{ "id": 421788, "title": "Advice to newbie about learning parameters", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/421788", "author_name": "", "post_date": "2023-07-06T18:53:02.439460200Z", "votes": 4, "comment_count": 6, "views": 0, "content": "

Hello everyone!
\nFirst of all, thank you everyone, who publishes their code with some models or preprocessing data, it is really useful especially if you have theory of deep learning but zero practicing.

\n

But how do you find the best learning parameters for some model ?
\nDo you only use only Kaggle GPU or you rent Google Colab or another cloud service for running different models in parallel ?
\nAnd how many epochs does need model to compare parameters of learning ?
\nI understand it depends on experience, but maybe somebody can give advice about it.

", "messages": [ { "id": "2333244", "postDate": "07/06/2023 18:53:02", "content": "

Hello everyone!
\nFirst of all, thank you everyone, who publishes their code with some models or preprocessing data, it is really useful especially if you have theory of deep learning but zero practicing.

\n

But how do you find the best learning parameters for some model ?
\nDo you only use only Kaggle GPU or you rent Google Colab or another cloud service for running different models in parallel ?
\nAnd how many epochs does need model to compare parameters of learning ?
\nI understand it depends on experience, but maybe somebody can give advice about it.

", "rawMarkdown": "Hello everyone!\nFirst of all, thank you everyone, who publishes their code with some models or preprocessing data, it is really useful especially if you have theory of deep learning but zero practicing.\n\nBut how do you find the best learning parameters for some model ?\nDo you only use only Kaggle GPU or you rent Google Colab or another cloud service for running different models in parallel ?\nAnd how many epochs does need model to compare parameters of learning ?\nI understand it depends on experience, but maybe somebody can give advice about it.", "votes": null }, { "id": "2333969", "postDate": "07/07/2023 10:25:00", "content": "

At first I used only Kaggle GPU and now I am using a local RTX3090.
\nDifferent models have different appropriate parameters and I don't even know which model is best. Of course, I also don't know which augmentation works best.

\n

I do this by hand, looking at the results of each experiment one by one.
\nIn doing this, I refer to past Kaggle public notes and parameters from research papers.

\n

Optuna and grid search would be impractical for many, in terms of computational resources.
\nIf there is an efficient parameter tuning I would like to know too.
\nAnyone who knows more about this, please let me know.

", "rawMarkdown": "At first I used only Kaggle GPU and now I am using a local RTX3090.\nDifferent models have different appropriate parameters and I don't even know which model is best. Of course, I also don't know which augmentation works best.\n\nI do this by hand, looking at the results of each experiment one by one.\nIn doing this, I refer to past Kaggle public notes and parameters from research papers.\n\nOptuna and grid search would be impractical for many, in terms of computational resources.\nIf there is an efficient parameter tuning I would like to know too.\nAnyone who knows more about this, please let me know.", "votes": null }, { "id": "2334013", "postDate": "07/07/2023 11:29:44", "content": "

imo, Intuition based parameter tuning is probably the best for most of the deep learning tasks :))

", "rawMarkdown": "imo, **Intuition** based parameter tuning is probably the best for most of the deep learning tasks :))", "votes": null }, { "id": "2334161", "postDate": "07/07/2023 13:42:49", "content": "

Thank you for sharing your experience! It's really useful.
\nI agree with you about grid search.
\nHow do you compare results? Do you use w&b or do you check by leaderboard score ?

", "rawMarkdown": "Thank you for sharing your experience! It's really useful.\nI agree with you about grid search.\nHow do you compare results? Do you use w&b or do you check by leaderboard score ?", "votes": null }, { "id": "2334167", "postDate": "07/07/2023 13:45:09", "content": "

Oh, it looks like a true, but I hoped that maybe I've missed something :)

", "rawMarkdown": "Oh, it looks like a true, but I hoped that maybe I've missed something :)", "votes": null }, { "id": "2334736", "postDate": "07/07/2023 23:40:15", "content": "

I don't use wandb because sometimes I get unexplained connection errors and the experiment stops.
\nI use mmdetection for training, so it keeps logging.
\nI use test.py from the mmdetection tools to evaluate the model and compare the results with the leaderboard results.

", "rawMarkdown": "I don't use wandb because sometimes I get unexplained connection errors and the experiment stops.\nI use mmdetection for training, so it keeps logging.\nI use test.py from the mmdetection tools to evaluate the model and compare the results with the leaderboard results.", "votes": null }, { "id": "2334809", "postDate": "07/08/2023 02:29:32", "content": "

https://github.com/google-research/tuning_playbook

", "rawMarkdown": "https://github.com/google-research/tuning_playbook", "votes": null } ], "comments": [ { "id": 2333969, "author_name": "yosukeyama", "author_url": "", "post_date": "07/07/2023 10:25:00", "content": "

At first I used only Kaggle GPU and now I am using a local RTX3090.
\nDifferent models have different appropriate parameters and I don't even know which model is best. Of course, I also don't know which augmentation works best.

\n

I do this by hand, looking at the results of each experiment one by one.
\nIn doing this, I refer to past Kaggle public notes and parameters from research papers.

\n

Optuna and grid search would be impractical for many, in terms of computational resources.
\nIf there is an efficient parameter tuning I would like to know too.
\nAnyone who knows more about this, please let me know.

", "votes": null, "replies": [ { "id": 2334013, "author_name": "nischaydnk", "author_url": "", "post_date": "07/07/2023 11:29:44", "content": "

imo, Intuition based parameter tuning is probably the best for most of the deep learning tasks :))

", "votes": null, "replies": [ { "id": 2334167, "author_name": "jurassimo", "author_url": "", "post_date": "07/07/2023 13:45:09", "content": "

Oh, it looks like a true, but I hoped that maybe I've missed something :)

", "votes": null, "replies": [] } ] }, { "id": 2334161, "author_name": "jurassimo", "author_url": "", "post_date": "07/07/2023 13:42:49", "content": "

Thank you for sharing your experience! It's really useful.
\nI agree with you about grid search.
\nHow do you compare results? Do you use w&b or do you check by leaderboard score ?

", "votes": null, "replies": [ { "id": 2334736, "author_name": "yosukeyama", "author_url": "", "post_date": "07/07/2023 23:40:15", "content": "

I don't use wandb because sometimes I get unexplained connection errors and the experiment stops.
\nI use mmdetection for training, so it keeps logging.
\nI use test.py from the mmdetection tools to evaluate the model and compare the results with the leaderboard results.

", "votes": null, "replies": [] } ] } ] }, { "id": 2334809, "author_name": "atom1231", "author_url": "", "post_date": "07/08/2023 02:29:32", "content": "

https://github.com/google-research/tuning_playbook

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2333244": "Hello everyone!\nFirst of all, thank you everyone, who publishes their code with some models or preprocessing data, it is really useful especially if you have theory of deep learning but zero practicing.\n\nBut how do you find the best learning parameters for some model ?\nDo you only use only Kaggle GPU or you rent Google Colab or another cloud service for running different models in parallel ?\nAnd how many epochs does need model to compare parameters of learning ?\nI understand it depends on experience, but maybe somebody can give advice about it.", "2333969": "At first I used only Kaggle GPU and now I am using a local RTX3090.\nDifferent models have different appropriate parameters and I don't even know which model is best. Of course, I also don't know which augmentation works best.\n\nI do this by hand, looking at the results of each experiment one by one.\nIn doing this, I refer to past Kaggle public notes and parameters from research papers.\n\nOptuna and grid search would be impractical for many, in terms of computational resources.\nIf there is an efficient parameter tuning I would like to know too.\nAnyone who knows more about this, please let me know.", "2334013": "imo, **Intuition** based parameter tuning is probably the best for most of the deep learning tasks :))", "2334161": "Thank you for sharing your experience! It's really useful.\nI agree with you about grid search.\nHow do you compare results? Do you use w&b or do you check by leaderboard score ?", "2334167": "Oh, it looks like a true, but I hoped that maybe I've missed something :)", "2334736": "I don't use wandb because sometimes I get unexplained connection errors and the experiment stops.\nI use mmdetection for training, so it keeps logging.\nI use test.py from the mmdetection tools to evaluate the model and compare the results with the leaderboard results.", "2334809": "https://github.com/google-research/tuning_playbook" }, "source": "meta" }