I want to use transfer learning for this competition but i don't know is it ok or not??
", "messages": [ { "id": 1115734, "postDate": "2020-12-16T14:21:32.450Z", "content": "yes it is allowed to use transfer learning or pre trained models.
\nBut as internet is not allowed in this competition , what you need to do is , create another notebook to train your models which will have internet enabled and then once it is committed and saved , create a new dataset using that notebook.
Now create another notebook , just for inference, import that previously created dataset and use those model to perform inference on your test data.
\nIn this way you can easily perform transfer learning .
I hope this helps.
", "rawMarkdown": "yes it is allowed to use transfer learning or pre trained models.\nBut as internet is not allowed in this competition , what you need to do is , create another notebook to train your models which will have internet enabled and then once it is committed and saved , create a new dataset using that notebook.\n\nNow create another notebook , just for inference, import that previously created dataset and use those model to perform inference on your test data.\nIn this way you can easily perform transfer learning .\n\nI hope this helps.", "votes": 4 }, { "id": 1115572, "postDate": "2020-12-16T11:38:32.690Z", "content": "I want to use transfer learning for this competition but i don't know is it ok or not??
", "rawMarkdown": "I want to use transfer learning for this competition but i don't know is it ok or not??", "votes": 2 }, { "id": 1117486, "postDate": "2020-12-18T05:54:13.933Z", "content": "Why not; training your custom network may take a large amount of time and still you may not get the results you prefer.
\nThere are several good models you can use like efficientnet, se-resnext, resnext, VIT etc etc.
\nDo read their papers and understand how they work before implementation.
Download the pre trained model trained on ImageNet available in Github locally and upload them as dataset into the kernel. Further on, you can load the model offline and train using it without violation of the protocol.
", "rawMarkdown": "Download the pre trained model trained on ImageNet available in Github locally and upload them as dataset into the kernel. Further on, you can load the model offline and train using it without violation of the protocol. ", "votes": 2 }, { "id": 1117487, "postDate": "2020-12-18T05:54:13.933Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1116364, "postDate": "2020-12-17T05:35:57.763Z", "content": "@prashantarorat Thank for your help..
", "rawMarkdown": "@prashantarorat Thank for your help.." } ], "comments": [ { "id": 1115734, "author_name": "Prashant Arora", "author_url": "", "post_date": "2020-12-16T14:21:32.450000", "content": "yes it is allowed to use transfer learning or pre trained models.
\nBut as internet is not allowed in this competition , what you need to do is , create another notebook to train your models which will have internet enabled and then once it is committed and saved , create a new dataset using that notebook.
Now create another notebook , just for inference, import that previously created dataset and use those model to perform inference on your test data.
\nIn this way you can easily perform transfer learning .
I hope this helps.
", "votes": 4, "replies": [] }, { "id": 1117486, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2020-12-18T05:54:13.933000", "content": "Why not; training your custom network may take a large amount of time and still you may not get the results you prefer.
\nThere are several good models you can use like efficientnet, se-resnext, resnext, VIT etc etc.
\nDo read their papers and understand how they work before implementation.
Download the pre trained model trained on ImageNet available in Github locally and upload them as dataset into the kernel. Further on, you can load the model offline and train using it without violation of the protocol.
", "votes": 2, "replies": [] }, { "id": 1117487, "author_name": "", "author_url": "", "post_date": "2020-12-18T05:54:13.933000", "content": "", "votes": 0, "replies": [] }, { "id": 1116364, "author_name": "Jeelkumar Gondaliya", "author_url": "", "post_date": "2020-12-17T05:35:57.763000", "content": "@prashantarorat Thank for your help..
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1115734": "yes it is allowed to use transfer learning or pre trained models.\nBut as internet is not allowed in this competition , what you need to do is , create another notebook to train your models which will have internet enabled and then once it is committed and saved , create a new dataset using that notebook.\n\nNow create another notebook , just for inference, import that previously created dataset and use those model to perform inference on your test data.\nIn this way you can easily perform transfer learning .\n\nI hope this helps.", "1115572": "I want to use transfer learning for this competition but i don't know is it ok or not??", "1117486": "Why not; training your custom network may take a large amount of time and still you may not get the results you prefer.\nThere are several good models you can use like efficientnet, se-resnext, resnext, VIT etc etc.\nDo read their papers and understand how they work before implementation.", "1122948": "Download the pre trained model trained on ImageNet available in Github locally and upload them as dataset into the kernel. Further on, you can load the model offline and train using it without violation of the protocol. ", "1117487": "", "1116364": "@prashantarorat Thank for your help.." } }