Unfortunately I don't think I will have time to join this competition. But I was visualizing some Kernels and I noticed that most of them use pytorch, while almost none use Keras/TF. The question is, why?
", "messages": [ { "id": "705998", "postDate": "12/29/2019 19:49:25", "content": "Unfortunately I don't think I will have time to join this competition. But I was visualizing some Kernels and I noticed that most of them use pytorch, while almost none use Keras/TF. The question is, why?
", "rawMarkdown": "Unfortunately I don't think I will have time to join this competition. But I was visualizing some Kernels and I noticed that most of them use pytorch, while almost none use Keras/TF. The question is, why?", "votes": null }, { "id": "706117", "postDate": "12/29/2019 23:54:14", "content": "For me, it is convenient to look into the structures and to delicate manipulations. Just ideal tool to study NN for me :)
", "rawMarkdown": "For me, it is convenient to look into the structures and to delicate manipulations. Just ideal tool to study NN for me :)", "votes": null }, { "id": "706386", "postDate": "12/30/2019 10:07:45", "content": "I was going to actually ask as well, I'm pretty surprised to see that myself.\nI might have a few ideas of why that is:\n- Tensorflow recently switched to TF 2.0, and so did Kaggle for its use of Keras. However the Google documentation is pretty bad, to the point that it has even become a meme sometimes. There's a lack of communication IMHO from Google about how to do things we could do in TF 1.X in TF 2.0 and I think a lot of people are getting confused.\n- On the other hand, PyTorch seems to be much simpler with one way to do things. There are also people using fast.ai on top of it who also seem to be growing the number of people using PyTorch.
\n\nOverall, from what I've seen on Reddit, Twitter and here on Kaggle, a lot of people seem fed up with the poor documentation of the change to TF 2.0, myself included. I've been starting to get into PyTorch slowly for that very reason.
\n\nHope that provides some answers, and I'd be curious to here from others!
", "rawMarkdown": "I was going to actually ask as well, I'm pretty surprised to see that myself.\nI might have a few ideas of why that is:\n- Tensorflow recently switched to TF 2.0, and so did Kaggle for its use of Keras. However the Google documentation is pretty bad, to the point that it has even become a meme sometimes. There's a lack of communication IMHO from Google about how to do things we could do in TF 1.X in TF 2.0 and I think a lot of people are getting confused.\n- On the other hand, PyTorch seems to be much simpler with one way to do things. There are also people using fast.ai on top of it who also seem to be growing the number of people using PyTorch.\n\nOverall, from what I've seen on Reddit, Twitter and here on Kaggle, a lot of people seem fed up with the poor documentation of the change to TF 2.0, myself included. I've been starting to get into PyTorch slowly for that very reason.\n\nHope that provides some answers, and I'd be curious to here from others!", "votes": null }, { "id": "706424", "postDate": "12/30/2019 11:02:58", "content": "my two cents why pytorch is clear cut here (take it with a pinch of salt):\n- it is arguably easier to learn (perhaps due to its more pythonic nature and less changes over time)\n- it is faster (at least it used to be a few months back)\n- it is easier to implement new/novel ideas (i guess that's why academics prefer it)\n- then, more research is done using pytorch (recent SOTA techniques may help you win competitions)\n- most of the winning solutions on recent Kaggle competitions in computer vision are built on pytorch
\n\nI guess this explains why we see this trend here. But in some other type of tasks (other than computer vision) or with other goals (deployment) we might have a different picture...
", "rawMarkdown": "my two cents why pytorch is clear cut here (take it with a pinch of salt):\n- it is arguably easier to learn (perhaps due to its more pythonic nature and less changes over time)\n- it is faster (at least it used to be a few months back)\n- it is easier to implement new/novel ideas (i guess that's why academics prefer it)\n- then, more research is done using pytorch (recent SOTA techniques may help you win competitions)\n- most of the winning solutions on recent Kaggle competitions in computer vision are built on pytorch\n\nI guess this explains why we see this trend here. But in some other type of tasks (other than computer vision) or with other goals (deployment) we might have a different picture...", "votes": null }, { "id": "718555", "postDate": "01/14/2020 14:23:12", "content": "Thank you for you answers guys!
", "rawMarkdown": "Thank you for you answers guys!", "votes": null }, { "id": "718722", "postDate": "01/14/2020 17:31:48", "content": "I got into DL through fast.ai which is based on pyTorch. I've been using pyTorch ever since. I think same is the case for many people.
", "rawMarkdown": "I got into DL through fast.ai which is based on pyTorch. I've been using pyTorch ever since. I think same is the case for many people.", "votes": null } ], "comments": [ { "id": 706117, "author_name": "hanjoonchoe", "author_url": "", "post_date": "12/29/2019 23:54:14", "content": "For me, it is convenient to look into the structures and to delicate manipulations. Just ideal tool to study NN for me :)
", "votes": null, "replies": [] }, { "id": 706386, "author_name": "maxlenormand", "author_url": "", "post_date": "12/30/2019 10:07:45", "content": "I was going to actually ask as well, I'm pretty surprised to see that myself.\nI might have a few ideas of why that is:\n- Tensorflow recently switched to TF 2.0, and so did Kaggle for its use of Keras. However the Google documentation is pretty bad, to the point that it has even become a meme sometimes. There's a lack of communication IMHO from Google about how to do things we could do in TF 1.X in TF 2.0 and I think a lot of people are getting confused.\n- On the other hand, PyTorch seems to be much simpler with one way to do things. There are also people using fast.ai on top of it who also seem to be growing the number of people using PyTorch.
\n\nOverall, from what I've seen on Reddit, Twitter and here on Kaggle, a lot of people seem fed up with the poor documentation of the change to TF 2.0, myself included. I've been starting to get into PyTorch slowly for that very reason.
\n\nHope that provides some answers, and I'd be curious to here from others!
", "votes": null, "replies": [] }, { "id": 706424, "author_name": "valanm", "author_url": "", "post_date": "12/30/2019 11:02:58", "content": "my two cents why pytorch is clear cut here (take it with a pinch of salt):\n- it is arguably easier to learn (perhaps due to its more pythonic nature and less changes over time)\n- it is faster (at least it used to be a few months back)\n- it is easier to implement new/novel ideas (i guess that's why academics prefer it)\n- then, more research is done using pytorch (recent SOTA techniques may help you win competitions)\n- most of the winning solutions on recent Kaggle competitions in computer vision are built on pytorch
\n\nI guess this explains why we see this trend here. But in some other type of tasks (other than computer vision) or with other goals (deployment) we might have a different picture...
", "votes": null, "replies": [] }, { "id": 718555, "author_name": "clappis", "author_url": "", "post_date": "01/14/2020 14:23:12", "content": "Thank you for you answers guys!
", "votes": null, "replies": [] }, { "id": 718722, "author_name": "timetraveller98", "author_url": "", "post_date": "01/14/2020 17:31:48", "content": "I got into DL through fast.ai which is based on pyTorch. I've been using pyTorch ever since. I think same is the case for many people.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "705998": "Unfortunately I don't think I will have time to join this competition. But I was visualizing some Kernels and I noticed that most of them use pytorch, while almost none use Keras/TF. The question is, why?", "706117": "For me, it is convenient to look into the structures and to delicate manipulations. Just ideal tool to study NN for me :)", "706386": "I was going to actually ask as well, I'm pretty surprised to see that myself.\nI might have a few ideas of why that is:\n- Tensorflow recently switched to TF 2.0, and so did Kaggle for its use of Keras. However the Google documentation is pretty bad, to the point that it has even become a meme sometimes. There's a lack of communication IMHO from Google about how to do things we could do in TF 1.X in TF 2.0 and I think a lot of people are getting confused.\n- On the other hand, PyTorch seems to be much simpler with one way to do things. There are also people using fast.ai on top of it who also seem to be growing the number of people using PyTorch.\n\nOverall, from what I've seen on Reddit, Twitter and here on Kaggle, a lot of people seem fed up with the poor documentation of the change to TF 2.0, myself included. I've been starting to get into PyTorch slowly for that very reason.\n\nHope that provides some answers, and I'd be curious to here from others!", "706424": "my two cents why pytorch is clear cut here (take it with a pinch of salt):\n- it is arguably easier to learn (perhaps due to its more pythonic nature and less changes over time)\n- it is faster (at least it used to be a few months back)\n- it is easier to implement new/novel ideas (i guess that's why academics prefer it)\n- then, more research is done using pytorch (recent SOTA techniques may help you win competitions)\n- most of the winning solutions on recent Kaggle competitions in computer vision are built on pytorch\n\nI guess this explains why we see this trend here. But in some other type of tasks (other than computer vision) or with other goals (deployment) we might have a different picture...", "718555": "Thank you for you answers guys!", "718722": "I got into DL through fast.ai which is based on pyTorch. I've been using pyTorch ever since. I think same is the case for many people." }, "source": "meta" }