{ "id": 575166, "title": "InversionNet using \"CRFs\" ", "url": "/competitions/waveform-inversion/discussion/575166", "author_name": "", "post_date": "2025-04-26T13:48:35.489925100Z", "votes": 2, "comment_count": 5, "views": 0, "content": "

CRFs(Conditional Random Fields) is method which CNN draw picture smoothly.
\nIt is based on \"Pixels that are close to each other are probably same \".
\nif you think your model's output isn't smooth, you try it!!

", "messages": [ { "id": "3187768", "postDate": "04/26/2025 13:48:35", "content": "

CRFs(Conditional Random Fields) is method which CNN draw picture smoothly.
\nIt is based on \"Pixels that are close to each other are probably same \".
\nif you think your model's output isn't smooth, you try it!!

", "rawMarkdown": "CRFs(Conditional Random Fields) is method which CNN draw picture smoothly.\nIt is based on \"Pixels that are close to each other are probably same \".\nif you think your model's output isn't smooth, you try it!!", "votes": null }, { "id": "3190712", "postDate": "05/01/2025 02:42:01", "content": "

I improve score adding loss which is penalty by around pixel diff !

", "rawMarkdown": "I improve score adding loss which is penalty by around pixel diff !", "votes": null }, { "id": "3191058", "postDate": "05/01/2025 10:26:14", "content": "

Output should NOT be smooth.
\nSee for example FlatVelA.
\nIf it helps you, your model is not good enough.

", "rawMarkdown": "Output should NOT be smooth.\nSee for example FlatVelA.\nIf it helps you, your model is not good enough.", "votes": null }, { "id": "3191156", "postDate": "05/01/2025 13:31:50", "content": "

Thank you for your kind comment! And your score is really impressive, bro!

\n

Maybe my explanation was unclear (partly because my English isn’t very good), but what I meant by \"smooth\" is that the values are not mixed across different layers.

\n

For example, the feature you mentioned, FlatVelA, is just multiple strata overlapping — within each layer, the values are relatively similar.

\n

A “smooth” example: 1500, 1500, 1500, … An “unsmooth” example: 1567, 1490, 1504, …

\n

I believe the \"smoothness\" you referred to is about the entire output, right?

\n

If that’s the case, and you still think the output shouldn't be smooth, could you please share your reasoning? I’d really like to understand.

\n

Finally, here is some evidence supporting my perspective:

\n

Figure 1: A sample of one training output.
\n \"\"

\n

Figure2 : each class's pixel diff data
\n\"\"

", "rawMarkdown": "Thank you for your kind comment! And your score is really impressive, bro!\n\nMaybe my explanation was unclear (partly because my English isn’t very good), but what I meant by \"smooth\" is that the values are not mixed across different layers.\n\nFor example, the feature you mentioned, FlatVelA, is just multiple strata overlapping — within each layer, the values are relatively similar.\n\nA “smooth” example: 1500, 1500, 1500, ... An “unsmooth” example: 1567, 1490, 1504, ...\n\nI believe the \"smoothness\" you referred to is about the entire output, right?\n\nIf that’s the case, and you still think the output shouldn't be smooth, could you please share your reasoning? I’d really like to understand.\n\nFinally, here is some evidence supporting my perspective:\n\nFigure 1: A sample of one training output.\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21047063%2F1ec18f141d8140c6729729376a3efcee%2F2025-05-01%20221734.png?generation=1746105583105037&alt=media)\n\nFigure2 : each class's pixel diff data\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21047063%2Ff4b4a6e316c961cbf546d85f994a7bd9%2Fimage.png?generation=1746105768141199&alt=media)", "votes": null }, { "id": "3191160", "postDate": "05/01/2025 13:39:36", "content": "

If you apply 'smoothness' only across same layers (but not same coumuns) it will be good for FlatVel, but not for e.g. CurveVel.

", "rawMarkdown": "If you apply 'smoothness' only across same layers (but not same coumuns) it will be good for FlatVel, but not for e.g. CurveVel.", "votes": null }, { "id": "3191172", "postDate": "05/01/2025 13:55:06", "content": "

As you said, I also didn't give enough consideration to the curve at first, so my score didn't improve.
\nSo I took vertical, horizontal, and diagonal pixel differences into account and set an appropriate threshold (inferred from Graph 2 above), and the accuracy improved.
\nHowever, as you say, it's true that the accuracy of the model is not sufficient, so I plan to continue improving it.

", "rawMarkdown": "As you said, I also didn't give enough consideration to the curve at first, so my score didn't improve.\nSo I took vertical, horizontal, and diagonal pixel differences into account and set an appropriate threshold (inferred from Graph 2 above), and the accuracy improved.\nHowever, as you say, it's true that the accuracy of the model is not sufficient, so I plan to continue improving it.", "votes": null } ], "comments": [ { "id": 3190712, "author_name": "takayukiishibahsi", "author_url": "", "post_date": "05/01/2025 02:42:01", "content": "

I improve score adding loss which is penalty by around pixel diff !

", "votes": null, "replies": [ { "id": 3191058, "author_name": "shlomoron", "author_url": "", "post_date": "05/01/2025 10:26:14", "content": "

Output should NOT be smooth.
\nSee for example FlatVelA.
\nIf it helps you, your model is not good enough.

", "votes": null, "replies": [ { "id": 3191156, "author_name": "takayukiishibahsi", "author_url": "", "post_date": "05/01/2025 13:31:50", "content": "

Thank you for your kind comment! And your score is really impressive, bro!

\n

Maybe my explanation was unclear (partly because my English isn’t very good), but what I meant by \"smooth\" is that the values are not mixed across different layers.

\n

For example, the feature you mentioned, FlatVelA, is just multiple strata overlapping — within each layer, the values are relatively similar.

\n

A “smooth” example: 1500, 1500, 1500, … An “unsmooth” example: 1567, 1490, 1504, …

\n

I believe the \"smoothness\" you referred to is about the entire output, right?

\n

If that’s the case, and you still think the output shouldn't be smooth, could you please share your reasoning? I’d really like to understand.

\n

Finally, here is some evidence supporting my perspective:

\n

Figure 1: A sample of one training output.
\n \"\"

\n

Figure2 : each class's pixel diff data
\n\"\"

", "votes": null, "replies": [ { "id": 3191160, "author_name": "shlomoron", "author_url": "", "post_date": "05/01/2025 13:39:36", "content": "

If you apply 'smoothness' only across same layers (but not same coumuns) it will be good for FlatVel, but not for e.g. CurveVel.

", "votes": null, "replies": [ { "id": 3191172, "author_name": "takayukiishibahsi", "author_url": "", "post_date": "05/01/2025 13:55:06", "content": "

As you said, I also didn't give enough consideration to the curve at first, so my score didn't improve.
\nSo I took vertical, horizontal, and diagonal pixel differences into account and set an appropriate threshold (inferred from Graph 2 above), and the accuracy improved.
\nHowever, as you say, it's true that the accuracy of the model is not sufficient, so I plan to continue improving it.

", "votes": null, "replies": [] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "3187768": "CRFs(Conditional Random Fields) is method which CNN draw picture smoothly.\nIt is based on \"Pixels that are close to each other are probably same \".\nif you think your model's output isn't smooth, you try it!!", "3190712": "I improve score adding loss which is penalty by around pixel diff !", "3191058": "Output should NOT be smooth.\nSee for example FlatVelA.\nIf it helps you, your model is not good enough.", "3191156": "Thank you for your kind comment! And your score is really impressive, bro!\n\nMaybe my explanation was unclear (partly because my English isn’t very good), but what I meant by \"smooth\" is that the values are not mixed across different layers.\n\nFor example, the feature you mentioned, FlatVelA, is just multiple strata overlapping — within each layer, the values are relatively similar.\n\nA “smooth” example: 1500, 1500, 1500, ... An “unsmooth” example: 1567, 1490, 1504, ...\n\nI believe the \"smoothness\" you referred to is about the entire output, right?\n\nIf that’s the case, and you still think the output shouldn't be smooth, could you please share your reasoning? I’d really like to understand.\n\nFinally, here is some evidence supporting my perspective:\n\nFigure 1: A sample of one training output.\n ![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21047063%2F1ec18f141d8140c6729729376a3efcee%2F2025-05-01%20221734.png?generation=1746105583105037&alt=media)\n\nFigure2 : each class's pixel diff data\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F21047063%2Ff4b4a6e316c961cbf546d85f994a7bd9%2Fimage.png?generation=1746105768141199&alt=media)", "3191160": "If you apply 'smoothness' only across same layers (but not same coumuns) it will be good for FlatVel, but not for e.g. CurveVel.", "3191172": "As you said, I also didn't give enough consideration to the curve at first, so my score didn't improve.\nSo I took vertical, horizontal, and diagonal pixel differences into account and set an appropriate threshold (inferred from Graph 2 above), and the accuracy improved.\nHowever, as you say, it's true that the accuracy of the model is not sufficient, so I plan to continue improving it." }, "source": "meta" }