be told were inference pipelines with augmentation code which were \"shared\" with him by some friends again. This code improved upon my 0.314 by 0.001 and did not affect the private LB at all. Though a bit nervous about whether they meant shared the ideas or shared the entire code, I didn't really look into it since I was very busy with college apps around that time.

\n

\"Screenshot-2022-01-07-014926\"
\nlast day subs made by my teammate without coordinating

\n

\"Screenshot-2022-01-06-200028\"

\n

The next day upon checking the LB I was delighted to see that my 0.314 had been successful in scoring a rank of 124. Over the next few days I couldn't check Kaggle due to fine-tuning and sending out college apps. My teammate did not respond to my attempts at communicating him after the deadline. Yesterday after the organizers verification I was heartbroken to see I had been removed from the LB.

\n

While I don't hope to be reinstated (seeing that my teammate did probably copy code for the augmentation), I still feel this was a valuable lesson for us all who hope to team up from the diverse Kaggle community. While the rules don't allow code-sharing or any similar form of communication before teaming up, I still believe that you should evaluate the other person's effort in the competition by number of submissions, rank, past experience, institution, involvement on the forums or past familiarity. I hope that none of the people at this wonderful community go through a similar situation and also that Kaggle can introduce some way to disband teams during a competition as well so as to not disadvantage those who do anonymous team ups to later fall prey to similar situations.

\n

While I didn't win any medals or honors from this, this competition definitely gave me a wider understanding of detectron2 related object detection techniques and also \"life lessons\" in general :p .You can check out the code I used (mostly adapted from public notebooks as you'll see, except for some of pseudo-labelling portion I wrote myself) on GitHub used to get my 0.323 sub :)

\n

I do not wish my teammate any ill-will or hold any hard feelings either. I just hope that moving forward you will stick to the Kaggle ethics of code/idea sharing and team-up/contact people only if you think you may have done work equivalent to their level or potentially can do that level of work.

", "messages": [ { "id": "1640803", "postDate": "01/06/2022 20:25:53", "content": "

Sartorius Cell Instance Segmentation was one of my first computer vision detection competitions which I participated in seriously. Almost 1 and a half months of efforts of training, trying out different settings and researching new ways to improve my CV, I got to a rank under the ~70s in the last few weeks. I even subscribed to Colab Pro+ in order to gain better compute to handle the large train times warranted by the competition.

\n

At this stage, being a bit active in the discussions, I was met with 4-5 team merge requests in a short duration. Being very excited to join in with a team and revel in my first \"data science collaboration\", I accepted one of my earliest requests who was an undergrad student at around ~150 rank at the time. Thinking that this person must obviously be a bit more knowledgeable than me I was hopeful to maintain my position. However soon after teaming I realized that this had not been a successful team-up at all.

\n

While we communicated a bit through Instagram, my teammate wasn't very enthusiastic about discussing/developing of code outside of what was available on public notebooks and most of their replies to me were 1 word comments like \"ok\" or asking elementary questions about code already there. However the real thing that confirmed my belief this as a not a good team up was the fact that they sent me a pdf encompassing some methods that a team in the top 50s ranks was using who had shared it since they were \"friends\" around the last 10 days.

\n

\"image\"

\n

Scared of what I should do now since this person could clearly be using some unoriginal or copied code as well (and since I had no options to disband the team now), I told them that we can \"try it out\" but we should focus on the current plan and framework I had made. After a bit of issues in implementing my pseudo-labels pipeline I was able to create a submission which reached 0.314 on the Public LB (and 0.323 on private) without much input from my teammate and mostly relying on public code/my variations of hyper-parameters.

\n

Then from around the week of 20th of December I got a bit busy with my college applications. There was almost 0 contact between us till 29th when I asked they about whether they could implement some of the data augmentation code I had asked them to look into. Just a few hours on the day of deadline, this person made 2 submissions (without corresponding with me or telling me that were going to make them and their 3rd submission in the team overall) which I would later be told were inference pipelines with augmentation code which were \"shared\" with him by some friends again. This code improved upon my 0.314 by 0.001 and did not affect the private LB at all. Though a bit nervous about whether they meant shared the ideas or shared the entire code, I didn't really look into it since I was very busy with college apps around that time.

\n

\"Screenshot-2022-01-07-014926\"
\nlast day subs made by my teammate without coordinating

\n

\"Screenshot-2022-01-06-200028\"

\n

The next day upon checking the LB I was delighted to see that my 0.314 had been successful in scoring a rank of 124. Over the next few days I couldn't check Kaggle due to fine-tuning and sending out college apps. My teammate did not respond to my attempts at communicating him after the deadline. Yesterday after the organizers verification I was heartbroken to see I had been removed from the LB.

\n

While I don't hope to be reinstated (seeing that my teammate did probably copy code for the augmentation), I still feel this was a valuable lesson for us all who hope to team up from the diverse Kaggle community. While the rules don't allow code-sharing or any similar form of communication before teaming up, I still believe that you should evaluate the other person's effort in the competition by number of submissions, rank, past experience, institution, involvement on the forums or past familiarity. I hope that none of the people at this wonderful community go through a similar situation and also that Kaggle can introduce some way to disband teams during a competition as well so as to not disadvantage those who do anonymous team ups to later fall prey to similar situations.

\n

While I didn't win any medals or honors from this, this competition definitely gave me a wider understanding of detectron2 related object detection techniques and also \"life lessons\" in general :p .You can check out the code I used (mostly adapted from public notebooks as you'll see, except for some of pseudo-labelling portion I wrote myself) on GitHub used to get my 0.323 sub :)

\n

I do not wish my teammate any ill-will or hold any hard feelings either. I just hope that moving forward you will stick to the Kaggle ethics of code/idea sharing and team-up/contact people only if you think you may have done work equivalent to their level or potentially can do that level of work.

", "rawMarkdown": "Sartorius Cell Instance Segmentation was one of my first computer vision detection competitions which I participated in seriously. Almost 1 and a half months of efforts of training, trying out different settings and researching new ways to improve my CV, I got to a rank under the ~70s in the last few weeks. I even subscribed to Colab Pro+ in order to gain better compute to handle the large train times warranted by the competition.\n\nAt this stage, being a bit active in the discussions, I was met with 4-5 team merge requests in a short duration. Being very excited to join in with a team and revel in my first \"data science collaboration\", I accepted one of my earliest requests who was an undergrad student at around ~150 rank at the time. Thinking that this person must obviously be a bit more knowledgeable than me I was hopeful to maintain my position. However soon after teaming I realized that this had not been a successful team-up at all. \n\nWhile we communicated a bit through Instagram, my teammate wasn't very enthusiastic about discussing/developing of code outside of what was available on public notebooks and most of their replies to me were 1 word comments like \"ok\" or asking elementary questions about code already there. However the real thing that confirmed my belief this as a not a good team up was the fact that they sent me a pdf encompassing some methods that a team in the top 50s ranks was using who had shared it since they were \"friends\" around the last 10 days. \n\n\"image\"\n\nScared of what I should do now since this person could clearly be using some unoriginal or copied code as well (and since I had no options to disband the team now), I told them that we can \"try it out\" but we should focus on the current plan and framework I had made. After a bit of issues in implementing my pseudo-labels pipeline I was able to create a submission which reached 0.314 on the Public LB (and 0.323 on private) without much input from my teammate and mostly relying on public code/my variations of hyper-parameters. \n\nThen from around the week of 20th of December I got a bit busy with my college applications. There was almost 0 contact between us till 29th when I asked they about whether they could implement some of the data augmentation code I had asked them to look into. Just a few hours on the day of deadline, this person made 2 submissions (without corresponding with me or telling me that were going to make them and their 3rd submission in the team overall) which I would later be told were inference pipelines with augmentation code which were \"shared\" with him by some friends again. This code improved upon my 0.314 by 0.001 and did not affect the private LB at all. Though a bit nervous about whether they meant shared the ideas or shared the entire code, I didn't really look into it since I was very busy with college apps around that time.\n \n\"Screenshot-2022-01-07-014926\"\nlast day subs made by my teammate without coordinating\n\n\"Screenshot-2022-01-06-200028\"\n\nThe next day upon checking the LB I was delighted to see that my 0.314 had been successful in scoring a rank of 124. Over the next few days I couldn't check Kaggle due to fine-tuning and sending out college apps. My teammate did not respond to my attempts at communicating him after the deadline. Yesterday after the organizers verification I was heartbroken to see I had been removed from the LB. \n\nWhile I don't hope to be reinstated (seeing that my teammate did probably copy code for the augmentation), I still feel this was a valuable lesson for us all who hope to team up from the diverse Kaggle community. While the rules don't allow code-sharing or any similar form of communication before teaming up, I still believe that you should evaluate the other person's effort in the competition by number of submissions, rank, past experience, institution, involvement on the forums or past familiarity. I hope that none of the people at this wonderful community go through a similar situation and also that Kaggle can introduce some way to disband teams during a competition as well so as to not disadvantage those who do anonymous team ups to later fall prey to similar situations. \n\nWhile I didn't win any medals or honors from this, this competition definitely gave me a wider understanding of detectron2 related object detection techniques and also \"life lessons\" in general :p .You can check out the code I used (mostly adapted from public notebooks as you'll see, except for some of pseudo-labelling portion I wrote myself) on [GitHub](https://github.com/PolymathAB16/sartorius-cell-submission) used to get my 0.323 sub :)\n\nI do not wish my teammate any ill-will or hold any hard feelings either. I just hope that moving forward you will stick to the Kaggle ethics of code/idea sharing and team-up/contact people only if you think you may have done work equivalent to their level or potentially can do that level of work.", "votes": null }, { "id": "1649146", "postDate": "01/14/2022 02:17:13", "content": "

Keep working, you did a good job on this!

", "rawMarkdown": "Keep working, you did a good job on this!", "votes": null }, { "id": "1650546", "postDate": "01/15/2022 05:32:45", "content": "

Now you get your first medal in petfinder, congrats👍 @ferlockx

", "rawMarkdown": "Now you get your first medal in petfinder, congrats👍 @ferlockx", "votes": null }, { "id": "1650817", "postDate": "01/15/2022 10:51:12", "content": "

thanks :) after a shakedown from silver lmao ;.; but it's fine haha was kinda expecting a lot of ranking changes in this one. Just taught me to focus much more on stacking in this kind of competition from the next time. Congrats on your gold and latest bronze :)

", "rawMarkdown": "thanks :) after a shakedown from silver lmao ;.; but it's fine haha was kinda expecting a lot of ranking changes in this one. Just taught me to focus much more on stacking in this kind of competition from the next time. Congrats on your gold and latest bronze :)", "votes": null }, { "id": "1650821", "postDate": "01/15/2022 10:57:05", "content": "

Thanks a lot for your support :/ congrats on your silver!

", "rawMarkdown": "Thanks a lot for your support :/ congrats on your silver!", "votes": null } ], "comments": [ { "id": 1649146, "author_name": "charonwangg", "author_url": "", "post_date": "01/14/2022 02:17:13", "content": "

Keep working, you did a good job on this!

", "votes": null, "replies": [ { "id": 1650821, "author_name": "ferlockx", "author_url": "", "post_date": "01/15/2022 10:57:05", "content": "

Thanks a lot for your support :/ congrats on your silver!

", "votes": null, "replies": [] } ] }, { "id": 1650546, "author_name": "forcewithme", "author_url": "", "post_date": "01/15/2022 05:32:45", "content": "

Now you get your first medal in petfinder, congrats👍 @ferlockx

", "votes": null, "replies": [ { "id": 1650817, "author_name": "ferlockx", "author_url": "", "post_date": "01/15/2022 10:51:12", "content": "

thanks :) after a shakedown from silver lmao ;.; but it's fine haha was kinda expecting a lot of ranking changes in this one. Just taught me to focus much more on stacking in this kind of competition from the next time. Congrats on your gold and latest bronze :)

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1640803": "Sartorius Cell Instance Segmentation was one of my first computer vision detection competitions which I participated in seriously. Almost 1 and a half months of efforts of training, trying out different settings and researching new ways to improve my CV, I got to a rank under the ~70s in the last few weeks. I even subscribed to Colab Pro+ in order to gain better compute to handle the large train times warranted by the competition.\n\nAt this stage, being a bit active in the discussions, I was met with 4-5 team merge requests in a short duration. Being very excited to join in with a team and revel in my first \"data science collaboration\", I accepted one of my earliest requests who was an undergrad student at around ~150 rank at the time. Thinking that this person must obviously be a bit more knowledgeable than me I was hopeful to maintain my position. However soon after teaming I realized that this had not been a successful team-up at all. \n\nWhile we communicated a bit through Instagram, my teammate wasn't very enthusiastic about discussing/developing of code outside of what was available on public notebooks and most of their replies to me were 1 word comments like \"ok\" or asking elementary questions about code already there. However the real thing that confirmed my belief this as a not a good team up was the fact that they sent me a pdf encompassing some methods that a team in the top 50s ranks was using who had shared it since they were \"friends\" around the last 10 days. \n\n\"image\"\n\nScared of what I should do now since this person could clearly be using some unoriginal or copied code as well (and since I had no options to disband the team now), I told them that we can \"try it out\" but we should focus on the current plan and framework I had made. After a bit of issues in implementing my pseudo-labels pipeline I was able to create a submission which reached 0.314 on the Public LB (and 0.323 on private) without much input from my teammate and mostly relying on public code/my variations of hyper-parameters. \n\nThen from around the week of 20th of December I got a bit busy with my college applications. There was almost 0 contact between us till 29th when I asked they about whether they could implement some of the data augmentation code I had asked them to look into. Just a few hours on the day of deadline, this person made 2 submissions (without corresponding with me or telling me that were going to make them and their 3rd submission in the team overall) which I would later be told were inference pipelines with augmentation code which were \"shared\" with him by some friends again. This code improved upon my 0.314 by 0.001 and did not affect the private LB at all. Though a bit nervous about whether they meant shared the ideas or shared the entire code, I didn't really look into it since I was very busy with college apps around that time.\n \n\"Screenshot-2022-01-07-014926\"\nlast day subs made by my teammate without coordinating\n\n\"Screenshot-2022-01-06-200028\"\n\nThe next day upon checking the LB I was delighted to see that my 0.314 had been successful in scoring a rank of 124. Over the next few days I couldn't check Kaggle due to fine-tuning and sending out college apps. My teammate did not respond to my attempts at communicating him after the deadline. Yesterday after the organizers verification I was heartbroken to see I had been removed from the LB. \n\nWhile I don't hope to be reinstated (seeing that my teammate did probably copy code for the augmentation), I still feel this was a valuable lesson for us all who hope to team up from the diverse Kaggle community. While the rules don't allow code-sharing or any similar form of communication before teaming up, I still believe that you should evaluate the other person's effort in the competition by number of submissions, rank, past experience, institution, involvement on the forums or past familiarity. I hope that none of the people at this wonderful community go through a similar situation and also that Kaggle can introduce some way to disband teams during a competition as well so as to not disadvantage those who do anonymous team ups to later fall prey to similar situations. \n\nWhile I didn't win any medals or honors from this, this competition definitely gave me a wider understanding of detectron2 related object detection techniques and also \"life lessons\" in general :p .You can check out the code I used (mostly adapted from public notebooks as you'll see, except for some of pseudo-labelling portion I wrote myself) on [GitHub](https://github.com/PolymathAB16/sartorius-cell-submission) used to get my 0.323 sub :)\n\nI do not wish my teammate any ill-will or hold any hard feelings either. I just hope that moving forward you will stick to the Kaggle ethics of code/idea sharing and team-up/contact people only if you think you may have done work equivalent to their level or potentially can do that level of work.", "1649146": "Keep working, you did a good job on this!", "1650546": "Now you get your first medal in petfinder, congrats👍 @ferlockx", "1650817": "thanks :) after a shakedown from silver lmao ;.; but it's fine haha was kinda expecting a lot of ranking changes in this one. Just taught me to focus much more on stacking in this kind of competition from the next time. Congrats on your gold and latest bronze :)", "1650821": "Thanks a lot for your support :/ congrats on your silver!" }, "source": "meta" }