I don't know too much about this topic but I'd love to learn about this :)\nIf you know about this topic or you've read about this ... I'd like to know:
\n\nI guess is not possible, but even if I have to invest here all my hours in order to learn about this topic, I'll do. So I'd like to know your opinion.
\n\n
", "messages": [ { "id": "625154", "postDate": "09/12/2019 18:49:43", "content": "I don't know too much about this topic but I'd love to learn about this :)\nIf you know about this topic or you've read about this ... I'd like to know:
\n\nI guess is not possible, but even if I have to invest here all my hours in order to learn about this topic, I'll do. So I'd like to know your opinion.
\n\n
", "rawMarkdown": "I don't know too much about this topic but I'd love to learn about this :)\nIf you know about this topic or you've read about this ... I'd like to know:\n\n- which kind of resources do we need?\n- is possible to compete with the 30hr GPU/week ? or at least to participate?\n\nI guess is not possible, but even if I have to invest here all my hours in order to learn about this topic, I'll do. So I'd like to know your opinion.\n\n", "votes": null }, { "id": "625212", "postDate": "09/12/2019 19:45:23", "content": "I've worked on 3d object detection. My work was on lung nodule detection (3d bounding box prediction for cancerous nodules) in CT scans (512x512xn sized images where n was in range of 300-600), I used patches of 128^3 as input data in a RetinaNet framework (adapted to 3D domain). I remember, I used about batch size of 4/8 on a GTX 1080 Ti (11 GB) and it worked well.
\n\nAs far as this competition is concerned, I'll have to look at the data (Data download/kernel creation is not working properly as of now) before jumping to any conclusion regarding usability of Kaggle kernels for this competition.
", "rawMarkdown": "I've worked on 3d object detection. My work was on lung nodule detection (3d bounding box prediction for cancerous nodules) in CT scans (512x512xn sized images where n was in range of 300-600), I used patches of 128^3 as input data in a RetinaNet framework (adapted to 3D domain). I remember, I used about batch size of 4/8 on a GTX 1080 Ti (11 GB) and it worked well. \n\nAs far as this competition is concerned, I'll have to look at the data (Data download/kernel creation is not working properly as of now) before jumping to any conclusion regarding usability of Kaggle kernels for this competition.", "votes": null }, { "id": "625223", "postDate": "09/12/2019 19:58:23", "content": "@rishabhiitbhu thank you so much for sharing your experience. Really helpful!
", "rawMarkdown": "rishabhiitbhu thank you so much for sharing your experience. Really helpful!", "votes": null }, { "id": "625253", "postDate": "09/12/2019 20:36:11", "content": "I am not fortunate enough to own a GPU for Kaggle competitions. Will see if the new quota will be bearable (one error in your code and 1/4 of your weekly quota is gone... also can only experiment 2 models at the same time).
", "rawMarkdown": "I am not fortunate enough to own a GPU for Kaggle competitions. Will see if the new quota will be bearable (one error in your code and 1/4 of your weekly quota is gone... also can only experiment 2 models at the same time).", "votes": null }, { "id": "625379", "postDate": "09/13/2019 02:28:45", "content": "The crying frog makes me sad lol .... maybe they will distribute 1000$ GCP credit to help us get some hardware muscle ? Cross fingers!
", "rawMarkdown": "The crying frog makes me sad *lol* .... maybe they will distribute 1000$ GCP credit to help us get some hardware muscle ? Cross fingers!", "votes": null }, { "id": "625466", "postDate": "09/13/2019 05:37:58", "content": "My first ever competition was this one: https://blog.udacity.com/2017/03/udacity-didi-self-driving-car-challenge.html (I was just 3 months into DL) I had 1 x 1080 Ti card and that one also was about 3d detection using sensor fusion (lidar+cameras+etc). Dataset was in the order of 20-30 Gb IIRC. I managed to finish #4 (against 2000+ teams) but you have to play your cards very wisely. My expectatin is that in this competition computing resources are going to be a deciding factor for gold medals., especially because no restrictions have been posed about real-time inference, which I find very weird, see https://www.kaggle.com/c/3d-object-detection-for-autonomous-vehicles/discussion/108655#latest-625470
", "rawMarkdown": "My first ever competition was this one: https://blog.udacity.com/2017/03/udacity-didi-self-driving-car-challenge.html (I was just 3 months into DL) I had 1 x 1080 Ti card and that one also was about 3d detection using sensor fusion (lidar+cameras+etc). Dataset was in the order of 20-30 Gb IIRC. I managed to finish #4 (against 2000+ teams) but you have to play your cards very wisely. My expectatin is that in this competition computing resources are going to be a deciding factor for gold medals., especially because no restrictions have been posed about real-time inference, which I find very weird, see https://www.kaggle.com/c/3d-object-detection-for-autonomous-vehicles/discussion/108655#latest-625470", "votes": null }, { "id": "625508", "postDate": "09/13/2019 06:32:28", "content": "Thank you so much for share your experience! Good to know :)\nBtw, is your solution public? I mean , Udacity competition's solution.
", "rawMarkdown": "Thank you so much for share your experience! Good to know :)\nBtw, is your solution public? I mean , Udacity competition's solution.", "votes": null }, { "id": "625514", "postDate": "09/13/2019 06:39:33", "content": "Would Google Colab be a good alternatives?
", "rawMarkdown": "Would Google Colab be a good alternatives?", "votes": null }, { "id": "625542", "postDate": "09/13/2019 07:16:28", "content": "I don't think so. Read this thread: Why does Google Colab randomly disconnects before 12 hours\nfree TPUs is amazing, but I have no idea how to use them properly.
", "rawMarkdown": "I don't think so. Read this thread: [Why does Google Colab randomly disconnects before 12 hours](https://stackoverflow.com/questions/52670872/why-does-google-colab-randomly-disconnects-before-12-hours)\nfree TPUs is amazing, but I have no idea how to use them properly.", "votes": null }, { "id": "625543", "postDate": "09/13/2019 07:17:30", "content": "Not source code, but I've spoken about it publicly in some events, some which are available online (in Spanish). In the case of Didi it was a novel solution which allowed to hit #4 just with lidar (other teams used lidar + camera).
", "rawMarkdown": "Not source code, but I've spoken about it publicly in some events, some which are available online (in Spanish). In the case of Didi it was a novel solution which allowed to hit #4 just with lidar (other teams used lidar + camera).", "votes": null }, { "id": "625732", "postDate": "09/13/2019 11:38:12", "content": "TPU has a big restriction on coding. I receive lots of error and just give it up...
", "rawMarkdown": "TPU has a big restriction on coding. I receive lots of error and just give it up...", "votes": null }, { "id": "625918", "postDate": "09/13/2019 15:26:11", "content": "\n\n\nRTX 1080 Ti (11 GB)
\n
Interesting 😉
", "rawMarkdown": "> RTX 1080 Ti (11 GB)\n\nInteresting 😉", "votes": null }, { "id": "625919", "postDate": "09/13/2019 15:29:37", "content": "I'm 100% sure they will not give 1000$ , and without overfitting hahahahha :)\ncross the fingers and maybe wr get 300$. In this competition only around 80 -100teams are going to compete (because they can, check the Google comps)
", "rawMarkdown": "I'm 100% sure they will not give 1000$ , and without overfitting hahahahha :)\ncross the fingers and maybe wr get 300$. In this competition only around 80 -100teams are going to compete (because they can, check the Google comps)", "votes": null }, { "id": "625921", "postDate": "09/13/2019 15:30:29", "content": "True, clean the code and find bugs is crazy... and play with the precission... hope someone starts playing with TPU and post results/notebooks....
", "rawMarkdown": "True, clean the code and find bugs is crazy... and play with the precission... hope someone starts playing with TPU and post results/notebooks....", "votes": null }, { "id": "625922", "postDate": "09/13/2019 15:31:16", "content": "Fixed it :p
", "rawMarkdown": "Fixed it :p", "votes": null }, { "id": "626435", "postDate": "09/14/2019 10:31:56", "content": "I use google colab, I save the model every now and then, upload to the drive. When it disconnects, download the trained model from the drive and continue training. it helps when you do not have your own gpu.
", "rawMarkdown": "I use google colab, I save the model every now and then, upload to the drive. When it disconnects, download the trained model from the drive and continue training. it helps when you do not have your own gpu.", "votes": null }, { "id": "627320", "postDate": "09/15/2019 19:07:21", "content": "Thank you for sharing your experience. I want to work on this challenge and want to buy a pc with rtx 2080 Ti card. Do you think it is enough for such works? Or I have to get multiple GPUs?\nmaybe couple of 1080 Ti instead of one 2080 Ti.
", "rawMarkdown": "Thank you for sharing your experience. I want to work on this challenge and want to buy a pc with rtx 2080 Ti card. Do you think it is enough for such works? Or I have to get multiple GPUs?\nmaybe couple of 1080 Ti instead of one 2080 Ti.", "votes": null }, { "id": "631848", "postDate": "09/22/2019 19:59:44", "content": "Honestly, I'm trying to compete in Severstal with only 30 hours, but I'm start to think that is impossible seriously compete with that limitations. You can try if you want to learn something in application, but better to don't have much expectation about to finish in a highter position.
", "rawMarkdown": "Honestly, I'm trying to compete in Severstal with only 30 hours, but I'm start to think that is impossible seriously compete with that limitations. You can try if you want to learn something in application, but better to don't have much expectation about to finish in a highter position.", "votes": null } ], "comments": [ { "id": 625212, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "09/12/2019 19:45:23", "content": "I've worked on 3d object detection. My work was on lung nodule detection (3d bounding box prediction for cancerous nodules) in CT scans (512x512xn sized images where n was in range of 300-600), I used patches of 128^3 as input data in a RetinaNet framework (adapted to 3D domain). I remember, I used about batch size of 4/8 on a GTX 1080 Ti (11 GB) and it worked well.
\n\nAs far as this competition is concerned, I'll have to look at the data (Data download/kernel creation is not working properly as of now) before jumping to any conclusion regarding usability of Kaggle kernels for this competition.
", "votes": null, "replies": [ { "id": 625223, "author_name": "jesucristo", "author_url": "", "post_date": "09/12/2019 19:58:23", "content": "@rishabhiitbhu thank you so much for sharing your experience. Really helpful!
", "votes": null, "replies": [] }, { "id": 625918, "author_name": "init27", "author_url": "", "post_date": "09/13/2019 15:26:11", "content": "\n\n\nRTX 1080 Ti (11 GB)
\n
Interesting 😉
", "votes": null, "replies": [] }, { "id": 625922, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "09/13/2019 15:31:16", "content": "Fixed it :p
", "votes": null, "replies": [] } ] }, { "id": 625253, "author_name": "xhlulu", "author_url": "", "post_date": "09/12/2019 20:36:11", "content": "I am not fortunate enough to own a GPU for Kaggle competitions. Will see if the new quota will be bearable (one error in your code and 1/4 of your weekly quota is gone... also can only experiment 2 models at the same time).
", "votes": null, "replies": [] }, { "id": 625379, "author_name": "nyleve", "author_url": "", "post_date": "09/13/2019 02:28:45", "content": "The crying frog makes me sad lol .... maybe they will distribute 1000$ GCP credit to help us get some hardware muscle ? Cross fingers!
", "votes": null, "replies": [ { "id": 625919, "author_name": "jesucristo", "author_url": "", "post_date": "09/13/2019 15:29:37", "content": "I'm 100% sure they will not give 1000$ , and without overfitting hahahahha :)\ncross the fingers and maybe wr get 300$. In this competition only around 80 -100teams are going to compete (because they can, check the Google comps)
", "votes": null, "replies": [] } ] }, { "id": 625466, "author_name": "antorsae", "author_url": "", "post_date": "09/13/2019 05:37:58", "content": "My first ever competition was this one: https://blog.udacity.com/2017/03/udacity-didi-self-driving-car-challenge.html (I was just 3 months into DL) I had 1 x 1080 Ti card and that one also was about 3d detection using sensor fusion (lidar+cameras+etc). Dataset was in the order of 20-30 Gb IIRC. I managed to finish #4 (against 2000+ teams) but you have to play your cards very wisely. My expectatin is that in this competition computing resources are going to be a deciding factor for gold medals., especially because no restrictions have been posed about real-time inference, which I find very weird, see https://www.kaggle.com/c/3d-object-detection-for-autonomous-vehicles/discussion/108655#latest-625470
", "votes": null, "replies": [ { "id": 625508, "author_name": "jesucristo", "author_url": "", "post_date": "09/13/2019 06:32:28", "content": "Thank you so much for share your experience! Good to know :)\nBtw, is your solution public? I mean , Udacity competition's solution.
", "votes": null, "replies": [] }, { "id": 625543, "author_name": "antorsae", "author_url": "", "post_date": "09/13/2019 07:17:30", "content": "Not source code, but I've spoken about it publicly in some events, some which are available online (in Spanish). In the case of Didi it was a novel solution which allowed to hit #4 just with lidar (other teams used lidar + camera).
", "votes": null, "replies": [] }, { "id": 627320, "author_name": "kargarisaac", "author_url": "", "post_date": "09/15/2019 19:07:21", "content": "Thank you for sharing your experience. I want to work on this challenge and want to buy a pc with rtx 2080 Ti card. Do you think it is enough for such works? Or I have to get multiple GPUs?\nmaybe couple of 1080 Ti instead of one 2080 Ti.
", "votes": null, "replies": [] } ] }, { "id": 625514, "author_name": "chanway", "author_url": "", "post_date": "09/13/2019 06:39:33", "content": "Would Google Colab be a good alternatives?
", "votes": null, "replies": [ { "id": 625542, "author_name": "jesucristo", "author_url": "", "post_date": "09/13/2019 07:16:28", "content": "I don't think so. Read this thread: Why does Google Colab randomly disconnects before 12 hours\nfree TPUs is amazing, but I have no idea how to use them properly.
", "votes": null, "replies": [] }, { "id": 625732, "author_name": "gogo827jz", "author_url": "", "post_date": "09/13/2019 11:38:12", "content": "TPU has a big restriction on coding. I receive lots of error and just give it up...
", "votes": null, "replies": [] }, { "id": 625921, "author_name": "jesucristo", "author_url": "", "post_date": "09/13/2019 15:30:29", "content": "True, clean the code and find bugs is crazy... and play with the precission... hope someone starts playing with TPU and post results/notebooks....
", "votes": null, "replies": [] }, { "id": 626435, "author_name": "mariammohamed", "author_url": "", "post_date": "09/14/2019 10:31:56", "content": "I use google colab, I save the model every now and then, upload to the drive. When it disconnects, download the trained model from the drive and continue training. it helps when you do not have your own gpu.
", "votes": null, "replies": [] } ] }, { "id": 631848, "author_name": "enricobeltramo", "author_url": "", "post_date": "09/22/2019 19:59:44", "content": "Honestly, I'm trying to compete in Severstal with only 30 hours, but I'm start to think that is impossible seriously compete with that limitations. You can try if you want to learn something in application, but better to don't have much expectation about to finish in a highter position.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "625154": "I don't know too much about this topic but I'd love to learn about this :)\nIf you know about this topic or you've read about this ... I'd like to know:\n\n- which kind of resources do we need?\n- is possible to compete with the 30hr GPU/week ? or at least to participate?\n\nI guess is not possible, but even if I have to invest here all my hours in order to learn about this topic, I'll do. So I'd like to know your opinion.\n\n", "625212": "I've worked on 3d object detection. My work was on lung nodule detection (3d bounding box prediction for cancerous nodules) in CT scans (512x512xn sized images where n was in range of 300-600), I used patches of 128^3 as input data in a RetinaNet framework (adapted to 3D domain). I remember, I used about batch size of 4/8 on a GTX 1080 Ti (11 GB) and it worked well. \n\nAs far as this competition is concerned, I'll have to look at the data (Data download/kernel creation is not working properly as of now) before jumping to any conclusion regarding usability of Kaggle kernels for this competition.", "625223": "rishabhiitbhu thank you so much for sharing your experience. Really helpful!", "625253": "I am not fortunate enough to own a GPU for Kaggle competitions. Will see if the new quota will be bearable (one error in your code and 1/4 of your weekly quota is gone... also can only experiment 2 models at the same time).", "625379": "The crying frog makes me sad *lol* .... maybe they will distribute 1000$ GCP credit to help us get some hardware muscle ? Cross fingers!", "625466": "My first ever competition was this one: https://blog.udacity.com/2017/03/udacity-didi-self-driving-car-challenge.html (I was just 3 months into DL) I had 1 x 1080 Ti card and that one also was about 3d detection using sensor fusion (lidar+cameras+etc). Dataset was in the order of 20-30 Gb IIRC. I managed to finish #4 (against 2000+ teams) but you have to play your cards very wisely. My expectatin is that in this competition computing resources are going to be a deciding factor for gold medals., especially because no restrictions have been posed about real-time inference, which I find very weird, see https://www.kaggle.com/c/3d-object-detection-for-autonomous-vehicles/discussion/108655#latest-625470", "625508": "Thank you so much for share your experience! Good to know :)\nBtw, is your solution public? I mean , Udacity competition's solution.", "625514": "Would Google Colab be a good alternatives?", "625542": "I don't think so. Read this thread: [Why does Google Colab randomly disconnects before 12 hours](https://stackoverflow.com/questions/52670872/why-does-google-colab-randomly-disconnects-before-12-hours)\nfree TPUs is amazing, but I have no idea how to use them properly.", "625543": "Not source code, but I've spoken about it publicly in some events, some which are available online (in Spanish). In the case of Didi it was a novel solution which allowed to hit #4 just with lidar (other teams used lidar + camera).", "625732": "TPU has a big restriction on coding. I receive lots of error and just give it up...", "625918": "> RTX 1080 Ti (11 GB)\n\nInteresting 😉", "625919": "I'm 100% sure they will not give 1000$ , and without overfitting hahahahha :)\ncross the fingers and maybe wr get 300$. In this competition only around 80 -100teams are going to compete (because they can, check the Google comps)", "625921": "True, clean the code and find bugs is crazy... and play with the precission... hope someone starts playing with TPU and post results/notebooks....", "625922": "Fixed it :p", "626435": "I use google colab, I save the model every now and then, upload to the drive. When it disconnects, download the trained model from the drive and continue training. it helps when you do not have your own gpu.", "627320": "Thank you for sharing your experience. I want to work on this challenge and want to buy a pc with rtx 2080 Ti card. Do you think it is enough for such works? Or I have to get multiple GPUs?\nmaybe couple of 1080 Ti instead of one 2080 Ti.", "631848": "Honestly, I'm trying to compete in Severstal with only 30 hours, but I'm start to think that is impossible seriously compete with that limitations. You can try if you want to learn something in application, but better to don't have much expectation about to finish in a highter position." }, "source": "meta" }![https://media1.tenor.com/images/9af263232d5efef9c62679bc2d0b8ddd/tenor.gif?itemid=14627355](\"https://media1.tenor.com/images/9af263232d5efef9c62679bc2d0b8ddd/tenor.gif?itemid=14627355\"){kind=link}