i want to make a graph of ranking vs GPU. Can kagglers provided details? I start with mine as an example:
\n\nrank: 13-LB (0.9970), 14 models submitted. 4 machines: \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\"
", "messages": [ { "id": 224949, "postDate": "2017-09-28T00:35:55.307Z", "content": "i want to make a graph of ranking vs GPU. Can kagglers provided details? I start with mine as an example:
\n\nrank: 13-LB (0.9970), 14 models submitted. 4 machines: \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\"
", "rawMarkdown": "i want to make a graph of ranking vs GPU. Can kagglers provided details? I start with mine as an example:\n\nrank: 13-LB (0.9970), 14 models submitted. 4 machines: \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\"", "votes": 29 }, { "id": 225160, "postDate": "2017-09-28T11:50:39.053Z", "content": "After training nearly around the clock, I started getting these daily warnings from my power company.\n\"Appliances may be working harder than expected.\" Ummm.... yes.
", "rawMarkdown": "After training nearly around the clock, I started getting these daily warnings from my power company.\n\"Appliances may be working harder than expected.\" Ummm.... yes.\n\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/225160/7422/warning%20letter.jpg", "votes": 20, "replies": [ { "id": 249270, "postDate": "2017-11-28T03:44:28.310Z", "content": "THAT'S HILARIOUS!
", "rawMarkdown": "THAT'S HILARIOUS!" } ] }, { "id": 225088, "postDate": "2017-09-28T08:04:46.920Z", "content": "A quick sort out:
\n\n\n", "rawMarkdown": "A quick sort out:\n>Rank 1st, Roughly 20 GPUs total. ~200 TFlops \nRank 3 , 2x1080Ti. \nRank 4, 5x1080Ti, 2x1080, 1x1070 \nRank 5, 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070. Total = 23.5TFLOPS . \nRank 7, 6 x Titan X Maxwell 12GB \nRank 10, 2 machines: 1- 4x 1080Ti, 1- 2x 1080 Ti. \nRank 11, 1 x 1080 Ti \nRank 12, DGX-Station(4 x Tesla P100). \nrank: 13, \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\" \nRank 20, 1 x GTX 1060 \nRank 25, 1 machine: GTX 1060 \nRank 28, 1 gtx1080 + 2 gtx 1060s \nRank 39, 1 machine with 2*GTX 1080. \nRank 70, 1 x 980 Ti. ", "votes": 11 }, { "id": 225065, "postDate": "2017-09-28T06:37:23.547Z", "content": "Rank 1st, Roughly 20 GPUs total. ~200 TFlops
\n
\n Rank 3 , 2x1080Ti.
\n Rank 4, 5x1080Ti, 2x1080, 1x1070
\n Rank 5, 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070. Total = 23.5TFLOPS .
\n Rank 7, 6 x Titan X Maxwell 12GB
\n Rank 10, 2 machines: 1- 4x 1080Ti, 1- 2x 1080 Ti.
\n Rank 11, 1 x 1080 Ti
\n Rank 12, DGX-Station(4 x Tesla P100).
\n rank: 13, \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\"
\n Rank 20, 1 x GTX 1060
\n Rank 25, 1 machine: GTX 1060
\n Rank 28, 1 gtx1080 + 2 gtx 1060s
\n Rank 39, 1 machine with 2*GTX 1080.
\n Rank 70, 1 x 980 Ti.
Someone should ask nvidia to start a kaggle compeition with gpu as prizes
", "rawMarkdown": "Someone should ask nvidia to start a kaggle compeition with gpu as prizes", "votes": 12, "replies": [ { "id": 225069, "postDate": "2017-09-28T06:48:30.003Z", "content": "That would be, the best competition in the history of data science competitions. Since I think the turnout would be huge because compared to just money, some data-science related gift is highly motivating for beginners and masters alike. I am excluding Grandmasters since these guys, according to me, are freaking gods, they do not need all this stuff for motivation ;) .
", "rawMarkdown": "That would be, the best competition in the history of data science competitions. Since I think the turnout would be huge because compared to just money, some data-science related gift is highly motivating for beginners and masters alike. I am excluding Grandmasters since these guys, according to me, are freaking gods, they do not need all this stuff for motivation ;) ." }, { "id": 225070, "postDate": "2017-09-28T06:52:10.760Z", "content": "Win a dgx here!
\n\nhttps://www.crowdai.org/challenges/nips-2017-learning-to-run
", "rawMarkdown": "Win a dgx here!\n\nhttps://www.crowdai.org/challenges/nips-2017-learning-to-run", "votes": 2 }, { "id": 225209, "postDate": "2017-09-28T14:12:12.037Z", "content": "\n\n\nHeng CherKeng wrote
\n \n\n\n \nWin a dgx here!
\nhttps://www.crowdai.org/challenges/nips-2017-learning-to-run
\n
Brb, upgrading my computer.
", "rawMarkdown": "\n> **Heng CherKeng wrote**\n> \n> > Win a dgx here!\n> \n> https://www.crowdai.org/challenges/nips-2017-learning-to-run\n\nBrb, upgrading my computer.", "votes": 3 }, { "id": 231079, "postDate": "2017-10-13T16:21:16.083Z", "content": "heh :)
", "rawMarkdown": "heh :)" } ] }, { "id": 224951, "postDate": "2017-09-28T00:40:46.813Z", "content": "Rank 11, LB 0.9971. Ensemble of 7 folds from one model. Single 1080 Ti
", "rawMarkdown": "Rank 11, LB 0.9971. Ensemble of 7 folds from one model. Single 1080 Ti", "votes": 12, "replies": [ { "id": 224954, "postDate": "2017-09-28T00:52:12.903Z", "content": "That is quite impressive. Would you please mind sharing a bit details of your model, or at least some lessons you learned in this competition? Thanks!
", "rawMarkdown": "That is quite impressive. Would you please mind sharing a bit details of your model, or at least some lessons you learned in this competition? Thanks!", "votes": 6 }, { "id": 225005, "postDate": "2017-09-28T03:30:47.060Z", "content": "interested in your 7 folds ensemble while my trail is just an average of 4 runs.
", "rawMarkdown": "interested in your 7 folds ensemble while my trail is just an average of 4 runs.", "votes": 2 }, { "id": 225017, "postDate": "2017-09-28T03:55:46.657Z", "content": "I used 10-fold split for train/val.\nAnd I got my best single model just 5 days ago and it took 1.5 days to train. I just needed another model to ensemble but didn't have enough time to train another model again from scratch.
\n\nI understand the possibility of heavy overfitting, but I decided to just reuse weight from fold0 with fold 1 train data (took 6 hour to plateau). Luckily, each additional fold seems to improve the public score. And I was able to train 6 more models during last 5 days.
", "rawMarkdown": "I used 10-fold split for train/val.\nAnd I got my best single model just 5 days ago and it took 1.5 days to train. I just needed another model to ensemble but didn't have enough time to train another model again from scratch. \n\nI understand the possibility of heavy overfitting, but I decided to just reuse weight from fold0 with fold 1 train data (took 6 hour to plateau). Luckily, each additional fold seems to improve the public score. And I was able to train 6 more models during last 5 days. ", "votes": 4 } ] }, { "id": 225146, "postDate": "2017-09-28T11:23:36.573Z", "content": "Rank 20, LB (0.9970). Ensemble of 4 models. 1 x GTX 1060
", "rawMarkdown": "Rank 20, LB (0.9970). Ensemble of 4 models. 1 x GTX 1060", "votes": 10, "replies": [ { "id": 225179, "postDate": "2017-09-28T12:44:08.453Z", "content": "Impressive!
", "rawMarkdown": "Impressive!" } ] }, { "id": 225079, "postDate": "2017-09-28T07:25:23.213Z", "content": "3rd place (0.9972), mainly used 2x1080Ti.
", "rawMarkdown": "3rd place (0.9972), mainly used 2x1080Ti.", "votes": 10, "replies": [ { "id": 225180, "postDate": "2017-09-28T12:45:02.773Z", "content": "Also really impressive!
", "rawMarkdown": "Also really impressive!" } ] }, { "id": 224976, "postDate": "2017-09-28T02:12:24.723Z", "content": "1st Place. Roughly 20 GPUs total in our team. ~200 TFlops
", "rawMarkdown": "1st Place. Roughly 20 GPUs total in our team. ~200 TFlops", "votes": 8, "replies": [ { "id": 225159, "postDate": "2017-09-28T11:47:25.730Z", "content": "I like that you don't even know the exact number of GPUs :) Soon people will say: \"I used 4 dozen GPUs in this comp\".
\n\nP.S.: congrats on the first place!
", "rawMarkdown": "I like that you don't even know the exact number of GPUs :) Soon people will say: \"I used 4 dozen GPUs in this comp\".\n\nP.S.: congrats on the first place!", "votes": 11 }, { "id": 225666, "postDate": "2017-09-29T18:43:34.103Z", "content": "LOL :))\nThanks @Konstantin!
", "rawMarkdown": "LOL :))\nThanks @Konstantin! ", "votes": 1 } ] }, { "id": 225888, "postDate": "2017-09-30T08:51:11.203Z", "content": "I have achieved 9969 with no gpu but on 16-core cpu
", "rawMarkdown": "I have achieved 9969 with no gpu but on 16-core cpu", "votes": 7, "replies": [ { "id": 225995, "postDate": "2017-09-30T16:13:22.707Z", "content": "How long did your training take? And also how long did prediction take on your machine?
", "rawMarkdown": "How long did your training take? And also how long did prediction take on your machine?" }, { "id": 226031, "postDate": "2017-09-30T19:30:01.120Z", "content": "8 hours for 1 epoch\ncant tell you exactly about prediction time - but much faster
", "rawMarkdown": "8 hours for 1 epoch\ncant tell you exactly about prediction time - but much faster", "votes": -2 }, { "id": 226032, "postDate": "2017-09-30T19:30:27.120Z", "rawMarkdown": "", "votes": -1 }, { "id": 226135, "postDate": "2017-10-01T05:45:11.827Z", "content": "Why didn't you make a submission?
", "rawMarkdown": "Why didn't you make a submission?", "votes": -2 }, { "id": 226153, "postDate": "2017-10-01T07:31:18.913Z", "content": "Actually out team took a 6-th place
", "rawMarkdown": "Actually out team took a 6-th place" } ] }, { "id": 225225, "postDate": "2017-09-28T14:55:54.053Z", "content": "Rank 87: (LB 0.9966) Single 1280 U-net, old GTX 970.
", "rawMarkdown": "Rank 87: (LB 0.9966) Single 1280 U-net, old GTX 970.", "votes": 6, "replies": [ { "id": 225293, "postDate": "2017-09-28T17:17:58.257Z", "content": "What! This is way more impressive.
\n\nGTX 970 is only .7 TFLOPS!
", "rawMarkdown": "What! This is way more impressive.\n\nGTX 970 is only .7 TFLOPS! ", "votes": 2 }, { "id": 225353, "postDate": "2017-09-28T20:21:25.210Z", "content": "Isn't it 3.5 TFLOPS though? https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#GeForce_900_series\nhttps://www.vrfocus.com/2016/05/nvidia-geforce-gtx-1080-1070-980-ti-980-970-comparison-guide/
", "rawMarkdown": "Isn't it 3.5 TFLOPS though? https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#GeForce_900_series\nhttps://www.vrfocus.com/2016/05/nvidia-geforce-gtx-1080-1070-980-ti-980-970-comparison-guide/" }, { "id": 225359, "postDate": "2017-09-28T20:30:20.910Z", "content": "Yes, you're right. I mis-read the table.
", "rawMarkdown": "Yes, you're right. I mis-read the table. \n\n" }, { "id": 225684, "postDate": "2017-09-29T20:16:18.490Z", "content": "Wow, same 970 here, but only got LB 0.9838. I have to resized each image to 128*128. How did you make it using 970???
", "rawMarkdown": "Wow, same 970 here, but only got LB 0.9838. I have to resized each image to 128*128. How did you make it using 970???" } ] }, { "id": 225121, "postDate": "2017-09-28T09:54:16.127Z", "content": "4th, 0.9972. As our team name says, we did it using roughly 80 TFlops (5x1080Ti, 2x1080, 1x1070). At some point at the end of the competition, we were even considering on using a 60x1070 mining rig for TTA :) But it turned out being not suitable for Deep Learning, so that was transformed into a local meme and didn't go further :)
", "rawMarkdown": "4th, 0.9972. As our team name says, we did it using roughly 80 TFlops (5x1080Ti, 2x1080, 1x1070). At some point at the end of the competition, we were even considering on using a 60x1070 mining rig for TTA :) But it turned out being not suitable for Deep Learning, so that was transformed into a local meme and didn't go further :) ", "votes": 6 }, { "id": 225175, "postDate": "2017-09-28T12:28:35.693Z", "content": "Rank: 126 (LB 0.9965) Single 1024 Unet with some test augmentation.
\n\nI don't have GPU so I rented ec2 p2 with K80. I found that the K80 was quite slow compared to other reported training times. This was my first deep learning competition so I had quite a few challenges. It was surprising that the lack of GPU was among my biggest bottlenecks. In other competitions I got used to be able to rent quite competitive machines from aws (Memory >= 100GB vCPU>=16 ~ 2 USD/day).
", "rawMarkdown": "Rank: 126 (LB 0.9965) Single 1024 Unet with some test augmentation.\n\nI don't have GPU so I rented ec2 p2 with K80. I found that the K80 was quite slow compared to other reported training times. This was my first deep learning competition so I had quite a few challenges. It was surprising that the lack of GPU was among my biggest bottlenecks. In other competitions I got used to be able to rent quite competitive machines from aws (Memory >= 100GB vCPU>=16 ~ 2 USD/day).", "votes": 3, "replies": [ { "id": 225358, "postDate": "2017-09-28T20:29:44.190Z", "content": "K80 should be way more faster than a regular GPU, but the K80 in EC2 p2 is actually a kind of virtual one. Even GTX 1080 is four times faster than a AWS K80.
", "rawMarkdown": "K80 should be way more faster than a regular GPU, but the K80 in EC2 p2 is actually a kind of virtual one. Even GTX 1080 is four times faster than a AWS K80.", "votes": 1 }, { "id": 225362, "postDate": "2017-09-28T20:37:33.883Z", "content": "Not exactly like that. Despite being freakingly expensive K80 isn't faster that usual gaming GPU's when we talk about DL. No matter physical or virtual (there is very little overhead, in range of single percents). This is the case because we need single precision fp32 flops. Double precision 64bit calculations is a whole another story, K80 is about 30 times faster than any GTX GPU. And it has several other features important in server/cloud computing, like ECC ram and special form-factor, etc.
", "rawMarkdown": "Not exactly like that. Despite being freakingly expensive K80 isn't faster that usual gaming GPU's when we talk about DL. No matter physical or virtual (there is very little overhead, in range of single percents). This is the case because we need single precision fp32 flops. Double precision 64bit calculations is a whole another story, K80 is about 30 times faster than any GTX GPU. And it has several other features important in server/cloud computing, like ECC ram and special form-factor, etc. ", "votes": 3 }, { "id": 225369, "postDate": "2017-09-28T21:00:07.737Z", "content": "So you mean even we have a physical K80 in our local computer, it wouldn't be faster than a GTX 1080TI in terms of DL computing?
", "rawMarkdown": "So you mean even we have a physical K80 in our local computer, it wouldn't be faster than a GTX 1080TI in terms of DL computing?" }, { "id": 225482, "postDate": "2017-09-29T07:42:17.370Z", "content": "Yes, it would run much slower unless you have 64bit computing task.
", "rawMarkdown": "Yes, it would run much slower unless you have 64bit computing task.", "votes": 1 }, { "id": 225593, "postDate": "2017-09-29T15:25:13.930Z", "content": "Got it. Thanks!
", "rawMarkdown": "Got it. Thanks!" } ] }, { "id": 225012, "postDate": "2017-09-28T03:50:05.817Z", "content": "5th - Private 0.9972. 7 models ensemble. 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070 (used this for parallel TTA inference only since it was a slow PC). Total = 11.3 + 9 + 6.5/2 = 23.5TFLOPS . Surely there is a strong positive correlation between GPU power and rank in this competition (with some exceptions) ! And I regret not having teamed up :)
\n\nI'll write up soon since I have plenty of time to burn these days.
", "rawMarkdown": "5th - Private 0.9972. 7 models ensemble. 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070 (used this for parallel TTA inference only since it was a slow PC). Total = 11.3 + 9 + 6.5/2 = 23.5TFLOPS . Surely there is a strong positive correlation between GPU power and rank in this competition (with some exceptions) ! And I regret not having teamed up :)\n\nI'll write up soon since I have plenty of time to burn these days.", "votes": 4 }, { "id": 225010, "postDate": "2017-09-28T03:44:29.810Z", "content": "WTF the hardware resources here seem astronomical. Do you guys own all these hardware or was it just rented hardware?
", "rawMarkdown": "WTF the hardware resources here seem astronomical. Do you guys own all these hardware or was it just rented hardware?", "votes": 4, "replies": [ { "id": 225029, "postDate": "2017-09-28T04:16:10.410Z", "content": "I have 2 x Pascal Titan X and 4 x GTX 1080 Ti
\n\nIt may sound like a lot, but sadly I pay $4000 per month for rent for a one bedroom appartment :( and with respect to this investing into GPUs does not look that expensive.
\n\nAnd one may think about it as about investment, this year I won:
\n\nSo, in terms of money, it does not look that bad.
\n\nI still had hope that it may help in terms of a job search, but it is not the case so far :( I get emails like this more often than I would like:
\n\n\nA Googler recently referred you for the Research Scientist, Google Brain (United States) role. We carefully reviewed your background and experience, and decided not to proceed with your application at this time.\n
That last bit of email was indeed painful. But I think you will definitely land something pretty soon :) (If you haven't already).
", "rawMarkdown": "That last bit of email was indeed painful. But I think you will definitely land something pretty soon :) (If you haven't already). ", "votes": 1 } ] }, { "id": 224958, "postDate": "2017-09-28T01:14:02.433Z", "content": "Rank 10, LB 0.9971. Ensemble of 3 different Unet models: 1. predict background stage from inverted mask then re-invert result , 2. predict car, 3. split image top/bottom. 2 machines: 1- 4x 1080Ti (had all 4 watercooled during competition to prevent thermal throttling), 1 -2x 1080 Ti.
", "rawMarkdown": "Rank 10, LB 0.9971. Ensemble of 3 different Unet models: 1. predict background stage from inverted mask then re-invert result , 2. predict car, 3. split image top/bottom. 2 machines: 1- 4x 1080Ti (had all 4 watercooled during competition to prevent thermal throttling), 1 -2x 1080 Ti.", "votes": 4 }, { "id": 227102, "postDate": "2017-10-03T18:05:40.393Z", "content": "rank: 62-Private (0.9967), 3 models ensembled. 1 machine w/ 1x1080
\n\nBTW, this competition is less powerplay than I thought... (So many 1060s!)
", "rawMarkdown": "rank: 62-Private (0.9967), 3 models ensembled. 1 machine w/ 1x1080\n\nBTW, this competition is less powerplay than I thought... (So many 1060s!)", "votes": 1 }, { "id": 225337, "postDate": "2017-09-28T19:10:40.783Z", "content": "78th TitanX, simple unet without using cv2.resize fonction at training or testing.
", "rawMarkdown": "78th TitanX, simple unet without using cv2.resize fonction at training or testing.", "votes": 1, "replies": [ { "id": 225497, "postDate": "2017-09-29T08:52:11.350Z", "content": "How did you upsample back to the proper size?
", "rawMarkdown": "How did you upsample back to the proper size?" }, { "id": 225564, "postDate": "2017-09-29T14:15:11.423Z", "content": "Take original picture 1918x1280, add 1 or 65 (to give more context to unet) horizontally using cv2.copyMakeBorder(img,0,0,1,1 ,cv2.BORDER_REFLECT_101) and use Vladimir Iglovikov' trick using Cropping2D (1 or 65) within the model to reduce the vector to the final mask size of 1918x1280. The advantage of adding 65 horizontally is to avoid the weighted dice metric since the border is cut out before calculating the loss.
\n\nI unfortunately only had time to run one fold and it killed me on private lb as expected but gave a nice jump in score.
", "rawMarkdown": "Take original picture 1918x1280, add 1 or 65 (to give more context to unet) horizontally using cv2.copyMakeBorder(img,0,0,1,1 ,cv2.BORDER_REFLECT_101) and use Vladimir Iglovikov' [trick][1] using Cropping2D (1 or 65) within the model to reduce the vector to the final mask size of 1918x1280. The advantage of adding 65 horizontally is to avoid the weighted dice metric since the border is cut out before calculating the loss.\n\nI unfortunately only had time to run one fold and it killed me on private lb as expected but gave a nice jump in score.\n\n\n [1]: https://www.kaggle.com/c/dstl-satellite-imagery-feature-detection/discussion/29789#166405" }, { "id": 226921, "postDate": "2017-10-03T10:43:56.910Z", "content": "hi absolutelyNoWarranty
\n\nSorry for my newbie question, I am just wondering why it has to be unsampling at the end and can't we instead make masks at cropped images and \"sew\" the together at the end for the whole image?
", "rawMarkdown": "hi absolutelyNoWarranty\n\nSorry for my newbie question, I am just wondering why it has to be unsampling at the end and can't we instead make masks at cropped images and \"sew\" the together at the end for the whole image?" }, { "id": 226994, "postDate": "2017-10-03T14:04:45.107Z", "content": "I don't have a good answer to your question; the goal of the competition is to find what methods work best and are the most robust. \"crop and \"sew\"\" didn't work for me in this particular competition but worked in a previous competition. Try something, if it fails, try something different.
", "rawMarkdown": "I don't have a good answer to your question; the goal of the competition is to find what methods work best and are the most robust. \"crop and \"sew\"\" didn't work for me in this particular competition but worked in a previous competition. Try something, if it fails, try something different." }, { "id": 227045, "postDate": "2017-10-03T15:51:29.197Z", "content": "that's right of course. i am asking because upsampling approach seems to be the standard approach and \"crop and sew\" was not even mentioned.
\n\n\"Intuitively\" i imagine \"crop and sew\" may yield more accurate results but apparently my intuition is very wrong and i am just trying to figure out .
", "rawMarkdown": "that's right of course. i am asking because upsampling approach seems to be the standard approach and \"crop and sew\" was not even mentioned. \n\n\"Intuitively\" i imagine \"crop and sew\" may yield more accurate results but apparently my intuition is very wrong and i am just trying to figure out ." } ] }, { "id": 225263, "postDate": "2017-09-28T16:03:29.557Z", "content": "Rank 70 - LB (0.9966), 1x1080Ti, 5 model ensemble (unet and densenet)
", "rawMarkdown": "Rank 70 - LB (0.9966), 1x1080Ti, 5 model ensemble (unet and densenet)", "votes": 1 }, { "id": 225134, "postDate": "2017-09-28T10:34:40.710Z", "content": "Rank - 42. 1 x GTX 1060.\nSingle model:\nPublic LB - 0.9968, Private LB - 0.9967
\n\nMade a submission using an ensemble of three models but missed the deadline :(\nEnsemble: \nPublic LB - 0.9969, Private LB - 0.9968
", "rawMarkdown": "Rank - 42. 1 x GTX 1060.\nSingle model:\nPublic LB - 0.9968, Private LB - 0.9967\n\nMade a submission using an ensemble of three models but missed the deadline :(\nEnsemble: \nPublic LB - 0.9969, Private LB - 0.9968", "votes": 1 }, { "id": 225107, "postDate": "2017-09-28T09:16:57.090Z", "content": "Public LB: Rank 64/0.9969 Private LB: Rank 33/0.9968. Unet with multiscale and one CVPR2017 model. 2 x gtx 1080
", "rawMarkdown": "Public LB: Rank 64/0.9969 Private LB: Rank 33/0.9968. Unet with multiscale and one CVPR2017 model. 2 x gtx 1080", "votes": 1 }, { "id": 225006, "postDate": "2017-09-28T03:32:55.387Z", "content": "28 - single model, 1 gtx1080 + 2 gtx 1060s
", "rawMarkdown": "28 - single model, 1 gtx1080 + 2 gtx 1060s", "votes": 1 }, { "id": 224999, "postDate": "2017-09-28T03:13:19.107Z", "content": "Public LB: Rank 87/0.9968; Private LB: Rank 129/0.9964. Ensemble of 5 folds from 3 different u-net models. One machine with 1*GTX 1070.
", "rawMarkdown": "Public LB: Rank 87/0.9968; Private LB: Rank 129/0.9964. Ensemble of 5 folds from 3 different u-net models. One machine with 1*GTX 1070.", "votes": 1 }, { "id": 226179, "postDate": "2017-10-01T09:44:01.960Z", "content": "6th:\n4x 1080ti,\n4x 1080ti,\n2x Titan X Maxwell,\n2x Titan X Pascal.\nAnd part time:\n4x 1080,\n4x 1080ti,\n2x Tesla P40
", "rawMarkdown": "6th:\n4x 1080ti,\n4x 1080ti,\n2x Titan X Maxwell,\n2x Titan X Pascal.\nAnd part time:\n4x 1080,\n4x 1080ti,\n2x Tesla P40", "votes": 2 }, { "id": 225806, "postDate": "2017-09-30T02:05:25.150Z", "content": "It seems that high gpu resources will become a must for some future kaggle compeition. Another example is https://www.kaggle.com/c/cdiscount-image-classification-challenge. Time to invest in gpu. But nvidia is upgrading their gpu very fast. If someone can provide gpu rental service specially designed for kagglers it would be nice
", "rawMarkdown": "It seems that high gpu resources will become a must for some future kaggle compeition. Another example is https://www.kaggle.com/c/cdiscount-image-classification-challenge. Time to invest in gpu. But nvidia is upgrading their gpu very fast. If someone can provide gpu rental service specially designed for kagglers it would be nice", "votes": 2 }, { "id": 225710, "postDate": "2017-09-29T21:24:35.333Z", "content": "3xGTX 980 Public (37th): 0.9970 -> Private 0.9966 (76th)
", "rawMarkdown": "3xGTX 980 Public (37th): 0.9970 -> Private 0.9966 (76th)", "votes": 2 }, { "id": 225563, "postDate": "2017-09-29T14:10:31.013Z", "content": "94th 1*1080Ti
", "rawMarkdown": "94th 1*1080Ti", "votes": 2 }, { "id": 225554, "postDate": "2017-09-29T13:39:00.390Z", "content": "9th, 2 x 1080Ti, 1280 U-net, 4 ensemble
", "rawMarkdown": "9th, 2 x 1080Ti, 1280 U-net, 4 ensemble", "votes": 2 }, { "id": 225449, "postDate": "2017-09-29T04:38:09.183Z", "content": "one 1080ti +one titan xp, final submission is single model,make a big mistake that be lost in overfit. no time to ensemble,its first time to do kaggle i will be better next time☺
", "rawMarkdown": "one 1080ti +one titan xp, final submission is single model,make a big mistake that be lost in overfit. no time to ensemble,its first time to do kaggle i will be better next time☺", "votes": 2 }, { "id": 225383, "postDate": "2017-09-28T21:36:12.830Z", "content": "Another interesting graph (although it would be very difficult to get the data) would be rank vs total power consumed for this contest (i.e., in kilowatt hours).
", "rawMarkdown": "Another interesting graph (although it would be very difficult to get the data) would be rank vs total power consumed for this contest (i.e., in kilowatt hours).", "votes": 2 }, { "id": 225267, "postDate": "2017-09-28T16:11:49.777Z", "content": "Public: rank 446 score 0.9933 Private: rank 428 score 0.9940\n1 GTX 660 Ti (2.7 TFlops with 3Gb of memory)
\n\nA simple average of 2 models, one very scaled down U-Net to make it fit mostly in GPU memory and another architecture based loosely on https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf \nHad things set up to do 5 training folds split by car but could only manage training on the first fold for each model before the competition ended.
", "rawMarkdown": "Public: rank 446 score 0.9933 Private: rank 428 score 0.9940\n1 GTX 660 Ti (2.7 TFlops with 3Gb of memory)\n\nA simple average of 2 models, one very scaled down U-Net to make it fit mostly in GPU memory and another architecture based loosely on https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf \nHad things set up to do 5 training folds split by car but could only manage training on the first fold for each model before the competition ended. ", "votes": 2 }, { "id": 225161, "postDate": "2017-09-28T11:52:45.017Z", "content": "7th private lb score: 6 x Titan X Maxwell 12GB
", "rawMarkdown": "7th private lb score: 6 x Titan X Maxwell 12GB", "votes": 2 }, { "id": 225141, "postDate": "2017-09-28T11:04:28.530Z", "content": "39th on private and 82nd on public (both cases 0.9968) - using 1 GTX 1070 and a single U-net, but tailored the loss s.t. boundary errors are weighted more heavily :).
", "rawMarkdown": "39th on private and 82nd on public (both cases 0.9968) - using 1 GTX 1070 and a single U-net, but tailored the loss s.t. boundary errors are weighted more heavily :).", "votes": 2 }, { "id": 225118, "postDate": "2017-09-28T09:45:03.883Z", "content": "rank: 38 (0.9968), 1 gtx 1080, public - 96 (0.9968). Ensemble of 7 Unet models, used bounding boxes with cars instead of whole images for learning.
", "rawMarkdown": "rank: 38 (0.9968), 1 gtx 1080, public - 96 (0.9968). Ensemble of 7 Unet models, used bounding boxes with cars instead of whole images for learning.", "votes": 2 }, { "id": 225066, "postDate": "2017-09-28T06:39:52.787Z", "content": "Public LB: Rank 70/0.9969; Private LB: Rank 71/0.9966. Single U-Net, 1 fold, 1 x 980 Ti.
", "rawMarkdown": "Public LB: Rank 70/0.9969; Private LB: Rank 71/0.9966. Single U-Net, 1 fold, 1 x 980 Ti.", "votes": 2 }, { "id": 225064, "postDate": "2017-09-28T06:17:31.637Z", "content": "Rank 12th (0.9970) on Private LB. DGX-Station(4 x Tesla P100). Ensemble of 6 different models.
", "rawMarkdown": "Rank 12th (0.9970) on Private LB. DGX-Station(4 x Tesla P100). Ensemble of 6 different models.", "votes": 2 }, { "id": 225041, "postDate": "2017-09-28T04:38:34.177Z", "content": "149 position. 2 GPU 1080
", "rawMarkdown": "149 position. 2 GPU 1080", "votes": 2 }, { "id": 224966, "postDate": "2017-09-28T01:52:37.067Z", "content": "rank 25-LB (0.9969). Ensemble of 5 folds from one model. 1 machine: GTX 1060
", "rawMarkdown": "rank 25-LB (0.9969). Ensemble of 5 folds from one model. 1 machine: GTX 1060", "votes": 2, "replies": [ { "id": 224972, "postDate": "2017-09-28T02:07:56.150Z", "content": "Eugene, which NN model have you used and what was the image input size?
", "rawMarkdown": "Eugene, which NN model have you used and what was the image input size?", "votes": 1 } ] }, { "id": 224952, "postDate": "2017-09-28T00:44:53.667Z", "content": "Public LB: Rank 39/0.9970; Private LB: Rank 29/0.9969. Ensembles of 8 models from 6 different models. One machine with 2*GTX 1080.
", "rawMarkdown": "Public LB: Rank 39/0.9970; Private LB: Rank 29/0.9969. Ensembles of 8 models from 6 different models. One machine with 2*GTX 1080.", "votes": 2 }, { "id": 227276, "postDate": "2017-10-04T04:49:37.077Z", "content": "I think you need to go about this in a much more stuctured way
", "rawMarkdown": "I think you need to go about this in a much more stuctured way" }, { "id": 226169, "postDate": "2017-10-01T08:54:54.093Z", "content": "nice job
", "rawMarkdown": "nice job" }, { "id": 226080, "postDate": "2017-10-01T00:24:42.880Z", "content": "good
", "rawMarkdown": "good" }, { "id": 225464, "postDate": "2017-09-29T05:42:35.780Z", "content": "Thanks @Heng for all your contributions.
\n\nRank 166 with single model.
\n\nI used floydhub GPU with Total memory: 11.17GiB\nMy best model is a unet 1280x1280, batch size 8 that converged after 29 epochs with\nloss: 0.0084 - dice_loss: 0.9957 - val_loss: 0.0108 - val_dice_loss: 0.9953
\n\nMy smallest model is a 512x512, batch size 12. Trained on Floyd but prediction was done on my CPU locally that took 8.5 days to complete. \nMy CPU device spec. :-\nIntel(R) Core(TM) i7-7500U CPU @ 2.70GHz, 2901 Mhz, 2 Core(s), 4 Logical Processor(s)\nInstalled Physical Memory (RAM) 12.0 GB
", "rawMarkdown": "Thanks @Heng for all your contributions.\n\nRank 166 with single model. \n\nI used floydhub GPU with Total memory: 11.17GiB\nMy best model is a unet 1280x1280, batch size 8 that converged after 29 epochs with\nloss: 0.0084 - dice_loss: 0.9957 - val_loss: 0.0108 - val_dice_loss: 0.9953\n\nMy smallest model is a 512x512, batch size 12. Trained on Floyd but prediction was done on my CPU locally that took 8.5 days to complete. \nMy CPU device spec. :-\nIntel(R) Core(TM) i7-7500U CPU @ 2.70GHz, 2901 Mhz, 2 Core(s), 4 Logical Processor(s)\nInstalled Physical Memory (RAM)\t12.0 GB\n\n" }, { "id": 226283, "postDate": "2017-10-01T17:42:05.197Z", "rawMarkdown": "", "votes": 3, "isDeleted": true, "replies": [ { "id": 226292, "postDate": "2017-10-01T18:05:38.443Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] } ], "comments": [ { "id": 225160, "author_name": "David Austin", "author_url": "", "post_date": "2017-09-28T11:50:39.053000", "content": "After training nearly around the clock, I started getting these daily warnings from my power company.\n\"Appliances may be working harder than expected.\" Ummm.... yes.
", "votes": 20, "replies": [ { "id": 249270, "author_name": "BrandonPippin", "author_url": "", "post_date": "2017-11-28T03:44:28.310000", "content": "THAT'S HILARIOUS!
", "votes": 0, "replies": [] } ] }, { "id": 225088, "author_name": "Eugene", "author_url": "", "post_date": "2017-09-28T08:04:46.920000", "content": "A quick sort out:
\n\n\n", "votes": 11, "replies": [] }, { "id": 225065, "author_name": "hengck23", "author_url": "", "post_date": "2017-09-28T06:37:23.547000", "content": "Rank 1st, Roughly 20 GPUs total. ~200 TFlops
\n
\n Rank 3 , 2x1080Ti.
\n Rank 4, 5x1080Ti, 2x1080, 1x1070
\n Rank 5, 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070. Total = 23.5TFLOPS .
\n Rank 7, 6 x Titan X Maxwell 12GB
\n Rank 10, 2 machines: 1- 4x 1080Ti, 1- 2x 1080 Ti.
\n Rank 11, 1 x 1080 Ti
\n Rank 12, DGX-Station(4 x Tesla P100).
\n rank: 13, \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\"
\n Rank 20, 1 x GTX 1060
\n Rank 25, 1 machine: GTX 1060
\n Rank 28, 1 gtx1080 + 2 gtx 1060s
\n Rank 39, 1 machine with 2*GTX 1080.
\n Rank 70, 1 x 980 Ti.
Someone should ask nvidia to start a kaggle compeition with gpu as prizes
", "votes": 12, "replies": [ { "id": 225069, "author_name": "remidi", "author_url": "", "post_date": "2017-09-28T06:48:30.003000", "content": "That would be, the best competition in the history of data science competitions. Since I think the turnout would be huge because compared to just money, some data-science related gift is highly motivating for beginners and masters alike. I am excluding Grandmasters since these guys, according to me, are freaking gods, they do not need all this stuff for motivation ;) .
", "votes": 0, "replies": [] }, { "id": 225070, "author_name": "hengck23", "author_url": "", "post_date": "2017-09-28T06:52:10.760000", "content": "Win a dgx here!
\n\nhttps://www.crowdai.org/challenges/nips-2017-learning-to-run
", "votes": 2, "replies": [] }, { "id": 225209, "author_name": "Alan Khoa Nguyen", "author_url": "", "post_date": "2017-09-28T14:12:12.037000", "content": "\n\n\nHeng CherKeng wrote
\n \n\n\n \nWin a dgx here!
\nhttps://www.crowdai.org/challenges/nips-2017-learning-to-run
\n
Brb, upgrading my computer.
", "votes": 3, "replies": [] }, { "id": 231079, "author_name": "Alon Burg", "author_url": "", "post_date": "2017-10-13T16:21:16.083000", "content": "heh :)
", "votes": 0, "replies": [] } ] }, { "id": 224951, "author_name": "Sukjae Cho", "author_url": "", "post_date": "2017-09-28T00:40:46.813000", "content": "Rank 11, LB 0.9971. Ensemble of 7 folds from one model. Single 1080 Ti
", "votes": 12, "replies": [ { "id": 224954, "author_name": "Hillview", "author_url": "", "post_date": "2017-09-28T00:52:12.903000", "content": "That is quite impressive. Would you please mind sharing a bit details of your model, or at least some lessons you learned in this competition? Thanks!
", "votes": 6, "replies": [] }, { "id": 225005, "author_name": "Canoe", "author_url": "", "post_date": "2017-09-28T03:30:47.060000", "content": "interested in your 7 folds ensemble while my trail is just an average of 4 runs.
", "votes": 2, "replies": [] }, { "id": 225017, "author_name": "Sukjae Cho", "author_url": "", "post_date": "2017-09-28T03:55:46.657000", "content": "I used 10-fold split for train/val.\nAnd I got my best single model just 5 days ago and it took 1.5 days to train. I just needed another model to ensemble but didn't have enough time to train another model again from scratch.
\n\nI understand the possibility of heavy overfitting, but I decided to just reuse weight from fold0 with fold 1 train data (took 6 hour to plateau). Luckily, each additional fold seems to improve the public score. And I was able to train 6 more models during last 5 days.
", "votes": 4, "replies": [] } ] }, { "id": 225146, "author_name": "Igor Praznik", "author_url": "", "post_date": "2017-09-28T11:23:36.573000", "content": "Rank 20, LB (0.9970). Ensemble of 4 models. 1 x GTX 1060
", "votes": 10, "replies": [ { "id": 225179, "author_name": "Tim Joseph", "author_url": "", "post_date": "2017-09-28T12:44:08.453000", "content": "Impressive!
", "votes": 0, "replies": [] } ] }, { "id": 225079, "author_name": "lyakaap", "author_url": "", "post_date": "2017-09-28T07:25:23.213000", "content": "3rd place (0.9972), mainly used 2x1080Ti.
", "votes": 10, "replies": [ { "id": 225180, "author_name": "Tim Joseph", "author_url": "", "post_date": "2017-09-28T12:45:02.773000", "content": "Also really impressive!
", "votes": 0, "replies": [] } ] }, { "id": 224976, "author_name": "Artem.Sanakoev", "author_url": "", "post_date": "2017-09-28T02:12:24.723000", "content": "1st Place. Roughly 20 GPUs total in our team. ~200 TFlops
", "votes": 8, "replies": [ { "id": 225159, "author_name": "Konstantin Lopukhin", "author_url": "", "post_date": "2017-09-28T11:47:25.730000", "content": "I like that you don't even know the exact number of GPUs :) Soon people will say: \"I used 4 dozen GPUs in this comp\".
\n\nP.S.: congrats on the first place!
", "votes": 11, "replies": [] }, { "id": 225666, "author_name": "Artem.Sanakoev", "author_url": "", "post_date": "2017-09-29T18:43:34.103000", "content": "LOL :))\nThanks @Konstantin!
", "votes": 1, "replies": [] } ] }, { "id": 225888, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2017-09-30T08:51:11.203000", "content": "I have achieved 9969 with no gpu but on 16-core cpu
", "votes": 7, "replies": [ { "id": 225995, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2017-09-30T16:13:22.707000", "content": "How long did your training take? And also how long did prediction take on your machine?
", "votes": 0, "replies": [] }, { "id": 226031, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2017-09-30T19:30:01.120000", "content": "8 hours for 1 epoch\ncant tell you exactly about prediction time - but much faster
", "votes": -2, "replies": [] }, { "id": 226032, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2017-09-30T19:30:27.120000", "content": "", "votes": -1, "replies": [] }, { "id": 226135, "author_name": "Joshua Wonser", "author_url": "", "post_date": "2017-10-01T05:45:11.827000", "content": "Why didn't you make a submission?
", "votes": -2, "replies": [] }, { "id": 226153, "author_name": "ValeriyBabushkin", "author_url": "", "post_date": "2017-10-01T07:31:18.913000", "content": "Actually out team took a 6-th place
", "votes": 0, "replies": [] } ] }, { "id": 225225, "author_name": "DmitryKustikov", "author_url": "", "post_date": "2017-09-28T14:55:54.053000", "content": "Rank 87: (LB 0.9966) Single 1280 U-net, old GTX 970.
", "votes": 6, "replies": [ { "id": 225293, "author_name": "Sukjae Cho", "author_url": "", "post_date": "2017-09-28T17:17:58.257000", "content": "What! This is way more impressive.
\n\nGTX 970 is only .7 TFLOPS!
", "votes": 2, "replies": [] }, { "id": 225353, "author_name": "xelibrion", "author_url": "", "post_date": "2017-09-28T20:21:25.210000", "content": "Isn't it 3.5 TFLOPS though? https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#GeForce_900_series\nhttps://www.vrfocus.com/2016/05/nvidia-geforce-gtx-1080-1070-980-ti-980-970-comparison-guide/
", "votes": 0, "replies": [] }, { "id": 225359, "author_name": "Sukjae Cho", "author_url": "", "post_date": "2017-09-28T20:30:20.910000", "content": "Yes, you're right. I mis-read the table.
", "votes": 0, "replies": [] }, { "id": 225684, "author_name": "xiqin", "author_url": "", "post_date": "2017-09-29T20:16:18.490000", "content": "Wow, same 970 here, but only got LB 0.9838. I have to resized each image to 128*128. How did you make it using 970???
", "votes": 0, "replies": [] } ] }, { "id": 225121, "author_name": "Nikolay Shebanov", "author_url": "", "post_date": "2017-09-28T09:54:16.127000", "content": "4th, 0.9972. As our team name says, we did it using roughly 80 TFlops (5x1080Ti, 2x1080, 1x1070). At some point at the end of the competition, we were even considering on using a 60x1070 mining rig for TTA :) But it turned out being not suitable for Deep Learning, so that was transformed into a local meme and didn't go further :)
", "votes": 6, "replies": [] }, { "id": 225175, "author_name": "beluga", "author_url": "", "post_date": "2017-09-28T12:28:35.693000", "content": "Rank: 126 (LB 0.9965) Single 1024 Unet with some test augmentation.
\n\nI don't have GPU so I rented ec2 p2 with K80. I found that the K80 was quite slow compared to other reported training times. This was my first deep learning competition so I had quite a few challenges. It was surprising that the lack of GPU was among my biggest bottlenecks. In other competitions I got used to be able to rent quite competitive machines from aws (Memory >= 100GB vCPU>=16 ~ 2 USD/day).
", "votes": 3, "replies": [ { "id": 225358, "author_name": "wzfstat", "author_url": "", "post_date": "2017-09-28T20:29:44.190000", "content": "K80 should be way more faster than a regular GPU, but the K80 in EC2 p2 is actually a kind of virtual one. Even GTX 1080 is four times faster than a AWS K80.
", "votes": 1, "replies": [] }, { "id": 225362, "author_name": "Cut Onion", "author_url": "", "post_date": "2017-09-28T20:37:33.883000", "content": "Not exactly like that. Despite being freakingly expensive K80 isn't faster that usual gaming GPU's when we talk about DL. No matter physical or virtual (there is very little overhead, in range of single percents). This is the case because we need single precision fp32 flops. Double precision 64bit calculations is a whole another story, K80 is about 30 times faster than any GTX GPU. And it has several other features important in server/cloud computing, like ECC ram and special form-factor, etc.
", "votes": 3, "replies": [] }, { "id": 225369, "author_name": "wzfstat", "author_url": "", "post_date": "2017-09-28T21:00:07.737000", "content": "So you mean even we have a physical K80 in our local computer, it wouldn't be faster than a GTX 1080TI in terms of DL computing?
", "votes": 0, "replies": [] }, { "id": 225482, "author_name": "Cut Onion", "author_url": "", "post_date": "2017-09-29T07:42:17.370000", "content": "Yes, it would run much slower unless you have 64bit computing task.
", "votes": 1, "replies": [] }, { "id": 225593, "author_name": "wzfstat", "author_url": "", "post_date": "2017-09-29T15:25:13.930000", "content": "Got it. Thanks!
", "votes": 0, "replies": [] } ] }, { "id": 225012, "author_name": "Kyle Lee", "author_url": "", "post_date": "2017-09-28T03:50:05.817000", "content": "5th - Private 0.9972. 7 models ensemble. 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070 (used this for parallel TTA inference only since it was a slow PC). Total = 11.3 + 9 + 6.5/2 = 23.5TFLOPS . Surely there is a strong positive correlation between GPU power and rank in this competition (with some exceptions) ! And I regret not having teamed up :)
\n\nI'll write up soon since I have plenty of time to burn these days.
", "votes": 4, "replies": [] }, { "id": 225010, "author_name": "kirk", "author_url": "", "post_date": "2017-09-28T03:44:29.810000", "content": "WTF the hardware resources here seem astronomical. Do you guys own all these hardware or was it just rented hardware?
", "votes": 4, "replies": [ { "id": 225029, "author_name": "Vladimir Iglovikov", "author_url": "", "post_date": "2017-09-28T04:16:10.410000", "content": "I have 2 x Pascal Titan X and 4 x GTX 1080 Ti
\n\nIt may sound like a lot, but sadly I pay $4000 per month for rent for a one bedroom appartment :( and with respect to this investing into GPUs does not look that expensive.
\n\nAnd one may think about it as about investment, this year I won:
\n\nSo, in terms of money, it does not look that bad.
\n\nI still had hope that it may help in terms of a job search, but it is not the case so far :( I get emails like this more often than I would like:
\n\n\nA Googler recently referred you for the Research Scientist, Google Brain (United States) role. We carefully reviewed your background and experience, and decided not to proceed with your application at this time.\n
That last bit of email was indeed painful. But I think you will definitely land something pretty soon :) (If you haven't already).
", "votes": 1, "replies": [] } ] }, { "id": 224958, "author_name": "David Austin", "author_url": "", "post_date": "2017-09-28T01:14:02.433000", "content": "Rank 10, LB 0.9971. Ensemble of 3 different Unet models: 1. predict background stage from inverted mask then re-invert result , 2. predict car, 3. split image top/bottom. 2 machines: 1- 4x 1080Ti (had all 4 watercooled during competition to prevent thermal throttling), 1 -2x 1080 Ti.
", "votes": 4, "replies": [] }, { "id": 227102, "author_name": "Chia-Ta Tsai", "author_url": "", "post_date": "2017-10-03T18:05:40.393000", "content": "rank: 62-Private (0.9967), 3 models ensembled. 1 machine w/ 1x1080
\n\nBTW, this competition is less powerplay than I thought... (So many 1060s!)
", "votes": 1, "replies": [] }, { "id": 225337, "author_name": "eagle4", "author_url": "", "post_date": "2017-09-28T19:10:40.783000", "content": "78th TitanX, simple unet without using cv2.resize fonction at training or testing.
", "votes": 1, "replies": [ { "id": 225497, "author_name": "absolutelyNoWarranty", "author_url": "", "post_date": "2017-09-29T08:52:11.350000", "content": "How did you upsample back to the proper size?
", "votes": 0, "replies": [] }, { "id": 225564, "author_name": "eagle4", "author_url": "", "post_date": "2017-09-29T14:15:11.423000", "content": "Take original picture 1918x1280, add 1 or 65 (to give more context to unet) horizontally using cv2.copyMakeBorder(img,0,0,1,1 ,cv2.BORDER_REFLECT_101) and use Vladimir Iglovikov' trick using Cropping2D (1 or 65) within the model to reduce the vector to the final mask size of 1918x1280. The advantage of adding 65 horizontally is to avoid the weighted dice metric since the border is cut out before calculating the loss.
\n\nI unfortunately only had time to run one fold and it killed me on private lb as expected but gave a nice jump in score.
", "votes": 0, "replies": [] }, { "id": 226921, "author_name": "YL", "author_url": "", "post_date": "2017-10-03T10:43:56.910000", "content": "hi absolutelyNoWarranty
\n\nSorry for my newbie question, I am just wondering why it has to be unsampling at the end and can't we instead make masks at cropped images and \"sew\" the together at the end for the whole image?
", "votes": 0, "replies": [] }, { "id": 226994, "author_name": "eagle4", "author_url": "", "post_date": "2017-10-03T14:04:45.107000", "content": "I don't have a good answer to your question; the goal of the competition is to find what methods work best and are the most robust. \"crop and \"sew\"\" didn't work for me in this particular competition but worked in a previous competition. Try something, if it fails, try something different.
", "votes": 0, "replies": [] }, { "id": 227045, "author_name": "YL", "author_url": "", "post_date": "2017-10-03T15:51:29.197000", "content": "that's right of course. i am asking because upsampling approach seems to be the standard approach and \"crop and sew\" was not even mentioned.
\n\n\"Intuitively\" i imagine \"crop and sew\" may yield more accurate results but apparently my intuition is very wrong and i am just trying to figure out .
", "votes": 0, "replies": [] } ] }, { "id": 225263, "author_name": "Matt Boyd", "author_url": "", "post_date": "2017-09-28T16:03:29.557000", "content": "Rank 70 - LB (0.9966), 1x1080Ti, 5 model ensemble (unet and densenet)
", "votes": 1, "replies": [] }, { "id": 225134, "author_name": "Arshjot Khehra", "author_url": "", "post_date": "2017-09-28T10:34:40.710000", "content": "Rank - 42. 1 x GTX 1060.\nSingle model:\nPublic LB - 0.9968, Private LB - 0.9967
\n\nMade a submission using an ensemble of three models but missed the deadline :(\nEnsemble: \nPublic LB - 0.9969, Private LB - 0.9968
", "votes": 1, "replies": [] }, { "id": 225107, "author_name": "Paul Kinpzz", "author_url": "", "post_date": "2017-09-28T09:16:57.090000", "content": "Public LB: Rank 64/0.9969 Private LB: Rank 33/0.9968. Unet with multiscale and one CVPR2017 model. 2 x gtx 1080
", "votes": 1, "replies": [] }, { "id": 225006, "author_name": "Alan Khoa Nguyen", "author_url": "", "post_date": "2017-09-28T03:32:55.387000", "content": "28 - single model, 1 gtx1080 + 2 gtx 1060s
", "votes": 1, "replies": [] }, { "id": 224999, "author_name": "James Requa", "author_url": "", "post_date": "2017-09-28T03:13:19.107000", "content": "Public LB: Rank 87/0.9968; Private LB: Rank 129/0.9964. Ensemble of 5 folds from 3 different u-net models. One machine with 1*GTX 1070.
", "votes": 1, "replies": [] }, { "id": 226179, "author_name": "n01z3", "author_url": "", "post_date": "2017-10-01T09:44:01.960000", "content": "6th:\n4x 1080ti,\n4x 1080ti,\n2x Titan X Maxwell,\n2x Titan X Pascal.\nAnd part time:\n4x 1080,\n4x 1080ti,\n2x Tesla P40
", "votes": 2, "replies": [] }, { "id": 225806, "author_name": "hengck23", "author_url": "", "post_date": "2017-09-30T02:05:25.150000", "content": "It seems that high gpu resources will become a must for some future kaggle compeition. Another example is https://www.kaggle.com/c/cdiscount-image-classification-challenge. Time to invest in gpu. But nvidia is upgrading their gpu very fast. If someone can provide gpu rental service specially designed for kagglers it would be nice
", "votes": 2, "replies": [] }, { "id": 225710, "author_name": "Rafał Jankowski", "author_url": "", "post_date": "2017-09-29T21:24:35.333000", "content": "3xGTX 980 Public (37th): 0.9970 -> Private 0.9966 (76th)
", "votes": 2, "replies": [] }, { "id": 225563, "author_name": "KanaChou", "author_url": "", "post_date": "2017-09-29T14:10:31.013000", "content": "94th 1*1080Ti
", "votes": 2, "replies": [] }, { "id": 225554, "author_name": "Torres", "author_url": "", "post_date": "2017-09-29T13:39:00.390000", "content": "9th, 2 x 1080Ti, 1280 U-net, 4 ensemble
", "votes": 2, "replies": [] }, { "id": 225449, "author_name": "ZhengHang", "author_url": "", "post_date": "2017-09-29T04:38:09.183000", "content": "one 1080ti +one titan xp, final submission is single model,make a big mistake that be lost in overfit. no time to ensemble,its first time to do kaggle i will be better next time☺
", "votes": 2, "replies": [] }, { "id": 225383, "author_name": "David J. Slate", "author_url": "", "post_date": "2017-09-28T21:36:12.830000", "content": "Another interesting graph (although it would be very difficult to get the data) would be rank vs total power consumed for this contest (i.e., in kilowatt hours).
", "votes": 2, "replies": [] }, { "id": 225267, "author_name": "Tim Anderton", "author_url": "", "post_date": "2017-09-28T16:11:49.777000", "content": "Public: rank 446 score 0.9933 Private: rank 428 score 0.9940\n1 GTX 660 Ti (2.7 TFlops with 3Gb of memory)
\n\nA simple average of 2 models, one very scaled down U-Net to make it fit mostly in GPU memory and another architecture based loosely on https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf \nHad things set up to do 5 training folds split by car but could only manage training on the first fold for each model before the competition ended.
", "votes": 2, "replies": [] }, { "id": 225161, "author_name": "", "author_url": "", "post_date": "2017-09-28T11:52:45.017000", "content": "", "votes": 2, "replies": [] }, { "id": 225141, "author_name": "", "author_url": "", "post_date": "2017-09-28T11:04:28.530000", "content": "", "votes": 2, "replies": [] }, { "id": 225118, "author_name": "", "author_url": "", "post_date": "2017-09-28T09:45:03.883000", "content": "", "votes": 2, "replies": [] }, { "id": 225066, "author_name": "", "author_url": "", "post_date": "2017-09-28T06:39:52.787000", "content": "", "votes": 2, "replies": [] }, { "id": 225064, "author_name": "", "author_url": "", "post_date": "2017-09-28T06:17:31.637000", "content": "", "votes": 2, "replies": [] }, { "id": 225041, "author_name": "", "author_url": "", "post_date": "2017-09-28T04:38:34.177000", "content": "", "votes": 2, "replies": [] }, { "id": 224966, "author_name": "", "author_url": "", "post_date": "2017-09-28T01:52:37.067000", "content": "", "votes": 2, "replies": [ { "id": 224972, "author_name": "", "author_url": "", "post_date": "2017-09-28T02:07:56.150000", "content": "", "votes": 1, "replies": [] } ] }, { "id": 224952, "author_name": "", "author_url": "", "post_date": "2017-09-28T00:44:53.667000", "content": "", "votes": 2, "replies": [] }, { "id": 227276, "author_name": "", "author_url": "", "post_date": "2017-10-04T04:49:37.077000", "content": "", "votes": 0, "replies": [] }, { "id": 226169, "author_name": "", "author_url": "", "post_date": "2017-10-01T08:54:54.093000", "content": "", "votes": 0, "replies": [] }, { "id": 226080, "author_name": "", "author_url": "", "post_date": "2017-10-01T00:24:42.880000", "content": "", "votes": 0, "replies": [] }, { "id": 225464, "author_name": "", "author_url": "", "post_date": "2017-09-29T05:42:35.780000", "content": "", "votes": 0, "replies": [] }, { "id": 226283, "author_name": "", "author_url": "", "post_date": "2017-10-01T17:42:05.197000", "content": "", "votes": 3, "replies": [ { "id": 226292, "author_name": "", "author_url": "", "post_date": "2017-10-01T18:05:38.443000", "content": "", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "224949": "i want to make a graph of ranking vs GPU. Can kagglers provided details? I start with mine as an example:\n\nrank: 13-LB (0.9970), 14 models submitted. 4 machines: \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\"", "225160": "After training nearly around the clock, I started getting these daily warnings from my power company.\n\"Appliances may be working harder than expected.\" Ummm.... yes.\n\n\n\n [1]: https://kaggle2.blob.core.windows.net/forum-message-attachments/225160/7422/warning%20letter.jpg", "225088": "A quick sort out:\n>Rank 1st, Roughly 20 GPUs total. ~200 TFlops \nRank 3 , 2x1080Ti. \nRank 4, 5x1080Ti, 2x1080, 1x1070 \nRank 5, 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070. Total = 23.5TFLOPS . \nRank 7, 6 x Titan X Maxwell 12GB \nRank 10, 2 machines: 1- 4x 1080Ti, 1- 2x 1080 Ti. \nRank 11, 1 x 1080 Ti \nRank 12, DGX-Station(4 x Tesla P100). \nrank: 13, \"2x1080Ti \" + \"1 pascal titanX\" + \"4x maxwell titanX\" + \" 4x maxwell titanX\" \nRank 20, 1 x GTX 1060 \nRank 25, 1 machine: GTX 1060 \nRank 28, 1 gtx1080 + 2 gtx 1060s \nRank 39, 1 machine with 2*GTX 1080. \nRank 70, 1 x 980 Ti. ", "225065": "Someone should ask nvidia to start a kaggle compeition with gpu as prizes", "224951": "Rank 11, LB 0.9971. Ensemble of 7 folds from one model. Single 1080 Ti", "225146": "Rank 20, LB (0.9970). Ensemble of 4 models. 1 x GTX 1060", "225079": "3rd place (0.9972), mainly used 2x1080Ti.", "224976": "1st Place. Roughly 20 GPUs total in our team. ~200 TFlops", "225888": "I have achieved 9969 with no gpu but on 16-core cpu", "225225": "Rank 87: (LB 0.9966) Single 1280 U-net, old GTX 970.", "225121": "4th, 0.9972. As our team name says, we did it using roughly 80 TFlops (5x1080Ti, 2x1080, 1x1070). At some point at the end of the competition, we were even considering on using a 60x1070 mining rig for TTA :) But it turned out being not suitable for Deep Learning, so that was transformed into a local meme and didn't go further :) ", "225175": "Rank: 126 (LB 0.9965) Single 1024 Unet with some test augmentation.\n\nI don't have GPU so I rented ec2 p2 with K80. I found that the K80 was quite slow compared to other reported training times. This was my first deep learning competition so I had quite a few challenges. It was surprising that the lack of GPU was among my biggest bottlenecks. In other competitions I got used to be able to rent quite competitive machines from aws (Memory >= 100GB vCPU>=16 ~ 2 USD/day).", "225012": "5th - Private 0.9972. 7 models ensemble. 2.5 GPUs/machines = 1x GTX1080Ti + 1x GTX1080 + 0.5x GTX1070 (used this for parallel TTA inference only since it was a slow PC). Total = 11.3 + 9 + 6.5/2 = 23.5TFLOPS . Surely there is a strong positive correlation between GPU power and rank in this competition (with some exceptions) ! And I regret not having teamed up :)\n\nI'll write up soon since I have plenty of time to burn these days.", "225010": "WTF the hardware resources here seem astronomical. Do you guys own all these hardware or was it just rented hardware?", "224958": "Rank 10, LB 0.9971. Ensemble of 3 different Unet models: 1. predict background stage from inverted mask then re-invert result , 2. predict car, 3. split image top/bottom. 2 machines: 1- 4x 1080Ti (had all 4 watercooled during competition to prevent thermal throttling), 1 -2x 1080 Ti.", "227102": "rank: 62-Private (0.9967), 3 models ensembled. 1 machine w/ 1x1080\n\nBTW, this competition is less powerplay than I thought... (So many 1060s!)", "225337": "78th TitanX, simple unet without using cv2.resize fonction at training or testing.", "225263": "Rank 70 - LB (0.9966), 1x1080Ti, 5 model ensemble (unet and densenet)", "225134": "Rank - 42. 1 x GTX 1060.\nSingle model:\nPublic LB - 0.9968, Private LB - 0.9967\n\nMade a submission using an ensemble of three models but missed the deadline :(\nEnsemble: \nPublic LB - 0.9969, Private LB - 0.9968", "225107": "Public LB: Rank 64/0.9969 Private LB: Rank 33/0.9968. Unet with multiscale and one CVPR2017 model. 2 x gtx 1080", "225006": "28 - single model, 1 gtx1080 + 2 gtx 1060s", "224999": "Public LB: Rank 87/0.9968; Private LB: Rank 129/0.9964. Ensemble of 5 folds from 3 different u-net models. One machine with 1*GTX 1070.", "226179": "6th:\n4x 1080ti,\n4x 1080ti,\n2x Titan X Maxwell,\n2x Titan X Pascal.\nAnd part time:\n4x 1080,\n4x 1080ti,\n2x Tesla P40", "225806": "It seems that high gpu resources will become a must for some future kaggle compeition. Another example is https://www.kaggle.com/c/cdiscount-image-classification-challenge. Time to invest in gpu. But nvidia is upgrading their gpu very fast. If someone can provide gpu rental service specially designed for kagglers it would be nice", "225710": "3xGTX 980 Public (37th): 0.9970 -> Private 0.9966 (76th)", "225563": "94th 1*1080Ti", "225554": "9th, 2 x 1080Ti, 1280 U-net, 4 ensemble", "225449": "one 1080ti +one titan xp, final submission is single model,make a big mistake that be lost in overfit. no time to ensemble,its first time to do kaggle i will be better next time☺", "225383": "Another interesting graph (although it would be very difficult to get the data) would be rank vs total power consumed for this contest (i.e., in kilowatt hours).", "225267": "Public: rank 446 score 0.9933 Private: rank 428 score 0.9940\n1 GTX 660 Ti (2.7 TFlops with 3Gb of memory)\n\nA simple average of 2 models, one very scaled down U-Net to make it fit mostly in GPU memory and another architecture based loosely on https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf \nHad things set up to do 5 training folds split by car but could only manage training on the first fold for each model before the competition ended. ", "225161": "7th private lb score: 6 x Titan X Maxwell 12GB", "225141": "39th on private and 82nd on public (both cases 0.9968) - using 1 GTX 1070 and a single U-net, but tailored the loss s.t. boundary errors are weighted more heavily :).", "225118": "rank: 38 (0.9968), 1 gtx 1080, public - 96 (0.9968). Ensemble of 7 Unet models, used bounding boxes with cars instead of whole images for learning.", "225066": "Public LB: Rank 70/0.9969; Private LB: Rank 71/0.9966. Single U-Net, 1 fold, 1 x 980 Ti.", "225064": "Rank 12th (0.9970) on Private LB. DGX-Station(4 x Tesla P100). Ensemble of 6 different models.", "225041": "149 position. 2 GPU 1080", "224966": "rank 25-LB (0.9969). Ensemble of 5 folds from one model. 1 machine: GTX 1060", "224952": "Public LB: Rank 39/0.9970; Private LB: Rank 29/0.9969. Ensembles of 8 models from 6 different models. One machine with 2*GTX 1080.", "227276": "I think you need to go about this in a much more stuctured way", "226169": "nice job", "226080": "good", "225464": "Thanks @Heng for all your contributions.\n\nRank 166 with single model. \n\nI used floydhub GPU with Total memory: 11.17GiB\nMy best model is a unet 1280x1280, batch size 8 that converged after 29 epochs with\nloss: 0.0084 - dice_loss: 0.9957 - val_loss: 0.0108 - val_dice_loss: 0.9953\n\nMy smallest model is a 512x512, batch size 12. Trained on Floyd but prediction was done on my CPU locally that took 8.5 days to complete. \nMy CPU device spec. :-\nIntel(R) Core(TM) i7-7500U CPU @ 2.70GHz, 2901 Mhz, 2 Core(s), 4 Logical Processor(s)\nInstalled Physical Memory (RAM)\t12.0 GB\n\n", "226283": "" } }