{ "id": 380657, "title": "Best NN Score?", "url": "/competitions/rsna-breast-cancer-detection/discussion/380657", "author_name": "Ali Abdin", "post_date": "2023-01-23T18:27:37.834000", "votes": 3, "comment_count": 6, "views": 0, "content": "

Just out of curiosity and so that people are able to compare their solutions.

\n

To everyone who tried a neural network based approach, following question:

\n", "messages": [ { "id": 2112650, "postDate": "2023-01-23T18:27:37.833Z", "content": "

Just out of curiosity and so that people are able to compare their solutions.

\n

To everyone who tried a neural network based approach, following question:

\n", "rawMarkdown": "Just out of curiosity and so that people are able to compare their solutions.\n\nTo everyone who tried a neural network based approach, following question:\n\n- What is the highest score you can achieve by using a single NN?\n- What is the highest score you can achieve by using an ensemble of NNs?\n", "votes": 3 }, { "id": 2112700, "postDate": "2023-01-23T19:09:41.263Z", "content": "

as far as I know - reading what people said in discussion topics:

\n\n

source: https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2110198

", "rawMarkdown": "as far as I know - reading what people said in discussion topics:\n- single CNN model: 0.58\n- single CNN model 4 folds blend - 0.63\n\nsource: https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2110198", "votes": 1 }, { "id": 2113328, "postDate": "2023-01-24T08:22:39.467Z", "content": "

Single ConvNextV2Base model with 1344x768 resolution images achieves LB 0.53

", "rawMarkdown": "Single ConvNextV2Base model with 1344x768 resolution images achieves LB 0.53", "votes": 2, "replies": [ { "id": 2113438, "postDate": "2023-01-24T09:31:24.770Z", "content": "

based on my experiments and the posts in the forum, i can conclude:

\n
    \n
  1. for resolution 1024, model (after ensemble) performance is limited to ~0.50.
    2. \n
  2. for single model above 0.50, one needs to go beyond 1024.
    3. \n
  3. generally better results for higher resolution and stronger model. (but i cannot go better beyond 1536~2048 which i don't know why … ). both transformer and CNN can achieve high score of 0.58.
    4. \n
", "rawMarkdown": "based on my experiments and the posts in the forum, i can conclude:\n\n1. for resolution 1024, model (after ensemble) performance is limited to ~0.50.\n2. for single model above 0.50, one needs to go beyond 1024.\n3. generally better results for higher resolution and stronger model. (but i cannot go better beyond 1536~2048 which i don't know why ... ). both transformer and CNN can achieve high score of 0.58. ", "votes": 2, "replies": [ { "id": 2114896, "postDate": "2023-01-25T10:03:51.493Z", "content": "

Actually in 1024x1024, I got 0.53 for single model, and 0.54 for ensemble 3 models (more model will exceed time.)

\n

But I also tried larger size, such as 1280x1280 (my machine only have 24G gpu mem), the cv is higher but LB is worse.

", "rawMarkdown": "Actually in 1024x1024, I got 0.53 for single model, and 0.54 for ensemble 3 models (more model will exceed time.)\n\nBut I also tried larger size, such as 1280x1280 (my machine only have 24G gpu mem), the cv is higher but LB is worse." } ] }, { "id": 2114901, "postDate": "2023-01-25T10:05:21.390Z", "content": "

Thx for sharing, so we can use ConvNextV2, regardless of the license?

", "rawMarkdown": "Thx for sharing, so we can use ConvNextV2, regardless of the license?", "replies": [ { "id": 2114950, "postDate": "2023-01-25T10:57:14.420Z", "content": "

if you win, i think ConvNextV2 solution cannot claim prize money.
\nthe shakeup could be big. who knows if your solution will be shakeup?

", "rawMarkdown": "if you win, i think ConvNextV2 solution cannot claim prize money.\nthe shakeup could be big. who knows if your solution will be shakeup?" } ] } ] } ], "comments": [ { "id": 2112700, "author_name": "Remek Kinas", "author_url": "", "post_date": "2023-01-23T19:09:41.263000", "content": "

as far as I know - reading what people said in discussion topics:

\n\n

source: https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2110198

", "votes": 1, "replies": [] }, { "id": 2113328, "author_name": "Mark Wijkhuizen", "author_url": "", "post_date": "2023-01-24T08:22:39.467000", "content": "

Single ConvNextV2Base model with 1344x768 resolution images achieves LB 0.53

", "votes": 2, "replies": [ { "id": 2113438, "author_name": "hengck23", "author_url": "", "post_date": "2023-01-24T09:31:24.770000", "content": "

based on my experiments and the posts in the forum, i can conclude:

\n
    \n
  1. for resolution 1024, model (after ensemble) performance is limited to ~0.50.
    2. \n
  2. for single model above 0.50, one needs to go beyond 1024.
    3. \n
  3. generally better results for higher resolution and stronger model. (but i cannot go better beyond 1536~2048 which i don't know why … ). both transformer and CNN can achieve high score of 0.58.
    4. \n
", "votes": 2, "replies": [ { "id": 2114896, "author_name": "Mr.Fire", "author_url": "", "post_date": "2023-01-25T10:03:51.493000", "content": "

Actually in 1024x1024, I got 0.53 for single model, and 0.54 for ensemble 3 models (more model will exceed time.)

\n

But I also tried larger size, such as 1280x1280 (my machine only have 24G gpu mem), the cv is higher but LB is worse.

", "votes": 0, "replies": [] } ] }, { "id": 2114901, "author_name": "Mr.Fire", "author_url": "", "post_date": "2023-01-25T10:05:21.390000", "content": "

Thx for sharing, so we can use ConvNextV2, regardless of the license?

", "votes": 0, "replies": [ { "id": 2114950, "author_name": "hengck23", "author_url": "", "post_date": "2023-01-25T10:57:14.420000", "content": "

if you win, i think ConvNextV2 solution cannot claim prize money.
\nthe shakeup could be big. who knows if your solution will be shakeup?

", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2112650": "Just out of curiosity and so that people are able to compare their solutions.\n\nTo everyone who tried a neural network based approach, following question:\n\n- What is the highest score you can achieve by using a single NN?\n- What is the highest score you can achieve by using an ensemble of NNs?\n", "2112700": "as far as I know - reading what people said in discussion topics:\n- single CNN model: 0.58\n- single CNN model 4 folds blend - 0.63\n\nsource: https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333#2110198", "2113328": "Single ConvNextV2Base model with 1344x768 resolution images achieves LB 0.53" } }