![\"\"](\"https://storage.googleapis.com/kaggle-forum-message-attachments/547783/13438/Screen%20Shot%202019-06-08%20at%204.57.14%20PM.png\")
You can test whether your kernel can be run successfully.
\n\nAs for one of our submissions’ result, the stage 2 running time is roughly 5.2 times more than stage 1.\nIn addition, the size of new submission.csv is also 5.2 times larger, thus I guess the distribution of testset is similar as the old one.
\n\n
You can test whether your kernel can be run successfully.
\n\nAs for one of our submissions’ result, the stage 2 running time is roughly 5.2 times more than stage 1.\nIn addition, the size of new submission.csv is also 5.2 times larger, thus I guess the distribution of testset is similar as the old one.
\n\n
Thanks for letting know.\nOne of my kernel successfully completes its run in 28163.9 seconds.
", "rawMarkdown": "Thanks for letting know.\nOne of my kernel successfully completes its run in 28163.9 seconds.", "votes": null }, { "id": "548146", "postDate": "06/08/2019 20:59:23", "content": "Not enough TTAs for sure :)
", "rawMarkdown": "Not enough TTAs for sure :)", "votes": null }, { "id": "548298", "postDate": "06/09/2019 06:19:28", "content": "In order to save the memory & time, our team uses \"tta by fold\", like for fold 0, infer with the original images, fold 1 with flipped images, ...
", "rawMarkdown": "In order to save the memory & time, our team uses \"tta by fold\", like for fold 0, infer with the original images, fold 1 with flipped images, ...", "votes": null }, { "id": "548429", "postDate": "06/09/2019 11:26:06", "content": "Nice idea. Did you also modify random crop too (if you used) so that it efficiently crops an image with smaller IOU among each fold? How much was the performace degraded compared to tta 5times or 10 times?
", "rawMarkdown": "Nice idea. Did you also modify random crop too (if you used) so that it efficiently crops an image with smaller IOU among each fold? How much was the performace degraded compared to tta 5times or 10 times?", "votes": null }, { "id": "548442", "postDate": "06/09/2019 11:43:37", "content": "We just use traditional 5 crops (4 corners and 1 center).
\n\nFor one single model, tta by fold is 0.001 lower than tta for each fold.
", "rawMarkdown": "We just use traditional 5 crops (4 corners and 1 center).\n\nFor one single model, tta by fold is 0.001 lower than tta for each fold.", "votes": null }, { "id": "548468", "postDate": "06/09/2019 12:45:06", "content": "yes, we use 10 fold thus \"10 tta by fold\": 2 flips * 5 crops (4 corners and 1 center). For our 10 fold seresnext101, non tta gets 653, 10 tta by fold gets 655, 10 tta for each fold gets 656.
", "rawMarkdown": "yes, we use 10 fold thus \"10 tta by fold\": 2 flips * 5 crops (4 corners and 1 center). For our 10 fold seresnext101, non tta gets 653, 10 tta by fold gets 655, 10 tta for each fold gets 656.", "votes": null }, { "id": "548469", "postDate": "06/09/2019 12:45:08", "content": "yes, we use 2 flips * 5 crops (4 corners and 1 center) for our 10 fold seresnext101\n- non tta gets 653\n- 10 tta by fold gets 655\n- 10 tta for each fold gets 656
", "rawMarkdown": "yes, we use 2 flips * 5 crops (4 corners and 1 center) for our 10 fold seresnext101\n- non tta gets 653\n- 10 tta by fold gets 655\n- 10 tta for each fold gets 656", "votes": null }, { "id": "548478", "postDate": "06/09/2019 13:01:15", "content": "Thank you for the detailed answer. Very informative.\nI've never come up with the idea of TTA by fold and nice to know. The performace degration seems relatively small considering it reduces inference cycles from 10 to 1 (in case of 10 tta).
", "rawMarkdown": "Thank you for the detailed answer. Very informative.\nI've never come up with the idea of TTA by fold and nice to know. The performace degration seems relatively small considering it reduces inference cycles from 10 to 1 (in case of 10 tta).", "votes": null }, { "id": "548518", "postDate": "06/09/2019 14:11:30", "content": "my network has some issues, thus the \"unsuccessful\" messages were sent successfully
", "rawMarkdown": "my network has some issues, thus the \"unsuccessful\" messages were sent successfully", "votes": null }, { "id": "548847", "postDate": "06/10/2019 02:31:55", "content": "@yiheng very nice idea!
", "rawMarkdown": "yiheng very nice idea!", "votes": null }, { "id": "548920", "postDate": "06/10/2019 05:27:35", "content": "what do you mean by \"10 tta by fold gets 655, 10 tta for each fold gets 656.\" could you explain the difference between these two?
", "rawMarkdown": "what do you mean by \"10 tta by fold gets 655, 10 tta for each fold gets 656.\" could you explain the difference between these two?", "votes": null }, { "id": "548960", "postDate": "06/10/2019 06:28:39", "content": "1) fold 0: tta0, fold 1: tta1,...\n2) fold 0: tta0, tta1, ..., tta9 fold 1: tta0, tta1, ..., tta9
", "rawMarkdown": "1) fold 0: tta0, fold 1: tta1,...\n2) fold 0: tta0, tta1, ..., tta9 fold 1: tta0, tta1, ..., tta9", "votes": null }, { "id": "548973", "postDate": "06/10/2019 06:43:19", "content": "I see I didn't know that people do first one. I only knew second one :)
", "rawMarkdown": "I see I didn't know that people do first one. I only knew second one :)", "votes": null }, { "id": "548974", "postDate": "06/10/2019 06:47:23", "content": "Actually, if we have enough inference time, we'll not choose 1)
", "rawMarkdown": "Actually, if we have enough inference time, we'll not choose 1)", "votes": null } ], "comments": [ { "id": 548039, "author_name": "appian", "author_url": "", "post_date": "06/08/2019 16:57:34", "content": "Thanks for letting know.\nOne of my kernel successfully completes its run in 28163.9 seconds.
", "votes": null, "replies": [ { "id": 548146, "author_name": "artyomp", "author_url": "", "post_date": "06/08/2019 20:59:23", "content": "Not enough TTAs for sure :)
", "votes": null, "replies": [] }, { "id": 548298, "author_name": "yiheng", "author_url": "", "post_date": "06/09/2019 06:19:28", "content": "In order to save the memory & time, our team uses \"tta by fold\", like for fold 0, infer with the original images, fold 1 with flipped images, ...
", "votes": null, "replies": [] }, { "id": 548429, "author_name": "appian", "author_url": "", "post_date": "06/09/2019 11:26:06", "content": "Nice idea. Did you also modify random crop too (if you used) so that it efficiently crops an image with smaller IOU among each fold? How much was the performace degraded compared to tta 5times or 10 times?
", "votes": null, "replies": [] }, { "id": 548442, "author_name": "yiheng", "author_url": "", "post_date": "06/09/2019 11:43:37", "content": "We just use traditional 5 crops (4 corners and 1 center).
\n\nFor one single model, tta by fold is 0.001 lower than tta for each fold.
", "votes": null, "replies": [] }, { "id": 548468, "author_name": "yiheng", "author_url": "", "post_date": "06/09/2019 12:45:06", "content": "yes, we use 10 fold thus \"10 tta by fold\": 2 flips * 5 crops (4 corners and 1 center). For our 10 fold seresnext101, non tta gets 653, 10 tta by fold gets 655, 10 tta for each fold gets 656.
", "votes": null, "replies": [] }, { "id": 548469, "author_name": "yiheng", "author_url": "", "post_date": "06/09/2019 12:45:08", "content": "yes, we use 2 flips * 5 crops (4 corners and 1 center) for our 10 fold seresnext101\n- non tta gets 653\n- 10 tta by fold gets 655\n- 10 tta for each fold gets 656
", "votes": null, "replies": [] }, { "id": 548478, "author_name": "appian", "author_url": "", "post_date": "06/09/2019 13:01:15", "content": "Thank you for the detailed answer. Very informative.\nI've never come up with the idea of TTA by fold and nice to know. The performace degration seems relatively small considering it reduces inference cycles from 10 to 1 (in case of 10 tta).
", "votes": null, "replies": [] }, { "id": 548518, "author_name": "yiheng", "author_url": "", "post_date": "06/09/2019 14:11:30", "content": "my network has some issues, thus the \"unsuccessful\" messages were sent successfully
", "votes": null, "replies": [] } ] }, { "id": 548847, "author_name": "projdev", "author_url": "", "post_date": "06/10/2019 02:31:55", "content": "@yiheng very nice idea!
", "votes": null, "replies": [] }, { "id": 548920, "author_name": "chrisyang7", "author_url": "", "post_date": "06/10/2019 05:27:35", "content": "what do you mean by \"10 tta by fold gets 655, 10 tta for each fold gets 656.\" could you explain the difference between these two?
", "votes": null, "replies": [ { "id": 548960, "author_name": "yiheng", "author_url": "", "post_date": "06/10/2019 06:28:39", "content": "1) fold 0: tta0, fold 1: tta1,...\n2) fold 0: tta0, tta1, ..., tta9 fold 1: tta0, tta1, ..., tta9
", "votes": null, "replies": [] }, { "id": 548973, "author_name": "chrisyang7", "author_url": "", "post_date": "06/10/2019 06:43:19", "content": "I see I didn't know that people do first one. I only knew second one :)
", "votes": null, "replies": [] }, { "id": 548974, "author_name": "yiheng", "author_url": "", "post_date": "06/10/2019 06:47:23", "content": "Actually, if we have enough inference time, we'll not choose 1)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "547783": "You can test whether your kernel can be run successfully. \n\nAs for one of our submissions’ result, the stage 2 running time is roughly 5.2 times more than stage 1.\nIn addition, the size of new submission.csv is also 5.2 times larger, thus I guess the distribution of testset is similar as the old one.\n\n", "548039": "Thanks for letting know.\nOne of my kernel successfully completes its run in 28163.9 seconds.", "548146": "Not enough TTAs for sure :)", "548298": "In order to save the memory & time, our team uses \"tta by fold\", like for fold 0, infer with the original images, fold 1 with flipped images, ...", "548429": "Nice idea. Did you also modify random crop too (if you used) so that it efficiently crops an image with smaller IOU among each fold? How much was the performace degraded compared to tta 5times or 10 times?", "548442": "We just use traditional 5 crops (4 corners and 1 center).\n\nFor one single model, tta by fold is 0.001 lower than tta for each fold.", "548468": "yes, we use 10 fold thus \"10 tta by fold\": 2 flips * 5 crops (4 corners and 1 center). For our 10 fold seresnext101, non tta gets 653, 10 tta by fold gets 655, 10 tta for each fold gets 656.", "548469": "yes, we use 2 flips * 5 crops (4 corners and 1 center) for our 10 fold seresnext101\n- non tta gets 653\n- 10 tta by fold gets 655\n- 10 tta for each fold gets 656", "548478": "Thank you for the detailed answer. Very informative.\nI've never come up with the idea of TTA by fold and nice to know. The performace degration seems relatively small considering it reduces inference cycles from 10 to 1 (in case of 10 tta).", "548518": "my network has some issues, thus the \"unsuccessful\" messages were sent successfully", "548847": "yiheng very nice idea!", "548920": "what do you mean by \"10 tta by fold gets 655, 10 tta for each fold gets 656.\" could you explain the difference between these two?", "548960": "1) fold 0: tta0, fold 1: tta1,...\n2) fold 0: tta0, tta1, ..., tta9 fold 1: tta0, tta1, ..., tta9", "548973": "I see I didn't know that people do first one. I only knew second one :)", "548974": "Actually, if we have enough inference time, we'll not choose 1)" }, "source": "meta" }