{ "id": 428451, "title": "39th public/363th private", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/428451", "author_name": "wpl", "post_date": "2023-08-01T13:12:54.744000", "votes": 10, "comment_count": 5, "views": 0, "content": "

To sum up

\n\n

batch:4,6,8

\n

epoch:100-250

\n

opt:AdamW

\n", "messages": [ { "id": 2368942, "postDate": "2023-08-01T13:12:54.743Z", "content": "

To sum up

\n\n

batch:4,6,8

\n

epoch:100-250

\n

opt:AdamW

\n", "rawMarkdown": "**To sum up**\n- **For model training, we only use yolov8x-seg, and the dataset only uses dataset1，only data augmentation is used**\n\nbatch:4,6,8\n\nepoch:100-250\n\nopt:AdamW\n\n- **About the size**: we tried to use various sizes for training and reasoning，from 256 to 1024.\n In the public data set, the** 512 **size is the highest score (**using dilate**),\n![512](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2F3a4a070be5b3449247b71146f6f57ac3%2F2023-08-01%202047081.png?generation=1690894746535707&alt=media)\nwhile in the private data, the **640** size reached a single model. The score of 0.547 (**without any Post-processing method**);\n![640](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2Ff9e9d0806a9eea04c266d00f45413377%2F2023-08-01%20204708.png?generation=1690894663315910&alt=media)\n- About post-processing: We tried to intersect the model masks with average scores to obtain accurate but small answers, and then combined with the best model results. With this method we achieved 0.537 on the public dataset (using mask dilate), and only 0.38 on the private data. When we removed the dilate it reached 0.488. Lost the medal because of it.\n![using dilate](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2F303f330099229ed4597c60514d508f1f%2F2023-08-01%20204934.png?generation=1690894783795205&alt=media)\n![without dilate](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2F4a0f2bd460a3633a30618db2a16bfe77%2F2023-08-01%20205028.png?generation=1690894798876799&alt=media)\n**For the mask, we should believe in the ability of the model itself. Random dilate and erode can be regarded as overfitting.**\n**Medals are just the most common rewards though, it's the abilities we've learned that count!😁**", "votes": 8 }, { "id": 2369690, "postDate": "2023-08-02T00:47:01.523Z", "content": "

I also experience a huuuge shake from 28th public/800+ private😂 I suppose the biggest problem is that I choose the models based on Public score. They overfit a lot. I realize the importance of model validation. It is my first time taking a part in Kaggle competition. Still a long way to go.

", "rawMarkdown": "I also experience a huuuge shake from 28th public/800+ private😂 I suppose the biggest problem is that I choose the models based on Public score. They overfit a lot. I realize the importance of model validation. It is my first time taking a part in Kaggle competition. Still a long way to go. ", "votes": 1, "replies": [ { "id": 2369803, "postDate": "2023-08-02T03:24:40.503Z", "content": "

I even only have a baseline in my first competition, and you are already very good.So,don't be discouraged, what we learn and experience of failure is also a reward😄

", "rawMarkdown": "I even only have a baseline in my first competition, and you are already very good.So,don't be discouraged, what we learn and experience of failure is also a reward😄" } ] }, { "id": 2370949, "postDate": "2023-08-02T18:21:08.687Z", "content": "

It is my first time taking a part in kaggle competition

", "rawMarkdown": "It is my first time taking a part in kaggle competition " }, { "id": 2369052, "postDate": "2023-08-01T14:06:48.187Z", "content": "

I admire your continuous learning and positive attitude! You will achieve more, bro！

", "rawMarkdown": "I admire your continuous learning and positive attitude! You will achieve more, bro！", "votes": 1, "isDeleted": true, "replies": [ { "id": 2369366, "postDate": "2023-08-01T17:42:25.640Z", "content": "

Thanks for your praise! Let's work hard together, bro！😊

", "rawMarkdown": "Thanks for your praise! Let's work hard together, bro！😊" } ] } ], "comments": [ { "id": 2369690, "author_name": "Blue He", "author_url": "", "post_date": "2023-08-02T00:47:01.523000", "content": "

I also experience a huuuge shake from 28th public/800+ private😂 I suppose the biggest problem is that I choose the models based on Public score. They overfit a lot. I realize the importance of model validation. It is my first time taking a part in Kaggle competition. Still a long way to go.

", "votes": 1, "replies": [ { "id": 2369803, "author_name": "wpl", "author_url": "", "post_date": "2023-08-02T03:24:40.503000", "content": "

I even only have a baseline in my first competition, and you are already very good.So,don't be discouraged, what we learn and experience of failure is also a reward😄

", "votes": 0, "replies": [] } ] }, { "id": 2370949, "author_name": "Bagal Vaishnavi", "author_url": "", "post_date": "2023-08-02T18:21:08.687000", "content": "

It is my first time taking a part in kaggle competition

", "votes": 0, "replies": [] }, { "id": 2369052, "author_name": "", "author_url": "", "post_date": "2023-08-01T14:06:48.187000", "content": "

I admire your continuous learning and positive attitude! You will achieve more, bro！

", "votes": 1, "replies": [ { "id": 2369366, "author_name": "wpl", "author_url": "", "post_date": "2023-08-01T17:42:25.640000", "content": "

Thanks for your praise! Let's work hard together, bro！😊

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2368942": "**To sum up**\n- **For model training, we only use yolov8x-seg, and the dataset only uses dataset1，only data augmentation is used**\n\nbatch:4,6,8\n\nepoch:100-250\n\nopt:AdamW\n\n- **About the size**: we tried to use various sizes for training and reasoning，from 256 to 1024.\n In the public data set, the** 512 **size is the highest score (**using dilate**),\n![512](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2F3a4a070be5b3449247b71146f6f57ac3%2F2023-08-01%202047081.png?generation=1690894746535707&alt=media)\nwhile in the private data, the **640** size reached a single model. The score of 0.547 (**without any Post-processing method**);\n![640](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2Ff9e9d0806a9eea04c266d00f45413377%2F2023-08-01%20204708.png?generation=1690894663315910&alt=media)\n- About post-processing: We tried to intersect the model masks with average scores to obtain accurate but small answers, and then combined with the best model results. With this method we achieved 0.537 on the public dataset (using mask dilate), and only 0.38 on the private data. When we removed the dilate it reached 0.488. Lost the medal because of it.\n![using dilate](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2F303f330099229ed4597c60514d508f1f%2F2023-08-01%20204934.png?generation=1690894783795205&alt=media)\n![without dilate](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F14279047%2F4a0f2bd460a3633a30618db2a16bfe77%2F2023-08-01%20205028.png?generation=1690894798876799&alt=media)\n**For the mask, we should believe in the ability of the model itself. Random dilate and erode can be regarded as overfitting.**\n**Medals are just the most common rewards though, it's the abilities we've learned that count!😁**", "2369690": "I also experience a huuuge shake from 28th public/800+ private😂 I suppose the biggest problem is that I choose the models based on Public score. They overfit a lot. I realize the importance of model validation. It is my first time taking a part in Kaggle competition. Still a long way to go. ", "2370949": "It is my first time taking a part in kaggle competition ", "2369052": "I admire your continuous learning and positive attitude! You will achieve more, bro！" } }