We'd like to invite everyone who has made meaningful contributions to sharing and education around use of TPUs to self-nominate for the TPU Star awards!
\n\nTo enter, please submit this form by July 8th 2020: https://www.kaggle.com/tpu-star-submission-jigsawmulti
\n\nSubmissions will then be evaluated by our in-house group of TPU experts, and we aim to announce & distribute those awards by the end of July!
", "messages": [ { "id": "897502", "postDate": "06/22/2020 23:01:13", "content": "We'd like to invite everyone who has made meaningful contributions to sharing and education around use of TPUs to self-nominate for the TPU Star awards!
\n\nTo enter, please submit this form by July 8th 2020: https://www.kaggle.com/tpu-star-submission-jigsawmulti
\n\nSubmissions will then be evaluated by our in-house group of TPU experts, and we aim to announce & distribute those awards by the end of July!
", "rawMarkdown": "We'd like to invite everyone who has made meaningful contributions to sharing and education around use of TPUs to self-nominate for the TPU Star awards!\n\nTo enter, please submit this form by July 8th 2020: https://www.kaggle.com/tpu-star-submission-jigsawmulti\n\nSubmissions will then be evaluated by our in-house group of TPU experts, and we aim to announce & distribute those awards by the end of July!", "votes": null }, { "id": "897629", "postDate": "06/23/2020 02:09:41", "content": "I upvote @shonenkov. Without your colab file, we are unable to finish with 19th.
", "rawMarkdown": "I upvote @shonenkov. Without your colab file, we are unable to finish with 19th.", "votes": null }, { "id": "898331", "postDate": "06/23/2020 13:07:18", "content": "My opinion, the one of an active user of others' kernels, on the contributions for harnessing TPU in the Jigsaw comp.:
\n\nDezso Ribli @riblidezso - the only one to open the door for many to train Roberta-XLM-Large with tf2 on free Colab with [Colab] Train XLM-R large with TPU v2-8 on Colab, and the only one to share finetuning MLM model on test set - Finetune XLM-Roberta on Jigsaw test data with MLM
Alex Shonenkov @shonenkov with Pytorch on TPU with data augmentation [TPU-Training] Super Fast XLMRoberta
DimitreOliveira @dimitreoliveira - the first to help others to use tf2 custom training loop with Jigsaw - TPU optimized training loops
Abhishek @abhishek with his invaluable 'real-time coding' youtube videos and I Like Clean TPU Training Kernels & I Can Not Lie
Xhlulu @xhlulu - Jigsaw TPU: XLM-Roberta
Michael Kazachok @miklgr500 with Jigsaw TPU: BERT with Huggingface and Keras
Apologies, I'd set the wrong deadline for TPU star submissions - it's been updated to the correct date: July 8th.
", "rawMarkdown": "Apologies, I'd set the wrong deadline for TPU star submissions - it's been updated to the correct date: July 8th.", "votes": null }, { "id": "898605", "postDate": "06/23/2020 16:13:33", "content": "Personally, I would suggest those who have managed to bring TPU functionality to Kaggle kernels and who shared both training and inference. If I would need to choose two winners I would pick:
\n\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta \n@xhlulu who was the first to show the utility of XLM-Roberta-Large and managed to fit it in kaggle kernel
\n\nhttps://www.kaggle.com/riblidezso/train-from-mlm-finetuned-xlm-roberta-large\n@riblidezso who published a very clean and well documented series of TPU kernels
\n\nEven though both of them are Tensorflow and not Pytorch :)
", "rawMarkdown": "Personally, I would suggest those who have managed to bring TPU functionality to Kaggle kernels and who shared both training and inference. If I would need to choose two winners I would pick:\n\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta \n@xhlulu who was the first to show the utility of XLM-Roberta-Large and managed to fit it in kaggle kernel\n\nhttps://www.kaggle.com/riblidezso/train-from-mlm-finetuned-xlm-roberta-large\n@riblidezso who published a very clean and well documented series of TPU kernels\n\nEven though both of them are Tensorflow and not Pytorch :)", "votes": null }, { "id": "898722", "postDate": "06/23/2020 17:32:55", "content": "@xhlulu 🙌 🙌 \nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta\nIt was his notebook that got my eye to using XML-Roberta-Large for this competition.
", "rawMarkdown": "xhlulu 🙌 🙌 \nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta\nIt was his notebook that got my eye to using XML-Roberta-Large for this competition.", "votes": null }, { "id": "899005", "postDate": "06/23/2020 22:20:46", "content": "Thanks a lot to these people:
\n\nYassine Alouini @yassinealouini
\nfor a nice introduction to TPU usage, etc. \nhttps://medium.com/@YassineAlouini/tpus-101-493301117eec\nhttps://medium.com/@YassineAlouini/roberta-meets-tpus-af839ce7c070
Xhlulu @xhlulu \nfor the XLM-RoBERTa kernel which let us start: Jigsaw TPU: XLM-Roberta
\n\nAlex Shonenkov @shonenkov \nfor shedding light on PyTorch/XLA issues around TPU usage [TPU-Training] Super Fast XLMRoberta
\n\nI am sorry if I forgot anyone who also (for sure) deserves the award.
", "rawMarkdown": "Thanks a lot to these people:\n\nYassine Alouini @yassinealouini \nfor a nice introduction to TPU usage, etc. \nhttps://medium.com/@YassineAlouini/tpus-101-493301117eec\nhttps://medium.com/@YassineAlouini/roberta-meets-tpus-af839ce7c070\n\nXhlulu @xhlulu \nfor the XLM-RoBERTa kernel which let us start: [Jigsaw TPU: XLM-Roberta](https://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta)\n\nAlex Shonenkov @shonenkov \nfor shedding light on PyTorch/XLA issues around TPU usage [ [TPU-Training] Super Fast XLMRoberta](https://www.kaggle.com/shonenkov/tpu-training-super-fast-xlmroberta)\n\nI am sorry if I forgot anyone who also (for sure) deserves the award.", "votes": null }, { "id": "901079", "postDate": "06/25/2020 08:42:07", "content": "Don't forget to mention those who initially sparked it all :)
", "rawMarkdown": "Don't forget to mention those who initially sparked it all :)", "votes": null }, { "id": "901112", "postDate": "06/25/2020 09:02:53", "content": "Absolutely true @adityaecdrid
", "rawMarkdown": "Absolutely true @adityaecdrid", "votes": null }, { "id": "901394", "postDate": "06/25/2020 12:45:34", "content": "Agreed
", "rawMarkdown": "Agreed", "votes": null }, { "id": "903402", "postDate": "06/26/2020 20:00:54", "content": "Agree as well 👍
", "rawMarkdown": "Agree as well 👍", "votes": null }, { "id": "909580", "postDate": "06/30/2020 17:07:22", "content": "I totally agree with Psi :-).
", "rawMarkdown": "I totally agree with Psi :-).", "votes": null }, { "id": "909583", "postDate": "06/30/2020 17:08:57", "content": "I vote for xhlulu🙏 :\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta
", "rawMarkdown": "I vote for xhlulu🙏 :\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta", "votes": null } ], "comments": [ { "id": 897629, "author_name": "veryrobustperson", "author_url": "", "post_date": "06/23/2020 02:09:41", "content": "I upvote @shonenkov. Without your colab file, we are unable to finish with 19th.
", "votes": null, "replies": [] }, { "id": 898331, "author_name": "isakev", "author_url": "", "post_date": "06/23/2020 13:07:18", "content": "My opinion, the one of an active user of others' kernels, on the contributions for harnessing TPU in the Jigsaw comp.:
\n\nDezso Ribli @riblidezso - the only one to open the door for many to train Roberta-XLM-Large with tf2 on free Colab with [Colab] Train XLM-R large with TPU v2-8 on Colab, and the only one to share finetuning MLM model on test set - Finetune XLM-Roberta on Jigsaw test data with MLM
Alex Shonenkov @shonenkov with Pytorch on TPU with data augmentation [TPU-Training] Super Fast XLMRoberta
DimitreOliveira @dimitreoliveira - the first to help others to use tf2 custom training loop with Jigsaw - TPU optimized training loops
Abhishek @abhishek with his invaluable 'real-time coding' youtube videos and I Like Clean TPU Training Kernels & I Can Not Lie
Xhlulu @xhlulu - Jigsaw TPU: XLM-Roberta
Michael Kazachok @miklgr500 with Jigsaw TPU: BERT with Huggingface and Keras
Apologies, I'd set the wrong deadline for TPU star submissions - it's been updated to the correct date: July 8th.
", "votes": null, "replies": [] }, { "id": 898605, "author_name": "philippsinger", "author_url": "", "post_date": "06/23/2020 16:13:33", "content": "Personally, I would suggest those who have managed to bring TPU functionality to Kaggle kernels and who shared both training and inference. If I would need to choose two winners I would pick:
\n\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta \n@xhlulu who was the first to show the utility of XLM-Roberta-Large and managed to fit it in kaggle kernel
\n\nhttps://www.kaggle.com/riblidezso/train-from-mlm-finetuned-xlm-roberta-large\n@riblidezso who published a very clean and well documented series of TPU kernels
\n\nEven though both of them are Tensorflow and not Pytorch :)
", "votes": null, "replies": [ { "id": 901394, "author_name": "sebastienm", "author_url": "", "post_date": "06/25/2020 12:45:34", "content": "Agreed
", "votes": null, "replies": [] }, { "id": 903402, "author_name": "yeayates21", "author_url": "", "post_date": "06/26/2020 20:00:54", "content": "Agree as well 👍
", "votes": null, "replies": [] } ] }, { "id": 898722, "author_name": "namanj27", "author_url": "", "post_date": "06/23/2020 17:32:55", "content": "@xhlulu 🙌 🙌 \nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta\nIt was his notebook that got my eye to using XML-Roberta-Large for this competition.
", "votes": null, "replies": [] }, { "id": 899005, "author_name": "muhakabartay", "author_url": "", "post_date": "06/23/2020 22:20:46", "content": "Thanks a lot to these people:
\n\nYassine Alouini @yassinealouini
\nfor a nice introduction to TPU usage, etc. \nhttps://medium.com/@YassineAlouini/tpus-101-493301117eec\nhttps://medium.com/@YassineAlouini/roberta-meets-tpus-af839ce7c070
Xhlulu @xhlulu \nfor the XLM-RoBERTa kernel which let us start: Jigsaw TPU: XLM-Roberta
\n\nAlex Shonenkov @shonenkov \nfor shedding light on PyTorch/XLA issues around TPU usage [TPU-Training] Super Fast XLMRoberta
\n\nI am sorry if I forgot anyone who also (for sure) deserves the award.
", "votes": null, "replies": [] }, { "id": 901079, "author_name": "adityaecdrid", "author_url": "", "post_date": "06/25/2020 08:42:07", "content": "Don't forget to mention those who initially sparked it all :)
", "votes": null, "replies": [ { "id": 901112, "author_name": "namanj27", "author_url": "", "post_date": "06/25/2020 09:02:53", "content": "Absolutely true @adityaecdrid
", "votes": null, "replies": [] } ] }, { "id": 909580, "author_name": "xiwuhan", "author_url": "", "post_date": "06/30/2020 17:07:22", "content": "I totally agree with Psi :-).
", "votes": null, "replies": [] }, { "id": 909583, "author_name": "xiwuhan", "author_url": "", "post_date": "06/30/2020 17:08:57", "content": "I vote for xhlulu🙏 :\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "897502": "We'd like to invite everyone who has made meaningful contributions to sharing and education around use of TPUs to self-nominate for the TPU Star awards!\n\nTo enter, please submit this form by July 8th 2020: https://www.kaggle.com/tpu-star-submission-jigsawmulti\n\nSubmissions will then be evaluated by our in-house group of TPU experts, and we aim to announce & distribute those awards by the end of July!", "897629": "I upvote @shonenkov. Without your colab file, we are unable to finish with 19th.", "898331": "My opinion, the one of an active user of others' kernels, on the contributions for harnessing TPU in the Jigsaw comp.:\n\n- Dezso Ribli @riblidezso - the only one to open the door for many to train Roberta-XLM-Large with tf2 on free Colab with [[Colab] Train XLM-R large with TPU v2-8 on Colab](https://www.kaggle.com/riblidezso/colab-train-xlm-r-large-with-tpu-v2-8-on-colab), and the only one to share finetuning MLM model on test set - [Finetune XLM-Roberta on Jigsaw test data with MLM](https://www.kaggle.com/riblidezso/finetune-xlm-roberta-on-jigsaw-test-data-with-mlm)\n\n- Alex Shonenkov @shonenkov with Pytorch on TPU with data augmentation [[TPU-Training] Super Fast XLMRoberta](https://www.kaggle.com/shonenkov/tpu-training-super-fast-xlmroberta)\n\n- DimitreOliveira @dimitreoliveira - the first to help others to use tf2 custom training loop with [Jigsaw - TPU optimized training loops](https://www.kaggle.com/dimitreoliveira/jigsaw-tpu-optimized-training-loops)\n\n- Abhishek @abhishek with his invaluable 'real-time coding' youtube videos and [I Like Clean TPU Training Kernels & I Can Not Lie](https://www.kaggle.com/abhishek/i-like-clean-tpu-training-kernels-i-can-not-lie)\n\n- Xhlulu @xhlulu - [Jigsaw TPU: XLM-Roberta](https://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta)\n\n- Michael Kazachok @miklgr500 with [Jigsaw TPU: BERT with Huggingface and Keras](https://www.kaggle.com/miklgr500/jigsaw-tpu-bert-with-huggingface-and-keras)", "898592": "Apologies, I'd set the wrong deadline for TPU star submissions - it's been updated to the correct date: July 8th.", "898605": "Personally, I would suggest those who have managed to bring TPU functionality to Kaggle kernels and who shared both training and inference. If I would need to choose two winners I would pick:\n\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta \n@xhlulu who was the first to show the utility of XLM-Roberta-Large and managed to fit it in kaggle kernel\n\nhttps://www.kaggle.com/riblidezso/train-from-mlm-finetuned-xlm-roberta-large\n@riblidezso who published a very clean and well documented series of TPU kernels\n\nEven though both of them are Tensorflow and not Pytorch :)", "898722": "xhlulu 🙌 🙌 \nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta\nIt was his notebook that got my eye to using XML-Roberta-Large for this competition.", "899005": "Thanks a lot to these people:\n\nYassine Alouini @yassinealouini \nfor a nice introduction to TPU usage, etc. \nhttps://medium.com/@YassineAlouini/tpus-101-493301117eec\nhttps://medium.com/@YassineAlouini/roberta-meets-tpus-af839ce7c070\n\nXhlulu @xhlulu \nfor the XLM-RoBERTa kernel which let us start: [Jigsaw TPU: XLM-Roberta](https://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta)\n\nAlex Shonenkov @shonenkov \nfor shedding light on PyTorch/XLA issues around TPU usage [ [TPU-Training] Super Fast XLMRoberta](https://www.kaggle.com/shonenkov/tpu-training-super-fast-xlmroberta)\n\nI am sorry if I forgot anyone who also (for sure) deserves the award.", "901079": "Don't forget to mention those who initially sparked it all :)", "901112": "Absolutely true @adityaecdrid", "901394": "Agreed", "903402": "Agree as well 👍", "909580": "I totally agree with Psi :-).", "909583": "I vote for xhlulu🙏 :\nhttps://www.kaggle.com/xhlulu/jigsaw-tpu-xlm-roberta" }, "source": "meta" }