Hi everyone,
\nThis is my first post here—please let me know if anything is off in terms of Kaggle etiquette or norms.
I tried running Bartley’s notebook and attempted some training, but the GPU computation required was quite intense. Eventually, I gave up on training the model myself, as it didn’t seem feasible to surpass a public score of 25.6 with my approach (I’m still a beginner, so it was a valuable learning experience regardless).
\nI'm curious—
\nFor those who achieved silver or bronze medals, did you rely mainly on the 45-hour/week Kaggle GPU for training? Or was it necessary to use your own GPU resources, collaborate with a team, or find other workarounds for the compute demands?
Would love to hear your experience—thanks in advance!
", "messages": [ { "id": "3237216", "postDate": "07/01/2025 01:04:14", "content": "Hi everyone,
\nThis is my first post here—please let me know if anything is off in terms of Kaggle etiquette or norms.
I tried running Bartley’s notebook and attempted some training, but the GPU computation required was quite intense. Eventually, I gave up on training the model myself, as it didn’t seem feasible to surpass a public score of 25.6 with my approach (I’m still a beginner, so it was a valuable learning experience regardless).
\nI'm curious—
\nFor those who achieved silver or bronze medals, did you rely mainly on the 45-hour/week Kaggle GPU for training? Or was it necessary to use your own GPU resources, collaborate with a team, or find other workarounds for the compute demands?
Would love to hear your experience—thanks in advance!
", "rawMarkdown": "Hi everyone,\nThis is my first post here—please let me know if anything is off in terms of Kaggle etiquette or norms.\n\nI tried running Bartley’s notebook and attempted some training, but the GPU computation required was quite intense. Eventually, I gave up on training the model myself, as it didn’t seem feasible to surpass a public score of 25.6 with my approach (I’m still a beginner, so it was a valuable learning experience regardless).\n\nI'm curious—\nFor those who achieved silver or bronze medals, did you rely mainly on the 45-hour/week Kaggle GPU for training? Or was it necessary to use your own GPU resources, collaborate with a team, or find other workarounds for the compute demands?\n\nWould love to hear your experience—thanks in advance!", "votes": null }, { "id": "3237234", "postDate": "07/01/2025 01:32:03", "content": "As a result, clustering to improve 25.6 solution to 25.4 didn't need GPUs and it deserve bronze medals :https://www.kaggle.com/competitions/waveform-inversion/discussion/587395
\nBut, I needed GPUs and TPUs to train DNN models and I rely on Google's TPU Research Cloud program, that offers TPUs for free.
", "rawMarkdown": "As a result, clustering to improve 25.6 solution to 25.4 didn't need GPUs and it deserve bronze medals :https://www.kaggle.com/competitions/waveform-inversion/discussion/587395\n\nBut, I needed GPUs and TPUs to train DNN models and I rely on Google's TPU Research Cloud program, that offers TPUs for free.", "votes": null }, { "id": "3237262", "postDate": "07/01/2025 02:22:23", "content": "I built a quick classifier early on for predicting the family because I saw that scores were so different between them. I had thought optimizing score on say Curve_Fault_B vs some of the others may be the breakthrough. It seems that split was not needed which I later found. Instead however, we were able to use the models that were made public and optimized weights based off families with these submissions. We figured there would be little shake with the amount of data available. We decided early on that the amount of compute resources did not align with our goals outside of Kaggle so we opted to go for the cheap route. Glad to hear some others had some cool ideas. Also interesting to hear how far some made it on 50 series cards. Next time I might have to put my 4080 super through the ringer haha
", "rawMarkdown": "I built a quick classifier early on for predicting the family because I saw that scores were so different between them. I had thought optimizing score on say Curve_Fault_B vs some of the others may be the breakthrough. It seems that split was not needed which I later found. Instead however, we were able to use the models that were made public and optimized weights based off families with these submissions. We figured there would be little shake with the amount of data available. We decided early on that the amount of compute resources did not align with our goals outside of Kaggle so we opted to go for the cheap route. Glad to hear some others had some cool ideas. Also interesting to hear how far some made it on 50 series cards. Next time I might have to put my 4080 super through the ringer haha", "votes": null }, { "id": "3237271", "postDate": "07/01/2025 02:29:52", "content": "By just finetuning the bartley's caformer with pseudo labels (about 65k samples), I can get 24.3.
", "rawMarkdown": "By just finetuning the bartley's caformer with pseudo labels (about 65k samples), I can get 24.3.", "votes": null }, { "id": "3237278", "postDate": "07/01/2025 02:35:03", "content": "Wanted to try this to edge into silver and just never got to it! Nice work!
", "rawMarkdown": "Wanted to try this to edge into silver and just never got to it! Nice work!", "votes": null }, { "id": "3238064", "postDate": "07/01/2025 14:06:50", "content": "Thank you for your comment. I will study the other discussion comments to see how far I could improve my score with Bartley's notebook.
", "rawMarkdown": "Thank you for your comment. I will study the other discussion comments to see how far I could improve my score with Bartley's notebook.", "votes": null }, { "id": "3238073", "postDate": "07/01/2025 14:17:11", "content": "I also applied noise reduction to the published csv and filtered it like image processing. But I couldn't come up with a clustering algorithm and thought this was smarter!
", "rawMarkdown": "I also applied noise reduction to the published csv and filtered it like image processing. But I couldn't come up with a clustering algorithm and thought this was smarter!", "votes": null } ], "comments": [ { "id": 3237234, "author_name": "haruiig", "author_url": "", "post_date": "07/01/2025 01:32:03", "content": "As a result, clustering to improve 25.6 solution to 25.4 didn't need GPUs and it deserve bronze medals :https://www.kaggle.com/competitions/waveform-inversion/discussion/587395
\nBut, I needed GPUs and TPUs to train DNN models and I rely on Google's TPU Research Cloud program, that offers TPUs for free.
", "votes": null, "replies": [ { "id": 3238073, "author_name": "shunsukehayashi1993", "author_url": "", "post_date": "07/01/2025 14:17:11", "content": "I also applied noise reduction to the published csv and filtered it like image processing. But I couldn't come up with a clustering algorithm and thought this was smarter!
", "votes": null, "replies": [] } ] }, { "id": 3237262, "author_name": "cody11null", "author_url": "", "post_date": "07/01/2025 02:22:23", "content": "I built a quick classifier early on for predicting the family because I saw that scores were so different between them. I had thought optimizing score on say Curve_Fault_B vs some of the others may be the breakthrough. It seems that split was not needed which I later found. Instead however, we were able to use the models that were made public and optimized weights based off families with these submissions. We figured there would be little shake with the amount of data available. We decided early on that the amount of compute resources did not align with our goals outside of Kaggle so we opted to go for the cheap route. Glad to hear some others had some cool ideas. Also interesting to hear how far some made it on 50 series cards. Next time I might have to put my 4080 super through the ringer haha
", "votes": null, "replies": [] }, { "id": 3237271, "author_name": "i2nfinit3y", "author_url": "", "post_date": "07/01/2025 02:29:52", "content": "By just finetuning the bartley's caformer with pseudo labels (about 65k samples), I can get 24.3.
", "votes": null, "replies": [ { "id": 3237278, "author_name": "cody11null", "author_url": "", "post_date": "07/01/2025 02:35:03", "content": "Wanted to try this to edge into silver and just never got to it! Nice work!
", "votes": null, "replies": [] }, { "id": 3238064, "author_name": "shunsukehayashi1993", "author_url": "", "post_date": "07/01/2025 14:06:50", "content": "Thank you for your comment. I will study the other discussion comments to see how far I could improve my score with Bartley's notebook.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3237216": "Hi everyone,\nThis is my first post here—please let me know if anything is off in terms of Kaggle etiquette or norms.\n\nI tried running Bartley’s notebook and attempted some training, but the GPU computation required was quite intense. Eventually, I gave up on training the model myself, as it didn’t seem feasible to surpass a public score of 25.6 with my approach (I’m still a beginner, so it was a valuable learning experience regardless).\n\nI'm curious—\nFor those who achieved silver or bronze medals, did you rely mainly on the 45-hour/week Kaggle GPU for training? Or was it necessary to use your own GPU resources, collaborate with a team, or find other workarounds for the compute demands?\n\nWould love to hear your experience—thanks in advance!", "3237234": "As a result, clustering to improve 25.6 solution to 25.4 didn't need GPUs and it deserve bronze medals :https://www.kaggle.com/competitions/waveform-inversion/discussion/587395\n\nBut, I needed GPUs and TPUs to train DNN models and I rely on Google's TPU Research Cloud program, that offers TPUs for free.", "3237262": "I built a quick classifier early on for predicting the family because I saw that scores were so different between them. I had thought optimizing score on say Curve_Fault_B vs some of the others may be the breakthrough. It seems that split was not needed which I later found. Instead however, we were able to use the models that were made public and optimized weights based off families with these submissions. We figured there would be little shake with the amount of data available. We decided early on that the amount of compute resources did not align with our goals outside of Kaggle so we opted to go for the cheap route. Glad to hear some others had some cool ideas. Also interesting to hear how far some made it on 50 series cards. Next time I might have to put my 4080 super through the ringer haha", "3237271": "By just finetuning the bartley's caformer with pseudo labels (about 65k samples), I can get 24.3.", "3237278": "Wanted to try this to edge into silver and just never got to it! Nice work!", "3238064": "Thank you for your comment. I will study the other discussion comments to see how far I could improve my score with Bartley's notebook.", "3238073": "I also applied noise reduction to the published csv and filtered it like image processing. But I couldn't come up with a clustering algorithm and thought this was smarter!" }, "source": "meta" }