Hello! While browsing through the networks of ML and notebooks, I have realized that there is immense popularity on both Tensorflow and pytorch as the new trend. I am wondering which framework is the \"future prospect\" for ML? I have looked at ImageAI's documentation, and it seems as if they are also going to shift into pytorch from tensor. I am wondering which is more efficient and powerful in terms of ML context.
\nThank you!
", "messages": [ { "id": "1156146", "postDate": "01/17/2021 00:29:30", "content": "Hello! While browsing through the networks of ML and notebooks, I have realized that there is immense popularity on both Tensorflow and pytorch as the new trend. I am wondering which framework is the \"future prospect\" for ML? I have looked at ImageAI's documentation, and it seems as if they are also going to shift into pytorch from tensor. I am wondering which is more efficient and powerful in terms of ML context.
\nThank you!
", "rawMarkdown": "Hello! While browsing through the networks of ML and notebooks, I have realized that there is immense popularity on both Tensorflow and pytorch as the new trend. I am wondering which framework is the \"future prospect\" for ML? I have looked at ImageAI's documentation, and it seems as if they are also going to shift into pytorch from tensor. I am wondering which is more efficient and powerful in terms of ML context. \n\nThank you!", "votes": null }, { "id": "1156580", "postDate": "01/17/2021 09:08:20", "content": "hi @andyjianzhou ,
\nThere is a discussing thread on the same topic: https://www.kaggle.com/discussion/203018
\nI bet it will be helpful as it did provide me with better insight.
\nHappy Learning!
Been involved with using software for almost 55 years. It has always been the case that a couple of competing methods could be used - and almost always the case that at some point in the future one version would flourish and the other would die.
\nHow to pick the winner - my recommendation is flip a coin. Or learn both - Or use the one you \"like\". Sadly for me, my age has reached the point where I cannot learn both - my old RAM needs too many epochs for new learning's. So I am hoping that tensorflow lasts longer than I do.
\nIF both survive there will be time periods where Pytorch is the better choice - but other time frames where Tensorflow will be the better choice.
", "rawMarkdown": "Been involved with using software for almost 55 years. It has always been the case that a couple of competing methods could be used - and almost always the case that at some point in the future one version would flourish and the other would die.\n\nHow to pick the winner - my recommendation is flip a coin. Or learn both - Or use the one you \"like\". Sadly for me, my age has reached the point where I cannot learn both - my old RAM needs too many epochs for new learning's. So I am hoping that tensorflow lasts longer than I do. \n\nIF both survive there will be time periods where Pytorch is the better choice - but other time frames where Tensorflow will be the better choice.", "votes": null }, { "id": "1163755", "postDate": "01/21/2021 22:47:24", "content": "Ah, alright thank you! I recently just started working with kaggle competitions, so far had some experiences with past projects but really only learned Keras. Thank you for this analysis, I will be looking forward to learning both while I have time.
", "rawMarkdown": "Ah, alright thank you! I recently just started working with kaggle competitions, so far had some experiences with past projects but really only learned Keras. Thank you for this analysis, I will be looking forward to learning both while I have time.", "votes": null }, { "id": "1163756", "postDate": "01/21/2021 22:47:31", "content": "Ah yes, I see now. Thank you!
", "rawMarkdown": "Ah yes, I see now. Thank you!", "votes": null }, { "id": "1163812", "postDate": "01/22/2021 00:28:55", "content": "I've been programming computers since 1962, PC Jimmmy, and I'm still trying to learn new stuff. For this contest I'm currently using 'R -> Keras -> Tensorflow -> CUDA', but that may yet change. There's an interesting article about learning, memory, and age titled \"Starting Fresh\" in the January 18, 2021 issue of \"The New Yorker\" magazine: https://www.newyorker.com/magazine/2021/01/18/is-it-really-too-late-to-learn-new-skills. I'd like to think, correctly or not, that the more one already knows, the easier it is to learn even more. And perhaps someday people will be able to consult their neurologist about getting a \"RAM upgrade\" for their aging brain.
", "rawMarkdown": "I've been programming computers since 1962, PC Jimmmy, and I'm still trying to learn new stuff. For this contest I'm currently using 'R -> Keras -> Tensorflow -> CUDA', but that may yet change. There's an interesting article about learning, memory, and age titled \"Starting Fresh\" in the January 18, 2021 issue of \"The New Yorker\" magazine: https://www.newyorker.com/magazine/2021/01/18/is-it-really-too-late-to-learn-new-skills. I'd like to think, correctly or not, that the more one already knows, the easier it is to learn even more. And perhaps someday people will be able to consult their neurologist about getting a \"RAM upgrade\" for their aging brain.", "votes": null }, { "id": "1163838", "postDate": "01/22/2021 01:26:57", "content": "Thanks for the link. For me it is truly like a model training. Years ago I could get to 97% accuracy in an epoch or two with anything I was learning. Now at 74 it seems I need to hit more like 20 epochs with a slower learning rate. I can still get to 97% but it takes so much longer and hence I learn less.
\nI started on Kaggle with R - found it almost useless for vision problems like this competition.
", "rawMarkdown": "Thanks for the link. For me it is truly like a model training. Years ago I could get to 97% accuracy in an epoch or two with anything I was learning. Now at 74 it seems I need to hit more like 20 epochs with a slower learning rate. I can still get to 97% but it takes so much longer and hence I learn less. \n\nI started on Kaggle with R - found it almost useless for vision problems like this competition.", "votes": null }, { "id": "1181510", "postDate": "02/02/2021 00:15:55", "content": "Ah yes, there was a recent update released. Pytorch is currently outpacing tensorflow in research papers, however tensorflow is currently being used a lot in long term projects still. Pytorch is growing rapidly, and pytorch is currently equal(or more popular) than tensorflow in terms of research. I don't know if that's true about projects. May someone confirm this? As I am currently learning Pytorch, but I have troubles choosing which framework to \"main\". Thanks!
", "rawMarkdown": "Ah yes, there was a recent update released. Pytorch is currently outpacing tensorflow in research papers, however tensorflow is currently being used a lot in long term projects still. Pytorch is growing rapidly, and pytorch is currently equal(or more popular) than tensorflow in terms of research. I don't know if that's true about projects. May someone confirm this? As I am currently learning Pytorch, but I have troubles choosing which framework to \"main\". Thanks!", "votes": null }, { "id": "1181531", "postDate": "02/02/2021 00:41:26", "content": "Even at Google, they seem not to bet on TF as the \"future prospect\" for ML.
\nThey don't write any research paper code in TF2. All the code are written either in TF1 (including the current SOTA on Imagenet published this year) or they move directly to JAX (like the ViT code) or even use Pytorch .
\nThey clearly do nothing to promote TF2
", "rawMarkdown": "Even at Google, they seem not to bet on TF as the \"future prospect\" for ML. \n\nThey don't write any research paper code in TF2. All the code are written either in TF1 (including the current SOTA on Imagenet published this year) or they move directly to JAX (like the ViT code) or even use Pytorch .\n\nThey clearly do nothing to promote TF2", "votes": null }, { "id": "1181600", "postDate": "02/02/2021 02:23:24", "content": "Ah, so shall I start learning pytorch and main that framework then?
", "rawMarkdown": "Ah, so shall I start learning pytorch and main that framework then?", "votes": null } ], "comments": [ { "id": 1156580, "author_name": "supplejade", "author_url": "", "post_date": "01/17/2021 09:08:20", "content": "hi @andyjianzhou ,
\nThere is a discussing thread on the same topic: https://www.kaggle.com/discussion/203018
\nI bet it will be helpful as it did provide me with better insight.
\nHappy Learning!
Ah yes, I see now. Thank you!
", "votes": null, "replies": [] } ] }, { "id": 1156598, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/17/2021 09:20:42", "content": "Been involved with using software for almost 55 years. It has always been the case that a couple of competing methods could be used - and almost always the case that at some point in the future one version would flourish and the other would die.
\nHow to pick the winner - my recommendation is flip a coin. Or learn both - Or use the one you \"like\". Sadly for me, my age has reached the point where I cannot learn both - my old RAM needs too many epochs for new learning's. So I am hoping that tensorflow lasts longer than I do.
\nIF both survive there will be time periods where Pytorch is the better choice - but other time frames where Tensorflow will be the better choice.
", "votes": null, "replies": [ { "id": 1163755, "author_name": "andyjianzhou", "author_url": "", "post_date": "01/21/2021 22:47:24", "content": "Ah, alright thank you! I recently just started working with kaggle competitions, so far had some experiences with past projects but really only learned Keras. Thank you for this analysis, I will be looking forward to learning both while I have time.
", "votes": null, "replies": [] }, { "id": 1163812, "author_name": "dslate", "author_url": "", "post_date": "01/22/2021 00:28:55", "content": "I've been programming computers since 1962, PC Jimmmy, and I'm still trying to learn new stuff. For this contest I'm currently using 'R -> Keras -> Tensorflow -> CUDA', but that may yet change. There's an interesting article about learning, memory, and age titled \"Starting Fresh\" in the January 18, 2021 issue of \"The New Yorker\" magazine: https://www.newyorker.com/magazine/2021/01/18/is-it-really-too-late-to-learn-new-skills. I'd like to think, correctly or not, that the more one already knows, the easier it is to learn even more. And perhaps someday people will be able to consult their neurologist about getting a \"RAM upgrade\" for their aging brain.
", "votes": null, "replies": [] }, { "id": 1163838, "author_name": "pcjimmmy", "author_url": "", "post_date": "01/22/2021 01:26:57", "content": "Thanks for the link. For me it is truly like a model training. Years ago I could get to 97% accuracy in an epoch or two with anything I was learning. Now at 74 it seems I need to hit more like 20 epochs with a slower learning rate. I can still get to 97% but it takes so much longer and hence I learn less.
\nI started on Kaggle with R - found it almost useless for vision problems like this competition.
", "votes": null, "replies": [] }, { "id": 1181510, "author_name": "andyjianzhou", "author_url": "", "post_date": "02/02/2021 00:15:55", "content": "Ah yes, there was a recent update released. Pytorch is currently outpacing tensorflow in research papers, however tensorflow is currently being used a lot in long term projects still. Pytorch is growing rapidly, and pytorch is currently equal(or more popular) than tensorflow in terms of research. I don't know if that's true about projects. May someone confirm this? As I am currently learning Pytorch, but I have troubles choosing which framework to \"main\". Thanks!
", "votes": null, "replies": [] } ] }, { "id": 1181531, "author_name": "serigne", "author_url": "", "post_date": "02/02/2021 00:41:26", "content": "Even at Google, they seem not to bet on TF as the \"future prospect\" for ML.
\nThey don't write any research paper code in TF2. All the code are written either in TF1 (including the current SOTA on Imagenet published this year) or they move directly to JAX (like the ViT code) or even use Pytorch .
\nThey clearly do nothing to promote TF2
", "votes": null, "replies": [ { "id": 1181600, "author_name": "andyjianzhou", "author_url": "", "post_date": "02/02/2021 02:23:24", "content": "Ah, so shall I start learning pytorch and main that framework then?
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1156146": "Hello! While browsing through the networks of ML and notebooks, I have realized that there is immense popularity on both Tensorflow and pytorch as the new trend. I am wondering which framework is the \"future prospect\" for ML? I have looked at ImageAI's documentation, and it seems as if they are also going to shift into pytorch from tensor. I am wondering which is more efficient and powerful in terms of ML context. \n\nThank you!", "1156580": "hi @andyjianzhou ,\n\n\nThere is a discussing thread on the same topic: https://www.kaggle.com/discussion/203018\nI bet it will be helpful as it did provide me with better insight.\nHappy Learning!", "1156598": "Been involved with using software for almost 55 years. It has always been the case that a couple of competing methods could be used - and almost always the case that at some point in the future one version would flourish and the other would die.\n\nHow to pick the winner - my recommendation is flip a coin. Or learn both - Or use the one you \"like\". Sadly for me, my age has reached the point where I cannot learn both - my old RAM needs too many epochs for new learning's. So I am hoping that tensorflow lasts longer than I do. \n\nIF both survive there will be time periods where Pytorch is the better choice - but other time frames where Tensorflow will be the better choice.", "1163755": "Ah, alright thank you! I recently just started working with kaggle competitions, so far had some experiences with past projects but really only learned Keras. Thank you for this analysis, I will be looking forward to learning both while I have time.", "1163756": "Ah yes, I see now. Thank you!", "1163812": "I've been programming computers since 1962, PC Jimmmy, and I'm still trying to learn new stuff. For this contest I'm currently using 'R -> Keras -> Tensorflow -> CUDA', but that may yet change. There's an interesting article about learning, memory, and age titled \"Starting Fresh\" in the January 18, 2021 issue of \"The New Yorker\" magazine: https://www.newyorker.com/magazine/2021/01/18/is-it-really-too-late-to-learn-new-skills. I'd like to think, correctly or not, that the more one already knows, the easier it is to learn even more. And perhaps someday people will be able to consult their neurologist about getting a \"RAM upgrade\" for their aging brain.", "1163838": "Thanks for the link. For me it is truly like a model training. Years ago I could get to 97% accuracy in an epoch or two with anything I was learning. Now at 74 it seems I need to hit more like 20 epochs with a slower learning rate. I can still get to 97% but it takes so much longer and hence I learn less. \n\nI started on Kaggle with R - found it almost useless for vision problems like this competition.", "1181510": "Ah yes, there was a recent update released. Pytorch is currently outpacing tensorflow in research papers, however tensorflow is currently being used a lot in long term projects still. Pytorch is growing rapidly, and pytorch is currently equal(or more popular) than tensorflow in terms of research. I don't know if that's true about projects. May someone confirm this? As I am currently learning Pytorch, but I have troubles choosing which framework to \"main\". Thanks!", "1181531": "Even at Google, they seem not to bet on TF as the \"future prospect\" for ML. \n\nThey don't write any research paper code in TF2. All the code are written either in TF1 (including the current SOTA on Imagenet published this year) or they move directly to JAX (like the ViT code) or even use Pytorch .\n\nThey clearly do nothing to promote TF2", "1181600": "Ah, so shall I start learning pytorch and main that framework then?" }, "source": "meta" }