hello, I already asked this question on community forum but there was no reply https://www.kaggle.com/discussion/239733
\nbasically I am interested do you think it is still worth to use PyTorch on Kaggle competition, or should we all use Tensorflow to be able to compete here? In my last competition I was trying to use PyTorch with TPU but instead focusing on competition itself I was busy with making TPU work, do you think it is worth now, or is it better to just use tf?
", "messages": [ { "id": "1314272", "postDate": "05/19/2021 04:41:05", "content": "hello, I already asked this question on community forum but there was no reply https://www.kaggle.com/discussion/239733
\nbasically I am interested do you think it is still worth to use PyTorch on Kaggle competition, or should we all use Tensorflow to be able to compete here? In my last competition I was trying to use PyTorch with TPU but instead focusing on competition itself I was busy with making TPU work, do you think it is worth now, or is it better to just use tf?
", "rawMarkdown": "hello, I already asked this question on community forum but there was no reply https://www.kaggle.com/discussion/239733\n\nbasically I am interested do you think it is still worth to use PyTorch on Kaggle competition, or should we all use Tensorflow to be able to compete here? In my last competition I was trying to use PyTorch with TPU but instead focusing on competition itself I was busy with making TPU work, do you think it is worth now, or is it better to just use tf?", "votes": null }, { "id": "1315936", "postDate": "05/20/2021 07:46:10", "content": "Hi Jacek, I made short Pytorch tutorial here and answer the question: Keras vs Pytorch. https://www.kaggle.com/remekkinas/tps-5-pytorch-nn-for-tabular-step-by-step
\nIn TPS-05 we observed that there is only my notebook with Pytorch and about 10 in Keras. Keras is much faster to prototype.
I think majority at the top will use Pytorch even though TPU give big edge to TF.
\nThe game changer for Google will be JAX in the coming years, which has both the flexiblity of Pytorch and native support of TPU.
\nJAX can clearly overtake both Pytorch and TF when backed by Google.
hello Remek,
\nI think your tutorial is about using PyTorch in general, not about TPU, I am asking about Kaggle competitions, I was able to use PyTorch in https://www.kaggle.com/c/severstal-steel-defect-detection but after that on Kaggle TPU started to be important and in my next competition I was too busy on making PyTorch work with TPU instead focusing on competition itself, so I wonder what is the current status of that here
", "rawMarkdown": "hello Remek,\n\nI think your tutorial is about using PyTorch in general, not about TPU, I am asking about Kaggle competitions, I was able to use PyTorch in https://www.kaggle.com/c/severstal-steel-defect-detection but after that on Kaggle TPU started to be important and in my next competition I was too busy on making PyTorch work with TPU instead focusing on competition itself, so I wonder what is the current status of that here", "votes": null }, { "id": "1316086", "postDate": "05/20/2021 09:43:39", "content": "HI Jacek, I understand :) Thank you for clarification! 👍
", "rawMarkdown": "HI Jacek, I understand :) Thank you for clarification! 👍", "votes": null }, { "id": "1316307", "postDate": "05/20/2021 12:52:43", "content": "Most of them probably may use Pytorch! in my case yes I will use pytorch
", "rawMarkdown": "Most of them probably may use Pytorch! in my case yes I will use pytorch", "votes": null }, { "id": "1316327", "postDate": "05/20/2021 13:10:48", "content": "hello FenomeN,
\ndo you train your models on local GPU or do you use Kaggle TPU with PyTorch?
", "rawMarkdown": "hello FenomeN,\n\ndo you train your models on local GPU or do you use Kaggle TPU with PyTorch?", "votes": null }, { "id": "1316328", "postDate": "05/20/2021 13:11:25", "content": "Where do you train your PyTorch models?
", "rawMarkdown": "Where do you train your PyTorch models?", "votes": null }, { "id": "1316568", "postDate": "05/20/2021 16:57:13", "content": "I did not join yet this competition. But I use colab Pro.
\nWith V100 on colab Pro, I don't feel the need anymore to use TPU for Pytorch.
I had trouble in making PyTorch on TPU. I use colab GPU or Kaggle GPU for me. if i have a teammate with GPU in local i would ask help from him
", "rawMarkdown": "I had trouble in making PyTorch on TPU. I use colab GPU or Kaggle GPU for me. if i have a teammate with GPU in local i would ask help from him", "votes": null } ], "comments": [ { "id": 1315936, "author_name": "remekkinas", "author_url": "", "post_date": "05/20/2021 07:46:10", "content": "Hi Jacek, I made short Pytorch tutorial here and answer the question: Keras vs Pytorch. https://www.kaggle.com/remekkinas/tps-5-pytorch-nn-for-tabular-step-by-step
\nIn TPS-05 we observed that there is only my notebook with Pytorch and about 10 in Keras. Keras is much faster to prototype.
hello Remek,
\nI think your tutorial is about using PyTorch in general, not about TPU, I am asking about Kaggle competitions, I was able to use PyTorch in https://www.kaggle.com/c/severstal-steel-defect-detection but after that on Kaggle TPU started to be important and in my next competition I was too busy on making PyTorch work with TPU instead focusing on competition itself, so I wonder what is the current status of that here
", "votes": null, "replies": [] }, { "id": 1316086, "author_name": "remekkinas", "author_url": "", "post_date": "05/20/2021 09:43:39", "content": "HI Jacek, I understand :) Thank you for clarification! 👍
", "votes": null, "replies": [] } ] }, { "id": 1316014, "author_name": "serigne", "author_url": "", "post_date": "05/20/2021 08:44:47", "content": "I think majority at the top will use Pytorch even though TPU give big edge to TF.
\nThe game changer for Google will be JAX in the coming years, which has both the flexiblity of Pytorch and native support of TPU.
\nJAX can clearly overtake both Pytorch and TF when backed by Google.
Where do you train your PyTorch models?
", "votes": null, "replies": [] }, { "id": 1316568, "author_name": "serigne", "author_url": "", "post_date": "05/20/2021 16:57:13", "content": "I did not join yet this competition. But I use colab Pro.
\nWith V100 on colab Pro, I don't feel the need anymore to use TPU for Pytorch.
Most of them probably may use Pytorch! in my case yes I will use pytorch
", "votes": null, "replies": [ { "id": 1316327, "author_name": "jacekpoplawski", "author_url": "", "post_date": "05/20/2021 13:10:48", "content": "hello FenomeN,
\ndo you train your models on local GPU or do you use Kaggle TPU with PyTorch?
", "votes": null, "replies": [] }, { "id": 1316581, "author_name": "morizin", "author_url": "", "post_date": "05/20/2021 17:13:29", "content": "I had trouble in making PyTorch on TPU. I use colab GPU or Kaggle GPU for me. if i have a teammate with GPU in local i would ask help from him
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1314272": "hello, I already asked this question on community forum but there was no reply https://www.kaggle.com/discussion/239733\n\nbasically I am interested do you think it is still worth to use PyTorch on Kaggle competition, or should we all use Tensorflow to be able to compete here? In my last competition I was trying to use PyTorch with TPU but instead focusing on competition itself I was busy with making TPU work, do you think it is worth now, or is it better to just use tf?", "1315936": "Hi Jacek, I made short Pytorch tutorial here and answer the question: Keras vs Pytorch. https://www.kaggle.com/remekkinas/tps-5-pytorch-nn-for-tabular-step-by-step \nIn TPS-05 we observed that there is only my notebook with Pytorch and about 10 in Keras. Keras is much faster to prototype.", "1316014": "I think majority at the top will use Pytorch even though TPU give big edge to TF.\nThe game changer for Google will be JAX in the coming years, which has both the flexiblity of Pytorch and native support of TPU. \nJAX can clearly overtake both Pytorch and TF when backed by Google.", "1316082": "hello Remek,\n\nI think your tutorial is about using PyTorch in general, not about TPU, I am asking about Kaggle competitions, I was able to use PyTorch in https://www.kaggle.com/c/severstal-steel-defect-detection but after that on Kaggle TPU started to be important and in my next competition I was too busy on making PyTorch work with TPU instead focusing on competition itself, so I wonder what is the current status of that here", "1316086": "HI Jacek, I understand :) Thank you for clarification! 👍", "1316307": "Most of them probably may use Pytorch! in my case yes I will use pytorch", "1316327": "hello FenomeN,\n\ndo you train your models on local GPU or do you use Kaggle TPU with PyTorch?", "1316328": "Where do you train your PyTorch models?", "1316568": "I did not join yet this competition. But I use colab Pro. \nWith V100 on colab Pro, I don't feel the need anymore to use TPU for Pytorch.", "1316581": "I had trouble in making PyTorch on TPU. I use colab GPU or Kaggle GPU for me. if i have a teammate with GPU in local i would ask help from him" }, "source": "meta" }