![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F7db639da289e00a67fac54c8e855ff75%2FSelection_999(3038).png?generation=1693789004947926&alt=media\")
to be updated as i progress along.
\nthe first steps are to repeat papers results.
\ne.g. GraphSAGE, segmented graph training …etc
to be updated as i progress along.
\nthe first steps are to repeat papers results.
\ne.g. GraphSAGE, segmented graph training …etc
Suggestion for pure beginner
\n[1] start with GraphSAGE (there are many resources)
\n[2] learned generally about computational graph compilation
\n(faster way is to read blogs and youtube)
[3] go to tensorflow XLA and HLO website and read generally
\n[4] Read competition papers and repeat results (use the github baseline to cross check your paper understanding and implementation)
\nif the workload is too heavy, find some buddy … study in a group …
\n[2],[3] are domain knowledge, so just read briefly.
\nfocus on [1],[4]
To start with GNNs, I just started with that blog and it's pretty nice:
\nhttps://mlabonne.github.io/blog/posts/2022-03-09-Graph_Attention_Network.html
", "rawMarkdown": "To start with GNNs, I just started with that blog and it's pretty nice:\n\nhttps://mlabonne.github.io/blog/posts/2022-03-09-Graph_Attention_Network.html\n", "votes": 2, "replies": [ { "id": 2466253, "postDate": "2023-10-03T17:06:11.730Z", "content": "have you did any changes or as it is you wrote
", "rawMarkdown": "have you did any changes or as it is you wrote", "votes": 1 } ] } ] }, { "id": 2515323, "postDate": "2023-11-06T21:10:03.217Z", "content": "xla-default cross validation results:
\ni haven't made submission yet, not sure if the results are correct.
\nmy models are modification from [1], the reference 3xSAGEConv with my own gnn normalisation
[1]Learning Large Graph Property Prediction via Graph Segment Training
\nhttps://github.com/kaidic/GST/tree/main
Are these kendall Tau score calculated using entire config runtime?? Amazing scores.
", "rawMarkdown": "Are these kendall Tau score calculated using entire config runtime?? Amazing scores. ", "replies": [ { "id": 2516880, "postDate": "2023-11-08T03:14:04.093Z", "content": "I assume these are Pearson, which would mean the Kendalls would be even higher
", "rawMarkdown": "I assume these are Pearson, which would mean the Kendalls would be even higher", "votes": 1 }, { "id": 2517172, "postDate": "2023-11-08T09:06:48.237Z", "content": "\"Are these kendall Tau score calculated using entire config runtime?? Amazing scores.\"
\nyes.
i think these are typical scores.
\noverall, i got current lb 0.675 with this
truth = result\n predict = result\n corr, _ = (predict, truth)\n\n (, f,f,)\nThanks @hengck23. Impressive Cv and nice jump on leaderboard.
", "rawMarkdown": "Thanks @hengck23. Impressive Cv and nice jump on leaderboard. ", "votes": 2 }, { "id": 2520895, "postDate": "2023-11-11T09:10:52.990Z", "content": "normalisation normalisation normalisation normalisation …. 😀
", "rawMarkdown": "normalisation normalisation normalisation normalisation .... 😀" } ] } ] }, { "id": 2424524, "postDate": "2023-09-05T10:54:13.193Z", "content": "Thanks for sharing, very useful information
", "rawMarkdown": "Thanks for sharing, very useful information" }, { "id": 2422358, "postDate": "2023-09-04T00:39:00.763Z", "content": "i am also working on a tool to visualize the results, e.g. (like dot for graphviz, or html) to draw the best configuration + tensor graph.
", "rawMarkdown": "i am also working on a tool to visualize the results, e.g. (like dot for graphviz, or html) to draw the best configuration + tensor graph." }, { "id": 2422357, "postDate": "2023-09-04T00:36:13.233Z", "content": "here are some good papers related to model runtime estimation for HLO:
\n1) OPTIMIZING DNN COMPILATION FOR DISTRIBUTED TRAINING WITH JOINT OP AND TENSOR FUSION
\nhttps://arxiv.org/pdf/2209.12769.pdf
http://mlforsystems.org/assets/papers/neurips2019/learned_tpu_kaufman_2019.pdf
\nLearned TPU Cost Model for XLA Tensor Programs
Suggestion for pure beginner
\n[1] start with GraphSAGE (there are many resources)
\n[2] learned generally about computational graph compilation
\n(faster way is to read blogs and youtube)
[3] go to tensorflow XLA and HLO website and read generally
\n[4] Read competition papers and repeat results (use the github baseline to cross check your paper understanding and implementation)
\nif the workload is too heavy, find some buddy … study in a group …
\n[2],[3] are domain knowledge, so just read briefly.
\nfocus on [1],[4]
To start with GNNs, I just started with that blog and it's pretty nice:
\nhttps://mlabonne.github.io/blog/posts/2022-03-09-Graph_Attention_Network.html
", "votes": 2, "replies": [ { "id": 2466253, "author_name": "Satheesh Bhukya", "author_url": "", "post_date": "2023-10-03T17:06:11.730000", "content": "have you did any changes or as it is you wrote
", "votes": 1, "replies": [] } ] } ] }, { "id": 2515323, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-06T21:10:03.217000", "content": "xla-default cross validation results:
\ni haven't made submission yet, not sure if the results are correct.
\nmy models are modification from [1], the reference 3xSAGEConv with my own gnn normalisation
[1]Learning Large Graph Property Prediction via Graph Segment Training
\nhttps://github.com/kaidic/GST/tree/main
Are these kendall Tau score calculated using entire config runtime?? Amazing scores.
", "votes": 0, "replies": [ { "id": 2516880, "author_name": "Rob Freeman", "author_url": "", "post_date": "2023-11-08T03:14:04.093000", "content": "I assume these are Pearson, which would mean the Kendalls would be even higher
", "votes": 1, "replies": [] }, { "id": 2517172, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-08T09:06:48.237000", "content": "\"Are these kendall Tau score calculated using entire config runtime?? Amazing scores.\"
\nyes.
i think these are typical scores.
\noverall, i got current lb 0.675 with this
truth = result\n predict = result\n corr, _ = (predict, truth)\n\n (, f,f,)\nThanks @hengck23. Impressive Cv and nice jump on leaderboard.
", "votes": 2, "replies": [] }, { "id": 2520895, "author_name": "hengck23", "author_url": "", "post_date": "2023-11-11T09:10:52.990000", "content": "normalisation normalisation normalisation normalisation …. 😀
", "votes": 0, "replies": [] } ] } ] }, { "id": 2424524, "author_name": "Yu Wu", "author_url": "", "post_date": "2023-09-05T10:54:13.193000", "content": "Thanks for sharing, very useful information
", "votes": 0, "replies": [] }, { "id": 2422358, "author_name": "hengck23", "author_url": "", "post_date": "2023-09-04T00:39:00.763000", "content": "i am also working on a tool to visualize the results, e.g. (like dot for graphviz, or html) to draw the best configuration + tensor graph.
", "votes": 0, "replies": [] }, { "id": 2422357, "author_name": "hengck23", "author_url": "", "post_date": "2023-09-04T00:36:13.233000", "content": "here are some good papers related to model runtime estimation for HLO:
\n1) OPTIMIZING DNN COMPILATION FOR DISTRIBUTED TRAINING WITH JOINT OP AND TENSOR FUSION
\nhttps://arxiv.org/pdf/2209.12769.pdf
http://mlforsystems.org/assets/papers/neurips2019/learned_tpu_kaufman_2019.pdf
\nLearned TPU Cost Model for XLA Tensor Programs