Hey, I am a newbie on Kaggle and trying to sort out what the workflow and process for competitions are. Can anybody help me with these questions?
\n\nI am trying to scour the forums and google for answers but I found it difficult to find definitive answer and the Kaggle official post are all that informative for a newbie.
\n\nHey, I am a newbie on Kaggle and trying to sort out what the workflow and process for competitions are. Can anybody help me with these questions?
\n\nI am trying to scour the forums and google for answers but I found it difficult to find definitive answer and the Kaggle official post are all that informative for a newbie.
\n\nHere are the answers of a few questions (I'm not sure about other questions):
\n\n2) Yes that's the correct approach and most of the competitors follows it.\n5) Yes the run time limit only applies to inference kernel, which you submit to the competition. This is one of the main reasons why people follow the approach mentioned in step 2.\n6) Private dataset will be available at the end the competition. All of your submissions will be ran automatically on that dataset and the scores will be available on private dataset. To make sure that your notebook/inference run correctly on private dataset, make sure you are creating the submission.csv file as given in the sample_submission.csv.
\n\nHope this helps!
", "rawMarkdown": "Here are the answers of a few questions (I'm not sure about other questions):\n\n2) Yes that's the correct approach and most of the competitors follows it.\n5) Yes the run time limit only applies to inference kernel, which you submit to the competition. This is one of the main reasons why people follow the approach mentioned in step 2.\n6) Private dataset will be available at the end the competition. All of your submissions will be ran automatically on that dataset and the scores will be available on private dataset. To make sure that your notebook/inference run correctly on private dataset, make sure you are creating the submission.csv file as given in the sample\\_submission.csv.\n\nHope this helps!", "votes": null }, { "id": "887484", "postDate": "06/15/2020 17:38:00", "content": "To fill in here:\n1) This has been a common issue brought up in the discussion because the rules aren't very clear. Since it doesn't seem the organizers are active in the discussion section, I have emailed Kaggle about this and will post their response in the discussion once I hear back.\n3) You can take that implementation and upload it as a Kaggle dataset to get access to it in the kernel without internet access. However, as I stated in 1 the rules surrounding this are unclear and we'll have to see what they say regarding this.\n4) You can use external datasets only if they are free, publicly accessible to everyone, and have some free commercial use license. However, I would be very weary of this. First, there doesn't appear to be too many good datasets for this. The GWHD dataset is the first of it's kind and has pretty specific requirements to match. Secondly, it seems risky in general to use External Data in Kaggle competitions now... In the Facebook Deepfake detection challenge, many competitors were indiscriminately disqualified including the first place solution for using scraped videos from Youtube, even though they had the proper license, because they didn't have explicit written consent from the people in those videos (See Here).
", "rawMarkdown": "To fill in here:\n1) This has been a common issue brought up in the discussion because the rules aren't very clear. Since it doesn't seem the organizers are active in the discussion section, I have emailed Kaggle about this and will post their response in the discussion once I hear back.\n3) You can take that implementation and upload it as a Kaggle dataset to get access to it in the kernel without internet access. However, as I stated in 1 the rules surrounding this are unclear and we'll have to see what they say regarding this.\n4) You can use external datasets only if they are free, publicly accessible to everyone, and have some free commercial use license. However, I would be very weary of this. First, there doesn't appear to be too many good datasets for this. The GWHD dataset is the first of it's kind and has pretty specific requirements to match. Secondly, it seems risky in general to use External Data in Kaggle competitions now... In the Facebook Deepfake detection challenge, many competitors were indiscriminately disqualified including the first place solution for using scraped videos from Youtube, even though they had the proper license, because they didn't have explicit written consent from the people in those videos ([See Here](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/157983)).", "votes": null }, { "id": "888624", "postDate": "06/16/2020 13:24:21", "content": "Organizer here (sorry for the recent lack of activity)
\n\nExternal data and library are allowed if available to other competitors and open-source. If you succeed to leverage open-source data to solve the generalization problem it's great ! However if you use private data that you can't share, your solution will likely to be disqualify. @matthewmasters I did not find the other discussions but would be happy to clarify this point for you.
", "rawMarkdown": "Organizer here (sorry for the recent lack of activity)\n\nExternal data and library are allowed if available to other competitors and open-source. If you succeed to leverage open-source data to solve the generalization problem it's great ! However if you use private data that you can't share, your solution will likely to be disqualify. @matthewmasters I did not find the other discussions but would be happy to clarify this point for you.", "votes": null } ], "comments": [ { "id": 887396, "author_name": "kaushal2896", "author_url": "", "post_date": "06/15/2020 16:44:53", "content": "Here are the answers of a few questions (I'm not sure about other questions):
\n\n2) Yes that's the correct approach and most of the competitors follows it.\n5) Yes the run time limit only applies to inference kernel, which you submit to the competition. This is one of the main reasons why people follow the approach mentioned in step 2.\n6) Private dataset will be available at the end the competition. All of your submissions will be ran automatically on that dataset and the scores will be available on private dataset. To make sure that your notebook/inference run correctly on private dataset, make sure you are creating the submission.csv file as given in the sample_submission.csv.
\n\nHope this helps!
", "votes": null, "replies": [ { "id": 887484, "author_name": "matthewmasters", "author_url": "", "post_date": "06/15/2020 17:38:00", "content": "To fill in here:\n1) This has been a common issue brought up in the discussion because the rules aren't very clear. Since it doesn't seem the organizers are active in the discussion section, I have emailed Kaggle about this and will post their response in the discussion once I hear back.\n3) You can take that implementation and upload it as a Kaggle dataset to get access to it in the kernel without internet access. However, as I stated in 1 the rules surrounding this are unclear and we'll have to see what they say regarding this.\n4) You can use external datasets only if they are free, publicly accessible to everyone, and have some free commercial use license. However, I would be very weary of this. First, there doesn't appear to be too many good datasets for this. The GWHD dataset is the first of it's kind and has pretty specific requirements to match. Secondly, it seems risky in general to use External Data in Kaggle competitions now... In the Facebook Deepfake detection challenge, many competitors were indiscriminately disqualified including the first place solution for using scraped videos from Youtube, even though they had the proper license, because they didn't have explicit written consent from the people in those videos (See Here).
", "votes": null, "replies": [] }, { "id": 888624, "author_name": "bendvd", "author_url": "", "post_date": "06/16/2020 13:24:21", "content": "Organizer here (sorry for the recent lack of activity)
\n\nExternal data and library are allowed if available to other competitors and open-source. If you succeed to leverage open-source data to solve the generalization problem it's great ! However if you use private data that you can't share, your solution will likely to be disqualify. @matthewmasters I did not find the other discussions but would be happy to clarify this point for you.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "886918": "Hey, I am a newbie on Kaggle and trying to sort out what the workflow and process for competitions are. Can anybody help me with these questions? \n\nI am trying to scour the forums and google for answers but I found it difficult to find definitive answer and the Kaggle official post are all that informative for a newbie.\n\n1. What does 'External data, freely & publicly available, is allowed. This includes pre-trained models.' and 'No custom packages enabled in kernels' in the code requirements section mean?\n2. What is the process for training a model for a competition? From what I can tell atm people use one notebook to train a model through transfer learning and save it and then in a new notebook add the dataset from the first model and use the saved model for inference and submission. is this the right workflow?\n3. Related to the previous two questions, can i add an github implementation of an ANN like [this one](https://github.com/xuannianz/EfficientDet)? how does this relate to the [External data thread](https://www.kaggle.com/c/global-wheat-detection/discussion/148561) post?\n4. Also can i add extra data for training i scrapped online?\n5. If the train+inference notebook workflow is correct are the run time limits (CPU/GPU) only applied to the inference step?\n6. What happens when the private test data becomes available? do we have to run inference on that new data manually? is it done automatically? does the notebook have to follow certain guidelines for it to run automatically?", "887396": "Here are the answers of a few questions (I'm not sure about other questions):\n\n2) Yes that's the correct approach and most of the competitors follows it.\n5) Yes the run time limit only applies to inference kernel, which you submit to the competition. This is one of the main reasons why people follow the approach mentioned in step 2.\n6) Private dataset will be available at the end the competition. All of your submissions will be ran automatically on that dataset and the scores will be available on private dataset. To make sure that your notebook/inference run correctly on private dataset, make sure you are creating the submission.csv file as given in the sample\\_submission.csv.\n\nHope this helps!", "887484": "To fill in here:\n1) This has been a common issue brought up in the discussion because the rules aren't very clear. Since it doesn't seem the organizers are active in the discussion section, I have emailed Kaggle about this and will post their response in the discussion once I hear back.\n3) You can take that implementation and upload it as a Kaggle dataset to get access to it in the kernel without internet access. However, as I stated in 1 the rules surrounding this are unclear and we'll have to see what they say regarding this.\n4) You can use external datasets only if they are free, publicly accessible to everyone, and have some free commercial use license. However, I would be very weary of this. First, there doesn't appear to be too many good datasets for this. The GWHD dataset is the first of it's kind and has pretty specific requirements to match. Secondly, it seems risky in general to use External Data in Kaggle competitions now... In the Facebook Deepfake detection challenge, many competitors were indiscriminately disqualified including the first place solution for using scraped videos from Youtube, even though they had the proper license, because they didn't have explicit written consent from the people in those videos ([See Here](https://www.kaggle.com/c/deepfake-detection-challenge/discussion/157983)).", "888624": "Organizer here (sorry for the recent lack of activity)\n\nExternal data and library are allowed if available to other competitors and open-source. If you succeed to leverage open-source data to solve the generalization problem it's great ! However if you use private data that you can't share, your solution will likely to be disqualify. @matthewmasters I did not find the other discussions but would be happy to clarify this point for you." }, "source": "meta" }