![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3736535%2F0a6d2e19631639f330f8a2ba170f0f7a%2Flocal_3days.png?generation=1749517526222321&alt=media\")
Hello! I have 9 hours available this week, and I have a single 4090 GPU, so I can run things locally. How about teaming up and working together?I have two codes for this event. GPU on Kaggle only allows for nine hours which is insufficient. I hope someone having GPU will work with me. Is there anybody, please?
", "messages": [ { "id": 3220349, "postDate": "2025-06-09T07:34:00.920Z", "content": "I have two codes for this event. GPU on Kaggle only allows for nine hours which is insufficient. I hope someone having GPU will work with me. Is there anybody, please?
", "rawMarkdown": "I have two codes for this event. GPU on Kaggle only allows for nine hours which is insufficient. I hope someone having GPU will work with me. Is there anybody, please? ", "votes": 3 }, { "id": 3221749, "postDate": "2025-06-11T12:01:19.477Z", "content": "https://salad.com/pricing
\nhttps://vast.ai/?gad_campaignid=20733719727
\nYou can rent a GPU cheaply using the website at the URL above
Hello! I have 9 hours available this week, and I have a single 4090 GPU, so I can run things locally. How about teaming up and working together?
For a single 4090 GPU i do not know exactly necessary time for compute.
\nCan you run it about 10 days?
\nBecause 150 epochs take 8 days.
\nFrom one discussion:
\nhttps://www.kaggle.com/competitions/waveform-inversion/discussion/582785
@overvalueawareness It depends on the code and performance. In my case, each epoch took around 3 hours.
", "rawMarkdown": "@overvalueawareness It depends on the code and performance. In my case, each epoch took around 3 hours.", "votes": -1 }, { "id": 3220828, "postDate": "2025-06-10T03:16:32.413Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3220833, "postDate": "2025-06-10T03:23:45.573Z", "content": "Please check your email.
", "rawMarkdown": "Please check your email." } ] }, { "id": 3224325, "postDate": "2025-06-14T17:05:37.303Z", "content": "How did you manage to achieve around 40 MAE after just a few epochs? Would you be willing to share the training parameters you used?
", "rawMarkdown": "How did you manage to achieve around 40 MAE after just a few epochs? Would you be willing to share the training parameters you used?", "replies": [ { "id": 3224590, "postDate": "2025-06-15T06:12:44.560Z", "content": "@jamalsaeedi I'm a beginner, so I wrote the code with the help of AI using publicly available sources. I'm not sure if it will be helpful, but please feel free to use it as a reference.
\nI can share the training parameters and strategies used to achieve strong performance in Full Waveform Inversion tasks.
\nEnsemble Configuration:
\nData Augmentation:
\nLoss Function Improvements:
\nRegularization Techniques:
\nThe approach used several key techniques to achieve rapid convergence[1]:
\nThe training was conducted on the OpenFWI dataset, which contains[4][5]:
\nThe comprehensive strategy was designed to achieve[1]:
\nTotal expected improvement: 0.12-0.33 MAE points beyond the baseline
\nIf you're seeing MAE values around 40 in early epochs, this likely represents the initial high error that rapidly decreases as the model learns the seismic-to-velocity mapping. The rapid convergence comes from using pre-trained features and the comprehensive training strategy outlined above.
\n[1] paste.txt
\n[2] https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fm-tuning-parameters.html?context=wx
\n[3] https://papers.nips.cc/paper_files/paper/2022/file/27d3ef263c7cb8d542c4f9815a49b69b-Supplemental-Datasets_and_Benchmarks.pdf
\n[4] https://arxiv.org/html/2412.19510v1
\n[5] https://www.nature.com/articles/s41598-024-68573-7
\n[6] https://openreview.net/forum?id=0nJt9aVGtl
\n[7] https://scispace.com/pdf/openfwi-large-scale-multi-structural-benchmark-datasets-for-3rg477j7.pdf
\n[8] https://colab.research.google.com/drive/17s5JmVs9ABl8MpmFlhWMSslj9_d5Atfx?usp=sharing
\n[9] https://openfwi-lanl.github.io/tutorial/
\n[10] https://pubmed.ncbi.nlm.nih.gov/39198496/
\n[11] https://docs.oracle.com/en-us/iaas/Content/generative-ai/fine-tuning-parameters.htm
you have 30 hours of gPU per week on Kaggle, not 9.
", "rawMarkdown": "you have 30 hours of gPU per week on Kaggle, not 9.", "replies": [ { "id": 3220699, "postDate": "2025-06-09T18:23:08.567Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 3221749, "author_name": "t fuku", "author_url": "", "post_date": "2025-06-11T12:01:19.477000", "content": "https://salad.com/pricing
\nhttps://vast.ai/?gad_campaignid=20733719727
\nYou can rent a GPU cheaply using the website at the URL above
Hello! I have 9 hours available this week, and I have a single 4090 GPU, so I can run things locally. How about teaming up and working together?
For a single 4090 GPU i do not know exactly necessary time for compute.
\nCan you run it about 10 days?
\nBecause 150 epochs take 8 days.
\nFrom one discussion:
\nhttps://www.kaggle.com/competitions/waveform-inversion/discussion/582785
@overvalueawareness It depends on the code and performance. In my case, each epoch took around 3 hours.
", "votes": -1, "replies": [] }, { "id": 3220828, "author_name": "", "author_url": "", "post_date": "2025-06-10T03:16:32.413000", "content": "", "votes": 0, "replies": [] }, { "id": 3220833, "author_name": "Focus", "author_url": "", "post_date": "2025-06-10T03:23:45.573000", "content": "Please check your email.
", "votes": 0, "replies": [] } ] }, { "id": 3224325, "author_name": "Cyrus", "author_url": "", "post_date": "2025-06-14T17:05:37.303000", "content": "How did you manage to achieve around 40 MAE after just a few epochs? Would you be willing to share the training parameters you used?
", "votes": 0, "replies": [ { "id": 3224590, "author_name": "Sung Heo", "author_url": "", "post_date": "2025-06-15T06:12:44.560000", "content": "@jamalsaeedi I'm a beginner, so I wrote the code with the help of AI using publicly available sources. I'm not sure if it will be helpful, but please feel free to use it as a reference.
\nI can share the training parameters and strategies used to achieve strong performance in Full Waveform Inversion tasks.
\nEnsemble Configuration:
\nData Augmentation:
\nLoss Function Improvements:
\nRegularization Techniques:
\nThe approach used several key techniques to achieve rapid convergence[1]:
\nThe training was conducted on the OpenFWI dataset, which contains[4][5]:
\nThe comprehensive strategy was designed to achieve[1]:
\nTotal expected improvement: 0.12-0.33 MAE points beyond the baseline
\nIf you're seeing MAE values around 40 in early epochs, this likely represents the initial high error that rapidly decreases as the model learns the seismic-to-velocity mapping. The rapid convergence comes from using pre-trained features and the comprehensive training strategy outlined above.
\n[1] paste.txt
\n[2] https://dataplatform.cloud.ibm.com/docs/content/wsj/analyze-data/fm-tuning-parameters.html?context=wx
\n[3] https://papers.nips.cc/paper_files/paper/2022/file/27d3ef263c7cb8d542c4f9815a49b69b-Supplemental-Datasets_and_Benchmarks.pdf
\n[4] https://arxiv.org/html/2412.19510v1
\n[5] https://www.nature.com/articles/s41598-024-68573-7
\n[6] https://openreview.net/forum?id=0nJt9aVGtl
\n[7] https://scispace.com/pdf/openfwi-large-scale-multi-structural-benchmark-datasets-for-3rg477j7.pdf
\n[8] https://colab.research.google.com/drive/17s5JmVs9ABl8MpmFlhWMSslj9_d5Atfx?usp=sharing
\n[9] https://openfwi-lanl.github.io/tutorial/
\n[10] https://pubmed.ncbi.nlm.nih.gov/39198496/
\n[11] https://docs.oracle.com/en-us/iaas/Content/generative-ai/fine-tuning-parameters.htm
you have 30 hours of gPU per week on Kaggle, not 9.
", "votes": 0, "replies": [ { "id": 3220699, "author_name": "", "author_url": "", "post_date": "2025-06-09T18:23:08.567000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "3220349": "I have two codes for this event. GPU on Kaggle only allows for nine hours which is insufficient. I hope someone having GPU will work with me. Is there anybody, please? ", "3221749": "https://salad.com/pricing\nhttps://vast.ai/?gad_campaignid=20733719727\nYou can rent a GPU cheaply using the website at the URL above", "3220782": "Hello! I have 9 hours available this week, and I have a single 4090 GPU, so I can run things locally. How about teaming up and working together?", "3220647": "you have 30 hours of gPU per week on Kaggle, not 9." } }