It is my first time taking a part in Kaggle competition. I learned a lot, though finally I lost the game.
\nHere, I'm going to present some mistakes I've made and the experience I've got from the failure. Hope to receive some suggestions.
\n1、The main reason of dropping so much must be the overfitting. And we mistakenly choose models based on Public LB scores. Although in the last week I realized using validation set should be a wiser choice, we did not make a change.
\n2、In the competition, we used ensemble strategy(NMS and WMF on Yolov7&mmdet), which largely help level up the scores on Public LB but on Private it is extremely low(here is my notebook). I suppose our ensemble function is OK and the problem should lie in choosing overfitting single model. Ridiculously, one of our single Yolov7 model, which gained the lowest point on Public, scores the best on Private.
\n3、In the last days, we tried to add TTA but we failed. The score drops dramatically. Still, I do not understand why. Hope to learn from good examples. I use different image (origin&rl_filp&ud_filp)as input and then ensemble them with WMF.
\n4、We also used dilation to process the mask. I guess it is another reason for the low score according to the discussion.
\n5、A lot of things I haven't try: multi-scale training and inference, data augmentation…
\nYesterday I was really upset about the huuuuge shake! Ranking 27 on Public, I thought I could win the competition. Well, too young too naive, ha-ha. This my first try. There is still a long way to go. I will learn harder and continue taking part in Kaggle competition. Any suggestion for learning on Kaggle is highly appreciated.
It is my first time taking a part in Kaggle competition. I learned a lot, though finally I lost the game.
\nHere, I'm going to present some mistakes I've made and the experience I've got from the failure. Hope to receive some suggestions.
\n1、The main reason of dropping so much must be the overfitting. And we mistakenly choose models based on Public LB scores. Although in the last week I realized using validation set should be a wiser choice, we did not make a change.
\n2、In the competition, we used ensemble strategy(NMS and WMF on Yolov7&mmdet), which largely help level up the scores on Public LB but on Private it is extremely low(here is my notebook). I suppose our ensemble function is OK and the problem should lie in choosing overfitting single model. Ridiculously, one of our single Yolov7 model, which gained the lowest point on Public, scores the best on Private.
\n3、In the last days, we tried to add TTA but we failed. The score drops dramatically. Still, I do not understand why. Hope to learn from good examples. I use different image (origin&rl_filp&ud_filp)as input and then ensemble them with WMF.
\n4、We also used dilation to process the mask. I guess it is another reason for the low score according to the discussion.
\n5、A lot of things I haven't try: multi-scale training and inference, data augmentation…
\nYesterday I was really upset about the huuuuge shake! Ranking 27 on Public, I thought I could win the competition. Well, too young too naive, ha-ha. This my first try. There is still a long way to go. I will learn harder and continue taking part in Kaggle competition. Any suggestion for learning on Kaggle is highly appreciated.
I had a fairly similar experience in my first go on kaggle. It does happen, the longer you are on the platform however, you start to learn techniques that cause you to leak more and more and eventually it will show. Consistency is key!
", "rawMarkdown": "I had a fairly similar experience in my first go on kaggle. It does happen, the longer you are on the platform however, you start to learn techniques that cause you to leak more and more and eventually it will show. Consistency is key!", "votes": null }, { "id": "2371162", "postDate": "08/03/2023 00:42:40", "content": "I got. Thanks!
", "rawMarkdown": "I got. Thanks!", "votes": null }, { "id": "2371342", "postDate": "08/03/2023 04:38:50", "content": "my local CV and public LB is work ok,but private LB is not good,I think the dilation convolution is the key to fail.when I don't use the dilation convolution, the private LB is also good
", "rawMarkdown": "my local CV and public LB is work ok,but private LB is not good,I think the dilation convolution is the key to fail.when I don't use the dilation convolution, the private LB is also good", "votes": null }, { "id": "2372404", "postDate": "08/03/2023 16:42:57", "content": "Yeah,yeah. I also suffer from dilation.
", "rawMarkdown": "Yeah,yeah. I also suffer from dilation.", "votes": null } ], "comments": [ { "id": 2370528, "author_name": "cody11null", "author_url": "", "post_date": "08/02/2023 13:28:48", "content": "I had a fairly similar experience in my first go on kaggle. It does happen, the longer you are on the platform however, you start to learn techniques that cause you to leak more and more and eventually it will show. Consistency is key!
", "votes": null, "replies": [ { "id": 2371162, "author_name": "bluehe", "author_url": "", "post_date": "08/03/2023 00:42:40", "content": "I got. Thanks!
", "votes": null, "replies": [ { "id": 2371342, "author_name": "chingllc", "author_url": "", "post_date": "08/03/2023 04:38:50", "content": "my local CV and public LB is work ok,but private LB is not good,I think the dilation convolution is the key to fail.when I don't use the dilation convolution, the private LB is also good
", "votes": null, "replies": [ { "id": 2372404, "author_name": "bluehe", "author_url": "", "post_date": "08/03/2023 16:42:57", "content": "Yeah,yeah. I also suffer from dilation.
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2369686": "It is my first time taking a part in Kaggle competition. I learned a lot, though finally I lost the game.\nHere, I'm going to present some mistakes I've made and the experience I've got from the failure. Hope to receive some suggestions.\n1、The main reason of dropping so much must be the overfitting. And we mistakenly choose models based on Public LB scores. Although in the last week I realized using validation set should be a wiser choice, we did not make a change.\n2、In the competition, we used ensemble strategy(NMS and WMF on Yolov7&mmdet), which largely help level up the scores on Public LB but on Private it is extremely low(here is my [notebook](https://www.kaggle.com/code/bluehe/hubmap-ensemble-yolov7-mmdet?scriptVersionId=138229964)). I suppose our ensemble function is OK and the problem should lie in choosing overfitting single model. Ridiculously, one of our single Yolov7 model, which gained the lowest point on Public, scores the best on Private.\n3、In the last days, we tried to add TTA but we failed. The score drops dramatically. Still, I do not understand why. Hope to learn from good examples. I use different image (origin&rl_filp&ud_filp)as input and then ensemble them with WMF.\n4、We also used dilation to process the mask. I guess it is another reason for the low score according to the discussion.\n5、A lot of things I haven't try: multi-scale training and inference, data augmentation...\nYesterday I was really upset about the huuuuge shake! Ranking 27 on Public, I thought I could win the competition. Well, too young too naive, ha-ha. This my first try. There is still a long way to go. I will learn harder and continue taking part in Kaggle competition. Any suggestion for learning on Kaggle is highly appreciated.", "2370528": "I had a fairly similar experience in my first go on kaggle. It does happen, the longer you are on the platform however, you start to learn techniques that cause you to leak more and more and eventually it will show. Consistency is key!", "2371162": "I got. Thanks!", "2371342": "my local CV and public LB is work ok,but private LB is not good,I think the dilation convolution is the key to fail.when I don't use the dilation convolution, the private LB is also good", "2372404": "Yeah,yeah. I also suffer from dilation." }, "source": "meta" }