{ "id": 129871, "title": "Can we use TPU to train models with PyTorch?", "url": "/competitions/deepfake-detection-challenge/discussion/129871", "author_name": "Hieu Phung", "post_date": "2020-02-11T04:17:53.851000", "votes": 8, "comment_count": 9, "views": 0, "content": "

So this morning, according to my time-zone, Kaggle've added TPU accelerator option in everyone's notebooks and each kaggler has provided 30 hours/week quota for using it. I know the easiest and most effective way to take advantage of TPU is constructing and training models by using Tensorflow or Keras; however, can someone who still wants to stick to using PyTorch utilize the TPU's power?

\n\n

UPDATE: Thanks to @adityaecdrid, we are coming closer to being able to use TPU with PyTorch. For more detail, let read the instructions here.

", "messages": [ { "id": 742083, "postDate": "2020-02-11T04:17:53.850Z", "content": "

So this morning, according to my time-zone, Kaggle've added TPU accelerator option in everyone's notebooks and each kaggler has provided 30 hours/week quota for using it. I know the easiest and most effective way to take advantage of TPU is constructing and training models by using Tensorflow or Keras; however, can someone who still wants to stick to using PyTorch utilize the TPU's power?

\n\n

UPDATE: Thanks to @adityaecdrid, we are coming closer to being able to use TPU with PyTorch. For more detail, let read the instructions here.

", "rawMarkdown": "So this morning, according to my time-zone, Kaggle've added TPU accelerator option in everyone's notebooks and each kaggler has provided 30 hours/week quota for using it. I know the easiest and most effective way to take advantage of TPU is constructing and training models by using Tensorflow or Keras; however, can someone who still wants to stick to using PyTorch utilize the TPU's power?\n\nUPDATE: Thanks to @adityaecdrid, we are coming closer to being able to use TPU with PyTorch. For more detail, let read the instructions [here](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/129820).", "votes": 8 }, { "id": 743328, "postDate": "2020-02-12T01:47:58.413Z", "content": "

Yes, it looks like others have gotten it to work in Kaggle Notebooks on the Flower Classification competition.

", "rawMarkdown": "Yes, it looks like others have [gotten it to work in Kaggle Notebooks](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/129820) on the Flower Classification competition.", "votes": 1, "replies": [ { "id": 743335, "postDate": "2020-02-12T01:56:19.633Z", "content": "

Looks promising! Thanks for his work, we can successfully install PyTorch-XLA in a kernel; however, it's not being able to use for now. Maybe, we need some more configurations and/or official support from Kaggle.

", "rawMarkdown": "Looks promising! Thanks for his work, we can successfully install PyTorch-XLA in a kernel; however, it's not being able to use for now. Maybe, we need some more configurations and/or official support from Kaggle." }, { "id": 744299, "postDate": "2020-02-12T17:54:01.227Z", "content": "

No unfortunately they have not. PyTorch support for TPUs is not something we have on Kaggle yet.

", "rawMarkdown": "No unfortunately they have not. PyTorch support for TPUs is not something we have on Kaggle yet.", "votes": 1 }, { "id": 744682, "postDate": "2020-02-13T03:12:24.117Z", "content": "

Supporting in the near future, maybe?

", "rawMarkdown": "Supporting in the near future, maybe?" } ] }, { "id": 742181, "postDate": "2020-02-11T06:07:39.147Z", "content": "

I remember once there is a news that we can use Google's TPU with pytorch .

\n\n

view this github

", "rawMarkdown": "I remember once there is a news that we can use Google's TPU with pytorch .\n\n[view this github ](https://github.com/pytorch/xla)", "votes": 1, "replies": [ { "id": 742542, "postDate": "2020-02-11T11:15:53.983Z", "content": "

Sadly, it doesn't work in Kaggle's notebooks 😂, for now.

", "rawMarkdown": "Sadly, it doesn't work in Kaggle's notebooks 😂, for now." }, { "id": 742551, "postDate": "2020-02-11T11:20:24.740Z", "content": "

Oh, that's so sad .

", "rawMarkdown": "Oh, that's so sad ." } ] }, { "id": 770144, "postDate": "2020-03-12T16:02:21.223Z", "content": "

More info here: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/133777\nI'll be posting updates there as work on PyTorch-XLA continues.

", "rawMarkdown": "More info here: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/133777\nI'll be posting updates there as work on PyTorch-XLA continues.", "votes": 1 }, { "id": 769565, "postDate": "2020-03-12T02:34:37.057Z", "content": "

I got same question since GPU credits have been used up.

", "rawMarkdown": "I got same question since GPU credits have been used up.", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 743328, "author_name": "Julia Elliott", "author_url": "", "post_date": "2020-02-12T01:47:58.413000", "content": "

Yes, it looks like others have gotten it to work in Kaggle Notebooks on the Flower Classification competition.

", "votes": 1, "replies": [ { "id": 743335, "author_name": "Hieu Phung", "author_url": "", "post_date": "2020-02-12T01:56:19.633000", "content": "

Looks promising! Thanks for his work, we can successfully install PyTorch-XLA in a kernel; however, it's not being able to use for now. Maybe, we need some more configurations and/or official support from Kaggle.

", "votes": 0, "replies": [] }, { "id": 744299, "author_name": "Martin Görner", "author_url": "", "post_date": "2020-02-12T17:54:01.227000", "content": "

No unfortunately they have not. PyTorch support for TPUs is not something we have on Kaggle yet.

", "votes": 1, "replies": [] }, { "id": 744682, "author_name": "Hieu Phung", "author_url": "", "post_date": "2020-02-13T03:12:24.117000", "content": "

Supporting in the near future, maybe?

", "votes": 0, "replies": [] } ] }, { "id": 742181, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2020-02-11T06:07:39.147000", "content": "

I remember once there is a news that we can use Google's TPU with pytorch .

\n\n

view this github

", "votes": 1, "replies": [ { "id": 742542, "author_name": "Hieu Phung", "author_url": "", "post_date": "2020-02-11T11:15:53.983000", "content": "

Sadly, it doesn't work in Kaggle's notebooks 😂, for now.

", "votes": 0, "replies": [] }, { "id": 742551, "author_name": "哈尔的移动城堡", "author_url": "", "post_date": "2020-02-11T11:20:24.740000", "content": "

Oh, that's so sad .

", "votes": 0, "replies": [] } ] }, { "id": 770144, "author_name": "Martin Görner", "author_url": "", "post_date": "2020-03-12T16:02:21.223000", "content": "

More info here: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/133777\nI'll be posting updates there as work on PyTorch-XLA continues.

", "votes": 1, "replies": [] }, { "id": 769565, "author_name": "", "author_url": "", "post_date": "2020-03-12T02:34:37.057000", "content": "

I got same question since GPU credits have been used up.

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "742083": "So this morning, according to my time-zone, Kaggle've added TPU accelerator option in everyone's notebooks and each kaggler has provided 30 hours/week quota for using it. I know the easiest and most effective way to take advantage of TPU is constructing and training models by using Tensorflow or Keras; however, can someone who still wants to stick to using PyTorch utilize the TPU's power?\n\nUPDATE: Thanks to @adityaecdrid, we are coming closer to being able to use TPU with PyTorch. For more detail, let read the instructions [here](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/129820).", "743328": "Yes, it looks like others have [gotten it to work in Kaggle Notebooks](https://www.kaggle.com/c/flower-classification-with-tpus/discussion/129820) on the Flower Classification competition.", "742181": "I remember once there is a news that we can use Google's TPU with pytorch .\n\n[view this github ](https://github.com/pytorch/xla)", "770144": "More info here: https://www.kaggle.com/c/flower-classification-with-tpus/discussion/133777\nI'll be posting updates there as work on PyTorch-XLA continues.", "769565": "I got same question since GPU credits have been used up." } }