Thanks to @shonenkov and his public kernel
\nI have set up the size of batches for my hardware and have used B7
Thanks to @shonenkov and his public kernel
\nI have set up the size of batches for my hardware and have used B7
If you don't mind me asking, can you tell me the difference between that kernel and the learning you have done locally?\nOnly EfficientNetB2→EfficientNetB7?
", "rawMarkdown": "If you don't mind me asking, can you tell me the difference between that kernel and the learning you have done locally?\nOnly EfficientNetB2→EfficientNetB7?", "votes": null }, { "id": "937374", "postDate": "07/21/2020 01:17:52", "content": "Hi @golubev,
\nCongratulations for your success..
\nCan you share with me the details of hardware (GPU/TPU, Colab/Kaggle Kernel) you used for training and inference. Also, did you use efficient net from pytorch or Keras. What was the batch size you used and the time taken to train.
All the best for future competitions.
", "rawMarkdown": "Hi @golubev, \n\nCongratulations for your success..\nCan you share with me the details of hardware (GPU/TPU, Colab/Kaggle Kernel) you used for training and inference. Also, did you use efficient net from pytorch or Keras. What was the batch size you used and the time taken to train.\n\nAll the best for future competitions.", "votes": null }, { "id": "937454", "postDate": "07/21/2020 03:17:34", "content": "Congrats on result @golubev! I really gave up on this competition because I didn't have enough hardware :((
", "rawMarkdown": "Congrats on result @golubev! I really gave up on this competition because I didn't have enough hardware :((", "votes": null }, { "id": "937506", "postDate": "07/21/2020 04:05:20", "content": "I can totally relate. I am a student from India trying to learn and grow., but I hVe to give up on competitions due to lack of hardware. I only have kaggle and colab at my disposal
", "rawMarkdown": "I can totally relate. I am a student from India trying to learn and grow., but I hVe to give up on competitions due to lack of hardware. I only have kaggle and colab at my disposal", "votes": null }, { "id": "937533", "postDate": "07/21/2020 04:27:33", "content": "yeah @tanulsingh077 this is true for me as well the key of getting a good rank was to train deeper models with enough time. same in my case too
", "rawMarkdown": "yeah @tanulsingh077 this is true for me as well the key of getting a good rank was to train deeper models with enough time. same in my case too", "votes": null }, { "id": "937692", "postDate": "07/21/2020 06:04:50", "content": "Congrats @golubev
\nB7 on this dataset might have taken a lot of time. I am really surprised to know that just a change in model made a huge difference.
", "rawMarkdown": "Congrats @golubev \n\nB7 on this dataset might have taken a lot of time. I am really surprised to know that just a change in model made a huge difference.", "votes": null }, { "id": "939308", "postDate": "07/22/2020 06:51:26", "content": "May I know the specs of this monster hardware of yours :-) B7 is no joke to train (I'm curious of the batch size you manage to squeeze)
", "rawMarkdown": "May I know the specs of this monster hardware of yours :-) B7 is no joke to train (I'm curious of the batch size you manage to squeeze)", "votes": null }, { "id": "939389", "postDate": "07/22/2020 07:57:10", "content": "hardware: 4x V100
\nbatch size: as maximum as possible
\n3.5days for 1 fold => I had got time only for 3 folds
best single models:\n| Type(xGPU) | CV | Public | Private |\n| --- | --- | --- | --- |\n| B0(x2) | 917 | 921 | 905 |\n| B2(Public kernel) | 912 | 921 | 904 |\n| B2(x4) | 918 | 927 | 909 |\n| B5(x1) | 922 | 929 | 911 |\n| B6(x2) | 925 | 927 | 915 |\n| B7(x4) | 927 | 941 | 921 |
", "rawMarkdown": "best single models:\n| Type(xGPU) | CV | Public | Private |\n| --- | --- | --- | --- |\n| B0(x2) | 917 | 921 | 905 |\n| B2(Public kernel) | 912 | 921 | 904 |\n| B2(x4) | 918 | 927 | 909 |\n| B5(x1) | 922 | 929 | 911 |\n| B6(x2) | 925 | 927 | 915 |\n| B7(x4) | 927 | 941 | 921 |", "votes": null }, { "id": "941859", "postDate": "07/23/2020 13:06:07", "content": "Thanks!
", "rawMarkdown": "Thanks!", "votes": null } ], "comments": [ { "id": 937374, "author_name": "shwetank3", "author_url": "", "post_date": "07/21/2020 01:17:52", "content": "Hi @golubev,
\nCongratulations for your success..
\nCan you share with me the details of hardware (GPU/TPU, Colab/Kaggle Kernel) you used for training and inference. Also, did you use efficient net from pytorch or Keras. What was the batch size you used and the time taken to train.
All the best for future competitions.
", "votes": null, "replies": [] }, { "id": 937454, "author_name": "duykhanh99", "author_url": "", "post_date": "07/21/2020 03:17:34", "content": "Congrats on result @golubev! I really gave up on this competition because I didn't have enough hardware :((
", "votes": null, "replies": [ { "id": 937506, "author_name": "tanulsingh077", "author_url": "", "post_date": "07/21/2020 04:05:20", "content": "I can totally relate. I am a student from India trying to learn and grow., but I hVe to give up on competitions due to lack of hardware. I only have kaggle and colab at my disposal
", "votes": null, "replies": [] }, { "id": 937533, "author_name": "pranshu29", "author_url": "", "post_date": "07/21/2020 04:27:33", "content": "yeah @tanulsingh077 this is true for me as well the key of getting a good rank was to train deeper models with enough time. same in my case too
", "votes": null, "replies": [] } ] }, { "id": 937692, "author_name": "vishnurapps", "author_url": "", "post_date": "07/21/2020 06:04:50", "content": "Congrats @golubev
\nB7 on this dataset might have taken a lot of time. I am really surprised to know that just a change in model made a huge difference.
", "votes": null, "replies": [] }, { "id": 939308, "author_name": "nyleve", "author_url": "", "post_date": "07/22/2020 06:51:26", "content": "May I know the specs of this monster hardware of yours :-) B7 is no joke to train (I'm curious of the batch size you manage to squeeze)
", "votes": null, "replies": [] }, { "id": 939389, "author_name": "golubev", "author_url": "", "post_date": "07/22/2020 07:57:10", "content": "hardware: 4x V100
\nbatch size: as maximum as possible
\n3.5days for 1 fold => I had got time only for 3 folds
If you don't mind me asking, can you tell me the difference between that kernel and the learning you have done locally?\nOnly EfficientNetB2→EfficientNetB7?
", "votes": null, "replies": [ { "id": 939556, "author_name": "golubev", "author_url": "", "post_date": "07/22/2020 09:55:55", "content": "best single models:\n| Type(xGPU) | CV | Public | Private |\n| --- | --- | --- | --- |\n| B0(x2) | 917 | 921 | 905 |\n| B2(Public kernel) | 912 | 921 | 904 |\n| B2(x4) | 918 | 927 | 909 |\n| B5(x1) | 922 | 929 | 911 |\n| B6(x2) | 925 | 927 | 915 |\n| B7(x4) | 927 | 941 | 921 |
", "votes": null, "replies": [] }, { "id": 941859, "author_name": "snakayama", "author_url": "", "post_date": "07/23/2020 13:06:07", "content": "Thanks!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "937347": "Thanks to @shonenkov and his [public kernel](https://www.kaggle.com/shonenkov/train-inference-gpu-baseline)\nI have set up the size of batches for my hardware and have used B7", "937371": "If you don't mind me asking, can you tell me the difference between that kernel and the learning you have done locally?\nOnly EfficientNetB2→EfficientNetB7?", "937374": "Hi @golubev, \n\nCongratulations for your success..\nCan you share with me the details of hardware (GPU/TPU, Colab/Kaggle Kernel) you used for training and inference. Also, did you use efficient net from pytorch or Keras. What was the batch size you used and the time taken to train.\n\nAll the best for future competitions.", "937454": "Congrats on result @golubev! I really gave up on this competition because I didn't have enough hardware :((", "937506": "I can totally relate. I am a student from India trying to learn and grow., but I hVe to give up on competitions due to lack of hardware. I only have kaggle and colab at my disposal", "937533": "yeah @tanulsingh077 this is true for me as well the key of getting a good rank was to train deeper models with enough time. same in my case too", "937692": "Congrats @golubev \n\nB7 on this dataset might have taken a lot of time. I am really surprised to know that just a change in model made a huge difference.", "939308": "May I know the specs of this monster hardware of yours :-) B7 is no joke to train (I'm curious of the batch size you manage to squeeze)", "939389": "hardware: 4x V100\nbatch size: as maximum as possible\n3.5days for 1 fold => I had got time only for 3 folds", "939556": "best single models:\n| Type(xGPU) | CV | Public | Private |\n| --- | --- | --- | --- |\n| B0(x2) | 917 | 921 | 905 |\n| B2(Public kernel) | 912 | 921 | 904 |\n| B2(x4) | 918 | 927 | 909 |\n| B5(x1) | 922 | 929 | 911 |\n| B6(x2) | 925 | 927 | 915 |\n| B7(x4) | 927 | 941 | 921 |", "941859": "Thanks!" }, "source": "meta" }