{ "id": 192926, "title": "Some suggestions for last 42 hours GPU", "url": "/competitions/rsna-str-pulmonary-embolism-detection/discussion/192926", "author_name": "Dewei Chen", "post_date": "2020-10-24T08:02:34.988000", "votes": 5, "comment_count": 3, "views": 0, "content": "

Firstly, just a little suggestion of the last 42 hours of usage for all team:

\n
    \n
  1. Save at least 10 hours for final submition. With that surprising kernel publish, we all need 2hours of GPU for one inference. We need to consider private data; hence we cannot copy and submit a result.
    2. \n
  2. If you find a way to reduce the inference time, I still suggest you keep 4 hours balance, you may need to stacking your model; Inference one signal model solution takes me 14 mins GPU time, so if you only doing averaging you may still need 2 hours for a single submission.
    3. \n
  3. For another 20-30 hours, finish your last model training.
    4. \n
\n

Secondly, There is two suggestion for Google and Kaggle:

\n
    \n

  1. For the competition with huge image data, we had two choices, download data to local, or using a notebook in the cloud for training purposes. For me, I spent 6 full days to download data to my local PC by using Kaggle API instruction; and 40 hours for a single model training.
    \nSo first suggestion is: Can Google increase the weekly time limit of GPU in this type of game? You can see there are 50% teams less than last year. I think it is the main reason.

    2. \n

  2. Just make a little change of notebook UI: Add a button beside Save & Run all with an accelerator selection button, we can only turn GPU when we run our code, it will be saved a lot of time.

    3. \n
\n

@juliaelliott

", "messages": [ { "id": 1058761, "postDate": "2020-10-24T08:02:34.987Z", "content": "

Firstly, just a little suggestion of the last 42 hours of usage for all team:

\n
    \n
  1. Save at least 10 hours for final submition. With that surprising kernel publish, we all need 2hours of GPU for one inference. We need to consider private data; hence we cannot copy and submit a result.
    2. \n
  2. If you find a way to reduce the inference time, I still suggest you keep 4 hours balance, you may need to stacking your model; Inference one signal model solution takes me 14 mins GPU time, so if you only doing averaging you may still need 2 hours for a single submission.
    3. \n
  3. For another 20-30 hours, finish your last model training.
    4. \n
\n

Secondly, There is two suggestion for Google and Kaggle:

\n
    \n

  1. For the competition with huge image data, we had two choices, download data to local, or using a notebook in the cloud for training purposes. For me, I spent 6 full days to download data to my local PC by using Kaggle API instruction; and 40 hours for a single model training.
    \nSo first suggestion is: Can Google increase the weekly time limit of GPU in this type of game? You can see there are 50% teams less than last year. I think it is the main reason.

    2. \n

  2. Just make a little change of notebook UI: Add a button beside Save & Run all with an accelerator selection button, we can only turn GPU when we run our code, it will be saved a lot of time.

    3. \n
\n

@juliaelliott

", "rawMarkdown": "Firstly, just a little suggestion of the last 42 hours of usage for all team: \n1. Save at least **10 hours for final submition**. With that surprising kernel publish, we all need 2hours of GPU for one inference. We need to consider private data; hence we cannot copy and submit a result. \n2. If you find a way to reduce the inference time, I still suggest you keep **4 hours** balance, you may need to **stacking** your model; Inference one signal model solution takes me **14 mins** GPU time, so if you only doing **averaging** you may still need **2 hours** for a single submission.\n3. For another 20-30 hours, finish your last model training. \n\n\nSecondly, There is two suggestion for Google and Kaggle: \n1. For the competition with huge image data, we had two choices, download data to local, or using a notebook in the cloud for training purposes. For me, I spent 6 full days to download data to my local PC by using Kaggle API instruction; and 40 hours for a single model training.\nSo first suggestion is: Can Google **increase the weekly time limit of GPU** in this type of game? You can see there are **50% teams less than last year**. I think it is the main reason.\n\n2. Just make a little change of notebook UI: **Add a button beside Save & Run** all with an **accelerator selection button**, we can only turn GPU when we run our code, it will be saved a lot of time.\n\n\n\n@juliaelliott ", "votes": 5 }, { "id": 1059386, "postDate": "2020-10-25T03:11:38.577Z", "content": "

Just fyi, you can commit a kernel within 1 minute and submit it with no issue. For the \"amazing kernel\" you talked about, you can still commit it in 1 minutes.

", "rawMarkdown": "Just fyi, you can commit a kernel within 1 minute and submit it with no issue. For the \"amazing kernel\" you talked about, you can still commit it in 1 minutes.", "votes": 1, "replies": [ { "id": 1059398, "postDate": "2020-10-25T03:36:17.377Z", "content": "

Agree with you, maybe using .head(2*batch_size) to make sure GPU RAM is enough when run with full data.

", "rawMarkdown": "Agree with you, maybe using `.head(2*batch_size)` to make sure GPU RAM is enough when run with full data.", "votes": 1 } ] }, { "id": 1059536, "postDate": "2020-10-25T08:05:33.927Z", "content": "

one addition, read this topic carefully: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/186982

\n

In this competition, There are two test dataset: Kaggle will using the real testset after you submit your code to leaderboad. Anothor one is what we saw, it will not be ran at leaderboard.
\nSo you don't need to run all test data when run an inference nootbook.

", "rawMarkdown": "one addition, read this topic carefully: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/186982\n\nIn this competition, There are two test dataset: Kaggle will using the real testset after you submit your code to leaderboad. Anothor one is what we saw, it will not be ran at leaderboard.\nSo you don't need to run all test data when run an inference nootbook." } ], "comments": [ { "id": 1059386, "author_name": "sin", "author_url": "", "post_date": "2020-10-25T03:11:38.577000", "content": "

Just fyi, you can commit a kernel within 1 minute and submit it with no issue. For the \"amazing kernel\" you talked about, you can still commit it in 1 minutes.

", "votes": 1, "replies": [ { "id": 1059398, "author_name": "Dewei Chen", "author_url": "", "post_date": "2020-10-25T03:36:17.377000", "content": "

Agree with you, maybe using .head(2*batch_size) to make sure GPU RAM is enough when run with full data.

", "votes": 1, "replies": [] } ] }, { "id": 1059536, "author_name": "Dewei Chen", "author_url": "", "post_date": "2020-10-25T08:05:33.927000", "content": "

one addition, read this topic carefully: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/186982

\n

In this competition, There are two test dataset: Kaggle will using the real testset after you submit your code to leaderboad. Anothor one is what we saw, it will not be ran at leaderboard.
\nSo you don't need to run all test data when run an inference nootbook.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1058761": "Firstly, just a little suggestion of the last 42 hours of usage for all team: \n1. Save at least **10 hours for final submition**. With that surprising kernel publish, we all need 2hours of GPU for one inference. We need to consider private data; hence we cannot copy and submit a result. \n2. If you find a way to reduce the inference time, I still suggest you keep **4 hours** balance, you may need to **stacking** your model; Inference one signal model solution takes me **14 mins** GPU time, so if you only doing **averaging** you may still need **2 hours** for a single submission.\n3. For another 20-30 hours, finish your last model training. \n\n\nSecondly, There is two suggestion for Google and Kaggle: \n1. For the competition with huge image data, we had two choices, download data to local, or using a notebook in the cloud for training purposes. For me, I spent 6 full days to download data to my local PC by using Kaggle API instruction; and 40 hours for a single model training.\nSo first suggestion is: Can Google **increase the weekly time limit of GPU** in this type of game? You can see there are **50% teams less than last year**. I think it is the main reason.\n\n2. Just make a little change of notebook UI: **Add a button beside Save & Run** all with an **accelerator selection button**, we can only turn GPU when we run our code, it will be saved a lot of time.\n\n\n\n@juliaelliott ", "1059386": "Just fyi, you can commit a kernel within 1 minute and submit it with no issue. For the \"amazing kernel\" you talked about, you can still commit it in 1 minutes.", "1059536": "one addition, read this topic carefully: https://www.kaggle.com/c/rsna-str-pulmonary-embolism-detection/discussion/186982\n\nIn this competition, There are two test dataset: Kaggle will using the real testset after you submit your code to leaderboad. Anothor one is what we saw, it will not be ran at leaderboard.\nSo you don't need to run all test data when run an inference nootbook." } }