{ "id": 375843, "title": "Train on local machine and import weights?", "url": "/competitions/rsna-breast-cancer-detection/discussion/375843", "author_name": "Vin Bhaskara", "post_date": "2023-01-03T18:40:05.200000", "votes": 6, "comment_count": 5, "views": 0, "content": "

Bonjour / Hi! This is my first code competition on Kaggle.

\n

I am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?

\n

Also, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?

\n

If that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions
\nWhat am I missing?

\n

Merci / Thanks!
\nHappy New Year / Bonne annee!

", "messages": [ { "id": 2084794, "postDate": "2023-01-03T18:40:05.200Z", "content": "

Bonjour / Hi! This is my first code competition on Kaggle.

\n

I am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?

\n

Also, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?

\n

If that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions
\nWhat am I missing?

\n

Merci / Thanks!
\nHappy New Year / Bonne annee!

", "rawMarkdown": "Bonjour / Hi! This is my first code competition on Kaggle. \n\nI am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?\n\nAlso, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?\n\nIf that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions \nWhat am I missing?\n\nMerci / Thanks! \nHappy New Year / Bonne annee!", "votes": 6 }, { "id": 2084840, "postDate": "2023-01-03T19:38:12.557Z", "content": "
\n

I am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?

\n
\n

Yes

\n
\n

Also, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?

\n
\n

No, you can keep your models as private datasets

\n
\n

If that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions

\n
\n

Inference resources are limited and same for all competitors, mentioned here - https://www.kaggle.com/competitions/rsna-breast-cancer-detection/overview/code-requirements

", "rawMarkdown": ">I am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?\n\nYes \n\n\n>Also, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?\n\nNo, you can keep your models as private datasets \n\n>If that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions\n\nInference resources are limited and same for all competitors, mentioned here - https://www.kaggle.com/competitions/rsna-breast-cancer-detection/overview/code-requirements\n\n ", "replies": [ { "id": 2084866, "postDate": "2023-01-03T19:52:05.143Z", "content": "

I see, thank you, that makes sense now. So the resources are capped only for inference but not for training. Thanks!

", "rawMarkdown": "I see, thank you, that makes sense now. So the resources are capped only for inference but not for training. Thanks!", "votes": 2, "replies": [ { "id": 2085938, "postDate": "2023-01-04T13:30:28.190Z", "content": "

It's a good question mate!!Frankly I had the same question for so long!!Finally it got cleared😊.Thank you!

", "rawMarkdown": "It's a good question mate!!Frankly I had the same question for so long!!Finally it got cleared😊.Thank you!" } ] }, { "id": 2085928, "postDate": "2023-01-04T13:25:38.400Z", "content": "

I am still confused about the rules.
\nBecause, if we can keep estimated models on personal machine as private data set, I don't understand the sentence:
\n\"Freely & publicly available external data is allowed, including pre-trained models\"

\n

To me, this implies that non-public models are forbidden, otherwise the sentence should be:
\n\"All external data is allowed, including personal trained models\"

\n

Moreover, the resources for the inference are much lower than for the training.

", "rawMarkdown": "I am still confused about the rules.\nBecause, if we can keep estimated models on personal machine as private data set, I don't understand the sentence:\n\"Freely & publicly available external data is allowed, including pre-trained models\"\n\nTo me, this implies that non-public models are forbidden, otherwise the sentence should be:\n\"All external data is allowed, including personal trained models\"\n\nMoreover, the resources for the inference are much lower than for the training.\n", "votes": 3, "replies": [ { "id": 2086687, "postDate": "2023-01-04T23:58:48.633Z", "content": "

Right. Maybe someone from Kaggle can confirm.

\n

@maggiemd - Could you please clarify the following for the rest of us:

\n
    \n
  1. Is it OK to download the training data and run experiments on our personal machines, and then just upload the trained model into an Inference Kaggle Notebook for submitting the predictions? While doing so, can we keep such locally trained models private? That is to say, use personal resources for training, and Kaggle's compute resources only for inference?
    2. \n
  2. Are the mentioned resource limits then only for inference (if one choose to train models on their personal resources)?
    3. \n
\n

Thanks a lot!

", "rawMarkdown": "Right. Maybe someone from Kaggle can confirm. \n\n@maggiemd - Could you please clarify the following for the rest of us:\n1. Is it OK to download the training data and run experiments on our personal machines, and then just upload the trained model into an Inference Kaggle Notebook for submitting the predictions? While doing so, can we keep such locally trained models private? That is to say, use personal resources for training, and Kaggle's compute resources only for inference?\n2. Are the mentioned resource limits then only for inference (if one choose to train models on their personal resources)?\n\nThanks a lot! ", "votes": 2 } ] } ] } ], "comments": [ { "id": 2084840, "author_name": "RB", "author_url": "", "post_date": "2023-01-03T19:38:12.557000", "content": "
\n

I am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?

\n
\n

Yes

\n
\n

Also, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?

\n
\n

No, you can keep your models as private datasets

\n
\n

If that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions

\n
\n

Inference resources are limited and same for all competitors, mentioned here - https://www.kaggle.com/competitions/rsna-breast-cancer-detection/overview/code-requirements

", "votes": 0, "replies": [ { "id": 2084866, "author_name": "Vin Bhaskara", "author_url": "", "post_date": "2023-01-03T19:52:05.143000", "content": "

I see, thank you, that makes sense now. So the resources are capped only for inference but not for training. Thanks!

", "votes": 2, "replies": [ { "id": 2085938, "author_name": "Senapati Rajesh", "author_url": "", "post_date": "2023-01-04T13:30:28.190000", "content": "

It's a good question mate!!Frankly I had the same question for so long!!Finally it got cleared😊.Thank you!

", "votes": 0, "replies": [] } ] }, { "id": 2085928, "author_name": "Joseph Rynkiewicz", "author_url": "", "post_date": "2023-01-04T13:25:38.400000", "content": "

I am still confused about the rules.
\nBecause, if we can keep estimated models on personal machine as private data set, I don't understand the sentence:
\n\"Freely & publicly available external data is allowed, including pre-trained models\"

\n

To me, this implies that non-public models are forbidden, otherwise the sentence should be:
\n\"All external data is allowed, including personal trained models\"

\n

Moreover, the resources for the inference are much lower than for the training.

", "votes": 3, "replies": [ { "id": 2086687, "author_name": "Vin Bhaskara", "author_url": "", "post_date": "2023-01-04T23:58:48.633000", "content": "

Right. Maybe someone from Kaggle can confirm.

\n

@maggiemd - Could you please clarify the following for the rest of us:

\n
    \n
  1. Is it OK to download the training data and run experiments on our personal machines, and then just upload the trained model into an Inference Kaggle Notebook for submitting the predictions? While doing so, can we keep such locally trained models private? That is to say, use personal resources for training, and Kaggle's compute resources only for inference?
    2. \n
  2. Are the mentioned resource limits then only for inference (if one choose to train models on their personal resources)?
    3. \n
\n

Thanks a lot!

", "votes": 2, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2084794": "Bonjour / Hi! This is my first code competition on Kaggle. \n\nI am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?\n\nAlso, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?\n\nIf that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions \nWhat am I missing?\n\nMerci / Thanks! \nHappy New Year / Bonne annee!", "2084840": ">I am not sure if I can download the data and train models on my personal machine, and then just import the model weights into a Kaggle notebook for submitting the predictions. Is that allowed?\n\nYes \n\n\n>Also, would such a \"pretrained model\" trained on my local machine need to be made publicly available in order to be used in my inference Kaggle notebook?\n\nNo, you can keep your models as private datasets \n\n>If that is indeed allowed, I don't quite see how code competitions are more balanced in terms of the hardware allowances as mentioned here: https://www.kaggle.com/docs/competitions\n\nInference resources are limited and same for all competitors, mentioned here - https://www.kaggle.com/competitions/rsna-breast-cancer-detection/overview/code-requirements\n\n " } }