\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5003978%2Fbe49144829fb999b091d0502ceec8237%2F2023-05-05%20%201.41.28.png?generation=1683218508312074&alt=media\")
training notebook:
\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-training
inference notebook:
\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-inference
original colab notebook I used:
\nhttps://github.com/hoyso48/Google---Isolated-Sign-Language-Recognition-1st-place-solution
Currently, Kaggle TPU-VM training in the training notebook is strangely slow (2x times slower than Colab). Let me know how to debug this issue if you are familiar with TPU and TensorFlow.
\nreproducible result from the above training notebook is as follows.
\n
single seed model of 17ms latency took around 20minutes and got 2nd place in private
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5003978%2F26e5323e751aa724bdd12daf140387d5%2F2023-05-05%20%201.43.46.png?generation=1683218642562166&alt=media\")
The 4-seed model often fails with submission scoring error. You can try a 3-seed model and may get a higher private score(?).
\nIf you have any questions or issues regarding these codes, please leave a comment below. Thanks!
", "messages": [ { "id": "2245887", "postDate": "05/04/2023 16:49:23", "content": "training notebook:
\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-training
inference notebook:
\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-inference
original colab notebook I used:
\nhttps://github.com/hoyso48/Google---Isolated-Sign-Language-Recognition-1st-place-solution
Currently, Kaggle TPU-VM training in the training notebook is strangely slow (2x times slower than Colab). Let me know how to debug this issue if you are familiar with TPU and TensorFlow.
\nreproducible result from the above training notebook is as follows.
\n
single seed model of 17ms latency took around 20minutes and got 2nd place in private
\n
The 4-seed model often fails with submission scoring error. You can try a 3-seed model and may get a higher private score(?).
\nIf you have any questions or issues regarding these codes, please leave a comment below. Thanks!
", "rawMarkdown": "training notebook:\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-training\n- inference notebook:\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-inference\n\n- original colab notebook I used:\nhttps://github.com/hoyso48/Google---Isolated-Sign-Language-Recognition-1st-place-solution\n\nCurrently, Kaggle TPU-VM training in the training notebook is strangely slow (2x times slower than Colab). Let me know how to debug this issue if you are familiar with TPU and TensorFlow.\n\nreproducible result from the above training notebook is as follows.\n\n\nsingle seed model of 17ms latency took around 20minutes and got 2nd place in private\n\n\n\n\nThe 4-seed model often fails with submission scoring error. You can try a 3-seed model and may get a higher private score(?).\n\nIf you have any questions or issues regarding these codes, please leave a comment below. Thanks!", "votes": null }, { "id": "2246088", "postDate": "05/04/2023 21:35:20", "content": "Thank you for publishing your code. It saves much time comparing to reproducing myself.
", "rawMarkdown": "Thank you for publishing your code. It saves much time comparing to reproducing myself.", "votes": null }, { "id": "2246626", "postDate": "05/05/2023 10:23:36", "content": "I have submitted your single model + our single model with 0.5/0.5 wieghts ensemble and got this score:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4212496%2F0287ac186cf09b4f961b3e5302f72894%2Ftg_image_449267593.jpeg?generation=1683282196716833&alt=media\")
0.82 public beaten ๐
", "rawMarkdown": "I have submitted your single model + our single model with 0.5/0.5 wieghts ensemble and got this score:\n\n\n\n0.82 public beaten ๐", "votes": null }, { "id": "2246756", "postDate": "05/05/2023 12:42:02", "content": "Wow. I didn't expect that much score improvement. Thank you for sharing! ๐
", "rawMarkdown": "Wow. I didn't expect that much score improvement. Thank you for sharing! ๐", "votes": null }, { "id": "2246883", "postDate": "05/05/2023 14:34:04", "content": "Thanks for posting the code!! I find it very useful for those who are just learning how to create models.
", "rawMarkdown": "Thanks for posting the code!! I find it very useful for those who are just learning how to create models.", "votes": null }, { "id": "2257044", "postDate": "05/13/2023 00:45:01", "content": "Thank a lot for sharing! I think, even competition host did not expect such result.
", "rawMarkdown": "Thank a lot for sharing! I think, even competition host did not expect such result.", "votes": null }, { "id": "2258019", "postDate": "05/13/2023 19:59:08", "content": "thnx that you published this post
", "rawMarkdown": "thnx that you published this post", "votes": null }, { "id": "2266464", "postDate": "05/20/2023 05:22:07", "content": "@hoyso48
\nThanks for sharing the code, that's really helpful!
\nHave you ever tried to run this code with GPU?
\nI've tried on CPU, GPU and TPU, but somehow only on GPU I got following error..
\nHave you experiend something like this?
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1554318%2F4d4259b1a846c774752737b583f3fbb5%2Fscreenshort.png?generation=1684560107872039&alt=media\")
Hi, @bamps23
\nThank you for sharing the issue! I just tested it with GPU(T4) in colab env, and It works fine without issue.
If you are using your local gpu, I would recommend you to check tensorflow version, or try out lastest docker container image of kaggle or colab(which installed with tensorflow==2.12.0).
\nDependency issues are very annoying working with tensorflow, but I feel less issues recently working with tensorflow 2.12.0. hope this helps :)
", "rawMarkdown": "Hi, @bamps23\nThank you for sharing the issue! I just tested it with GPU(T4) in colab env, and It works fine without issue. \n\nIf you are using your local gpu, I would recommend you to check tensorflow version, or try out lastest docker container image of kaggle or colab(which installed with tensorflow==2.12.0). \n\nDependency issues are very annoying working with tensorflow, but I feel less issues recently working with tensorflow 2.12.0. hope this helps :)", "votes": null }, { "id": "2266809", "postDate": "05/20/2023 11:39:40", "content": "Thanks @hoyso48!
\nI found that changing the tensorflow version itself caused the issue.
\nMy Kaggle docker default tensorflow version was 2.11.0, then I installed 2.12.0 as following the top instruction. But it causes the error. After I made it back to 2.11.0, it works.
\nI don't know how it is, but tensorflow in Kaggle docker has to be carefully installedโฆ
Anyway, thanks for the answer and the code:)
", "rawMarkdown": "Thanks @hoyso48!\nI found that changing the tensorflow version itself caused the issue.\nMy Kaggle docker default tensorflow version was 2.11.0, then I installed 2.12.0 as following the top instruction. But it causes the error. After I made it back to 2.11.0, it works.\nI don't know how it is, but tensorflow in Kaggle docker has to be carefully installed...\n\nAnyway, thanks for the answer and the code:)", "votes": null }, { "id": "2330908", "postDate": "07/05/2023 08:31:55", "content": "Hi @hoyso48, thank you so much for sharing the code, and congrats on this work! Do you think you could also share the code that you mentioned started with Pytorch? I would really appreciate it, even if it is a draw and not the definitive version. I am having some issues replicating the code from Keras to Pytorch.
\nThanks a lot!
Hi! I'm sorry but I don't think I can help with that. The Pytorch version doesn't have any overlap with the solution. It only contains implementations of GCN and pretrained 2D-CNN, and the content is completely different from the TensorFlow version. I recommend looking for a Pytorch implementation among other winner's solutions, and try to implement it by referring only to the model. Thanks!
", "rawMarkdown": "Hi! I'm sorry but I don't think I can help with that. The Pytorch version doesn't have any overlap with the solution. It only contains implementations of GCN and pretrained 2D-CNN, and the content is completely different from the TensorFlow version. I recommend looking for a Pytorch implementation among other winner's solutions, and try to implement it by referring only to the model. Thanks!", "votes": null }, { "id": "2331441", "postDate": "07/05/2023 14:49:32", "content": "Okay, thanks for the suggestion!
", "rawMarkdown": "Okay, thanks for the suggestion!", "votes": null } ], "comments": [ { "id": 2246088, "author_name": "tatamikenn", "author_url": "", "post_date": "05/04/2023 21:35:20", "content": "Thank you for publishing your code. It saves much time comparing to reproducing myself.
", "votes": null, "replies": [] }, { "id": 2246626, "author_name": "kolyaforrat", "author_url": "", "post_date": "05/05/2023 10:23:36", "content": "I have submitted your single model + our single model with 0.5/0.5 wieghts ensemble and got this score:
\n
0.82 public beaten ๐
", "votes": null, "replies": [ { "id": 2246756, "author_name": "hoyso48", "author_url": "", "post_date": "05/05/2023 12:42:02", "content": "Wow. I didn't expect that much score improvement. Thank you for sharing! ๐
", "votes": null, "replies": [] } ] }, { "id": 2246883, "author_name": "ivanisaev", "author_url": "", "post_date": "05/05/2023 14:34:04", "content": "Thanks for posting the code!! I find it very useful for those who are just learning how to create models.
", "votes": null, "replies": [] }, { "id": 2257044, "author_name": "maksymstetsenko", "author_url": "", "post_date": "05/13/2023 00:45:01", "content": "Thank a lot for sharing! I think, even competition host did not expect such result.
", "votes": null, "replies": [] }, { "id": 2258019, "author_name": "aisuluuulankyzy", "author_url": "", "post_date": "05/13/2023 19:59:08", "content": "thnx that you published this post
", "votes": null, "replies": [] }, { "id": 2266464, "author_name": "bamps53", "author_url": "", "post_date": "05/20/2023 05:22:07", "content": "@hoyso48
\nThanks for sharing the code, that's really helpful!
\nHave you ever tried to run this code with GPU?
\nI've tried on CPU, GPU and TPU, but somehow only on GPU I got following error..
\nHave you experiend something like this?
Hi, @bamps23
\nThank you for sharing the issue! I just tested it with GPU(T4) in colab env, and It works fine without issue.
If you are using your local gpu, I would recommend you to check tensorflow version, or try out lastest docker container image of kaggle or colab(which installed with tensorflow==2.12.0).
\nDependency issues are very annoying working with tensorflow, but I feel less issues recently working with tensorflow 2.12.0. hope this helps :)
", "votes": null, "replies": [ { "id": 2266809, "author_name": "bamps53", "author_url": "", "post_date": "05/20/2023 11:39:40", "content": "Thanks @hoyso48!
\nI found that changing the tensorflow version itself caused the issue.
\nMy Kaggle docker default tensorflow version was 2.11.0, then I installed 2.12.0 as following the top instruction. But it causes the error. After I made it back to 2.11.0, it works.
\nI don't know how it is, but tensorflow in Kaggle docker has to be carefully installedโฆ
Anyway, thanks for the answer and the code:)
", "votes": null, "replies": [] } ] } ] }, { "id": 2330908, "author_name": "laiatarrsbenet", "author_url": "", "post_date": "07/05/2023 08:31:55", "content": "Hi @hoyso48, thank you so much for sharing the code, and congrats on this work! Do you think you could also share the code that you mentioned started with Pytorch? I would really appreciate it, even if it is a draw and not the definitive version. I am having some issues replicating the code from Keras to Pytorch.
\nThanks a lot!
Hi! I'm sorry but I don't think I can help with that. The Pytorch version doesn't have any overlap with the solution. It only contains implementations of GCN and pretrained 2D-CNN, and the content is completely different from the TensorFlow version. I recommend looking for a Pytorch implementation among other winner's solutions, and try to implement it by referring only to the model. Thanks!
", "votes": null, "replies": [ { "id": 2331441, "author_name": "laiatarrsbenet", "author_url": "", "post_date": "07/05/2023 14:49:32", "content": "Okay, thanks for the suggestion!
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2245887": "training notebook:\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-training\n- inference notebook:\nhttps://www.kaggle.com/code/hoyso48/1st-place-solution-inference\n\n- original colab notebook I used:\nhttps://github.com/hoyso48/Google---Isolated-Sign-Language-Recognition-1st-place-solution\n\nCurrently, Kaggle TPU-VM training in the training notebook is strangely slow (2x times slower than Colab). Let me know how to debug this issue if you are familiar with TPU and TensorFlow.\n\nreproducible result from the above training notebook is as follows.\n\n\nsingle seed model of 17ms latency took around 20minutes and got 2nd place in private\n\n\n\n\nThe 4-seed model often fails with submission scoring error. You can try a 3-seed model and may get a higher private score(?).\n\nIf you have any questions or issues regarding these codes, please leave a comment below. Thanks!", "2246088": "Thank you for publishing your code. It saves much time comparing to reproducing myself.", "2246626": "I have submitted your single model + our single model with 0.5/0.5 wieghts ensemble and got this score:\n\n\n\n0.82 public beaten ๐", "2246756": "Wow. I didn't expect that much score improvement. Thank you for sharing! ๐", "2246883": "Thanks for posting the code!! I find it very useful for those who are just learning how to create models.", "2257044": "Thank a lot for sharing! I think, even competition host did not expect such result.", "2258019": "thnx that you published this post", "2266464": "hoyso48 \nThanks for sharing the code, that's really helpful!\nHave you ever tried to run this code with GPU?\nI've tried on CPU, GPU and TPU, but somehow only on GPU I got following error..\nHave you experiend something like this?\n\n", "2266579": "Hi, @bamps23\nThank you for sharing the issue! I just tested it with GPU(T4) in colab env, and It works fine without issue. \n\nIf you are using your local gpu, I would recommend you to check tensorflow version, or try out lastest docker container image of kaggle or colab(which installed with tensorflow==2.12.0). \n\nDependency issues are very annoying working with tensorflow, but I feel less issues recently working with tensorflow 2.12.0. hope this helps :)", "2266809": "Thanks @hoyso48!\nI found that changing the tensorflow version itself caused the issue.\nMy Kaggle docker default tensorflow version was 2.11.0, then I installed 2.12.0 as following the top instruction. But it causes the error. After I made it back to 2.11.0, it works.\nI don't know how it is, but tensorflow in Kaggle docker has to be carefully installed...\n\nAnyway, thanks for the answer and the code:)", "2330908": "Hi @hoyso48, thank you so much for sharing the code, and congrats on this work! Do you think you could also share the code that you mentioned started with Pytorch? I would really appreciate it, even if it is a draw and not the definitive version. I am having some issues replicating the code from Keras to Pytorch.\nThanks a lot!", "2331416": "Hi! I'm sorry but I don't think I can help with that. The Pytorch version doesn't have any overlap with the solution. It only contains implementations of GCN and pretrained 2D-CNN, and the content is completely different from the TensorFlow version. I recommend looking for a Pytorch implementation among other winner's solutions, and try to implement it by referring only to the model. Thanks!", "2331441": "Okay, thanks for the suggestion!" }, "source": "meta" }