{ "id": 307906, "title": "Wait for a pytorch inference notebook!", "url": "/competitions/happy-whale-and-dolphin/discussion/307906", "author_name": "", "post_date": "2022-02-16T06:45:46.417218700Z", "votes": 12, "comment_count": 15, "views": 0, "content": "

Most of the public inference notebooks are based on tensorflow, waiting for a genius to provide a notebook based on pytorch!

", "messages": [ { "id": "1692633", "postDate": "02/16/2022 06:45:46", "content": "

Most of the public inference notebooks are based on tensorflow, waiting for a genius to provide a notebook based on pytorch!

", "rawMarkdown": "Most of the public inference notebooks are based on tensorflow, waiting for a genius to provide a notebook based on pytorch!", "votes": null }, { "id": "1692638", "postDate": "02/16/2022 06:51:24", "content": "

\"\"

", "rawMarkdown": "![](https://i.imgur.com/hiYzgpn.png)", "votes": null }, { "id": "1692917", "postDate": "02/16/2022 10:43:46", "content": "

Make pytorch great again!💔

", "rawMarkdown": "Make pytorch great again!💔", "votes": null }, { "id": "1693244", "postDate": "02/16/2022 14:44:09", "content": "

TPU很猛，但是只有tf能用tpu😄

", "rawMarkdown": "TPU很猛，但是只有tf能用tpu😄", "votes": null }, { "id": "1693380", "postDate": "02/16/2022 16:26:52", "content": "

Why not make it yourself?

", "rawMarkdown": "Why not make it yourself?", "votes": null }, { "id": "1693412", "postDate": "02/16/2022 16:54:28", "content": "

Humanity? Always looking forward to be lazy.

", "rawMarkdown": "Humanity? Always looking forward to be lazy.", "votes": null }, { "id": "1693579", "postDate": "02/16/2022 18:59:19", "content": "

I think if you tried converting the existing tensorflow notebooks to be trained on GPU and then trained the model, you can expect 8-10 times the training times which are already pretty high.

\n

Even if we say that pytorch models might have a bit \"higher speed\" (using the term very loosely here), and you don't have a multi-gpu setup(which if you do wouldn't be shareable/ usable in Kaggle notebook env anyways), this kind of time would make training pretty much impractical

\n

You can check out tfimm if you want to implement more newer architectures (which are usually in pytorch) in tensorflow

", "rawMarkdown": "I think if you tried converting the existing tensorflow notebooks to be trained on GPU and then trained the model, you can expect 8-10 times the training times which are already pretty high.\n\nEven if we say that pytorch models might have a bit \"higher speed\" (using the term very loosely here), and you don't have a multi-gpu setup(which if you do wouldn't be shareable/ usable in Kaggle notebook env anyways), this kind of time would make training pretty much impractical\n\nYou can check out tfimm if you want to implement more newer architectures (which are usually in pytorch) in tensorflow", "votes": null }, { "id": "1693766", "postDate": "02/16/2022 23:41:12", "content": "

Started one here: https://www.kaggle.com/taindow/pytorch-metric-learning-effnet-arcface.

\n

Training + inference, will work on getting that score up.

", "rawMarkdown": "Started one here: https://www.kaggle.com/taindow/pytorch-metric-learning-effnet-arcface.\n\nTraining + inference, will work on getting that score up.", "votes": null }, { "id": "1693995", "postDate": "02/17/2022 05:27:04", "content": "

underrated comment

", "rawMarkdown": "underrated comment", "votes": null }, { "id": "1694164", "postDate": "02/17/2022 07:50:27", "content": "

It seems, that there are more PyTorch's notebooks than tensorflow's, doesn't it?

", "rawMarkdown": "It seems, that there are more PyTorch's notebooks than tensorflow's, doesn't it?", "votes": null }, { "id": "1695662", "postDate": "02/18/2022 09:44:42", "content": "

Made the base for another one here:
\nhttps://www.kaggle.com/vladvdv/pytorch-inference-notebok-arcface-gem-pooling
\nThere is a lot to improve, refactor and optimize.
\nlater edit: added the training notebook also https://www.kaggle.com/vladvdv/pytorch-train-notebook-arcface-gem-pooling

", "rawMarkdown": "Made the base for another one here: \nhttps://www.kaggle.com/vladvdv/pytorch-inference-notebok-arcface-gem-pooling\nThere is a lot to improve, refactor and optimize.\nlater edit: added the training notebook also https://www.kaggle.com/vladvdv/pytorch-train-notebook-arcface-gem-pooling", "votes": null }, { "id": "1695771", "postDate": "02/18/2022 11:03:54", "content": "

@vladvdv something wrong with the link

", "rawMarkdown": "vladvdv something wrong with the link", "votes": null }, { "id": "1695776", "postDate": "02/18/2022 11:12:09", "content": "

Updated it now. Thanks @pestipeti

", "rawMarkdown": "Updated it now. Thanks @pestipeti", "votes": null }, { "id": "1728945", "postDate": "03/19/2022 13:20:25", "content": "

Here is mine: https://www.kaggle.com/code/clemchris/pytorch-lightning-arcface-train-infer/notebook. Score will get better soon hopefully with backfins data.

", "rawMarkdown": "Here is mine: https://www.kaggle.com/code/clemchris/pytorch-lightning-arcface-train-infer/notebook. Score will get better soon hopefully with backfins data.", "votes": null }, { "id": "1729567", "postDate": "03/20/2022 08:09:46", "content": "

I look into your notebook and … one word came to my mind - outstanding. Thank you for providing this notebook - this is for me source of good Lightning learning. I forked it and trying to improve score … Let you know if will be able to jump a little bit.

\n

What do you mean \"backfins data\"?

", "rawMarkdown": "I look into your notebook and ... one word came to my mind - outstanding. Thank you for providing this notebook - this is for me source of good Lightning learning. I forked it and trying to improve score ... Let you know if will be able to jump a little bit.\n\nWhat do you mean \"backfins data\"?", "votes": null }, { "id": "1729601", "postDate": "03/20/2022 09:20:04", "content": "

Thanks again haha :). Looking forward to your improvements! I am currently evaluating other timm models and data augmentations. timm + Lightning is a strong combo.

\n

The backfins data comes from this notebook where I converted the backfintfrecords dataset.

", "rawMarkdown": "Thanks again haha :). Looking forward to your improvements! I am currently evaluating other timm models and data augmentations. timm + Lightning is a strong combo.\n\nThe backfins data comes from this [notebook](https://www.kaggle.com/code/clemchris/convert-backfintfrecords) where I converted the [backfintfrecords](https://www.kaggle.com/datasets/jpbremer/backfintfrecords) dataset.", "votes": null } ], "comments": [ { "id": 1692638, "author_name": "init27", "author_url": "", "post_date": "02/16/2022 06:51:24", "content": "

\"\"

", "votes": null, "replies": [ { "id": 1692917, "author_name": "jsrdcht", "author_url": "", "post_date": "02/16/2022 10:43:46", "content": "

Make pytorch great again!💔

", "votes": null, "replies": [] }, { "id": 1694164, "author_name": "vad13irt", "author_url": "", "post_date": "02/17/2022 07:50:27", "content": "

It seems, that there are more PyTorch's notebooks than tensorflow's, doesn't it?

", "votes": null, "replies": [] } ] }, { "id": 1693244, "author_name": "forcewithme", "author_url": "", "post_date": "02/16/2022 14:44:09", "content": "

TPU很猛，但是只有tf能用tpu😄

", "votes": null, "replies": [] }, { "id": 1693380, "author_name": "gunesevitan", "author_url": "", "post_date": "02/16/2022 16:26:52", "content": "

Why not make it yourself?

", "votes": null, "replies": [ { "id": 1693412, "author_name": "jsrdcht", "author_url": "", "post_date": "02/16/2022 16:54:28", "content": "

Humanity? Always looking forward to be lazy.

", "votes": null, "replies": [] }, { "id": 1693995, "author_name": "antonzv", "author_url": "", "post_date": "02/17/2022 05:27:04", "content": "

underrated comment

", "votes": null, "replies": [] } ] }, { "id": 1693579, "author_name": "ferlockx", "author_url": "", "post_date": "02/16/2022 18:59:19", "content": "

I think if you tried converting the existing tensorflow notebooks to be trained on GPU and then trained the model, you can expect 8-10 times the training times which are already pretty high.

\n

Even if we say that pytorch models might have a bit \"higher speed\" (using the term very loosely here), and you don't have a multi-gpu setup(which if you do wouldn't be shareable/ usable in Kaggle notebook env anyways), this kind of time would make training pretty much impractical

\n

You can check out tfimm if you want to implement more newer architectures (which are usually in pytorch) in tensorflow

", "votes": null, "replies": [] }, { "id": 1693766, "author_name": "taindow", "author_url": "", "post_date": "02/16/2022 23:41:12", "content": "

Started one here: https://www.kaggle.com/taindow/pytorch-metric-learning-effnet-arcface.

\n

Training + inference, will work on getting that score up.

", "votes": null, "replies": [] }, { "id": 1695662, "author_name": "vladvdv", "author_url": "", "post_date": "02/18/2022 09:44:42", "content": "

Made the base for another one here:
\nhttps://www.kaggle.com/vladvdv/pytorch-inference-notebok-arcface-gem-pooling
\nThere is a lot to improve, refactor and optimize.
\nlater edit: added the training notebook also https://www.kaggle.com/vladvdv/pytorch-train-notebook-arcface-gem-pooling

", "votes": null, "replies": [ { "id": 1695771, "author_name": "pestipeti", "author_url": "", "post_date": "02/18/2022 11:03:54", "content": "

@vladvdv something wrong with the link

", "votes": null, "replies": [] }, { "id": 1695776, "author_name": "vladvdv", "author_url": "", "post_date": "02/18/2022 11:12:09", "content": "

Updated it now. Thanks @pestipeti

", "votes": null, "replies": [] } ] }, { "id": 1728945, "author_name": "clemchris", "author_url": "", "post_date": "03/19/2022 13:20:25", "content": "

Here is mine: https://www.kaggle.com/code/clemchris/pytorch-lightning-arcface-train-infer/notebook. Score will get better soon hopefully with backfins data.

", "votes": null, "replies": [ { "id": 1729567, "author_name": "remekkinas", "author_url": "", "post_date": "03/20/2022 08:09:46", "content": "

I look into your notebook and … one word came to my mind - outstanding. Thank you for providing this notebook - this is for me source of good Lightning learning. I forked it and trying to improve score … Let you know if will be able to jump a little bit.

\n

What do you mean \"backfins data\"?

", "votes": null, "replies": [] }, { "id": 1729601, "author_name": "clemchris", "author_url": "", "post_date": "03/20/2022 09:20:04", "content": "

Thanks again haha :). Looking forward to your improvements! I am currently evaluating other timm models and data augmentations. timm + Lightning is a strong combo.

\n

The backfins data comes from this notebook where I converted the backfintfrecords dataset.

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1692633": "Most of the public inference notebooks are based on tensorflow, waiting for a genius to provide a notebook based on pytorch!", "1692638": "![](https://i.imgur.com/hiYzgpn.png)", "1692917": "Make pytorch great again!💔", "1693244": "TPU很猛，但是只有tf能用tpu😄", "1693380": "Why not make it yourself?", "1693412": "Humanity? Always looking forward to be lazy.", "1693579": "I think if you tried converting the existing tensorflow notebooks to be trained on GPU and then trained the model, you can expect 8-10 times the training times which are already pretty high.\n\nEven if we say that pytorch models might have a bit \"higher speed\" (using the term very loosely here), and you don't have a multi-gpu setup(which if you do wouldn't be shareable/ usable in Kaggle notebook env anyways), this kind of time would make training pretty much impractical\n\nYou can check out tfimm if you want to implement more newer architectures (which are usually in pytorch) in tensorflow", "1693766": "Started one here: https://www.kaggle.com/taindow/pytorch-metric-learning-effnet-arcface.\n\nTraining + inference, will work on getting that score up.", "1693995": "underrated comment", "1694164": "It seems, that there are more PyTorch's notebooks than tensorflow's, doesn't it?", "1695662": "Made the base for another one here: \nhttps://www.kaggle.com/vladvdv/pytorch-inference-notebok-arcface-gem-pooling\nThere is a lot to improve, refactor and optimize.\nlater edit: added the training notebook also https://www.kaggle.com/vladvdv/pytorch-train-notebook-arcface-gem-pooling", "1695771": "vladvdv something wrong with the link", "1695776": "Updated it now. Thanks @pestipeti", "1728945": "Here is mine: https://www.kaggle.com/code/clemchris/pytorch-lightning-arcface-train-infer/notebook. Score will get better soon hopefully with backfins data.", "1729567": "I look into your notebook and ... one word came to my mind - outstanding. Thank you for providing this notebook - this is for me source of good Lightning learning. I forked it and trying to improve score ... Let you know if will be able to jump a little bit.\n\nWhat do you mean \"backfins data\"?", "1729601": "Thanks again haha :). Looking forward to your improvements! I am currently evaluating other timm models and data augmentations. timm + Lightning is a strong combo.\n\nThe backfins data comes from this [notebook](https://www.kaggle.com/code/clemchris/convert-backfintfrecords) where I converted the [backfintfrecords](https://www.kaggle.com/datasets/jpbremer/backfintfrecords) dataset." }, "source": "meta" }