Putdown some Good to know facts regarding the common techniques for the starters in the community.
", "messages": [ { "id": "1101176", "postDate": "12/03/2020 17:34:56", "content": "Putdown some Good to know facts regarding the common techniques for the starters in the community.
", "rawMarkdown": "Putdown some Good to know facts regarding the common techniques for the starters in the community.", "votes": null }, { "id": "1101198", "postDate": "12/03/2020 17:51:13", "content": "The most important thing that every kaggler say is \" Always believe your CV \"; so preparing a good validation scheme is very important.
", "rawMarkdown": "The most important thing that every kaggler say is \" Always believe your CV \"; so preparing a good validation scheme is very important.", "votes": null }, { "id": "1101255", "postDate": "12/03/2020 18:50:33", "content": "Try everything.
", "rawMarkdown": "Try everything.", "votes": null }, { "id": "1101331", "postDate": "12/03/2020 20:07:07", "content": "Yeah, I do. Sometimes the network doesnt learn. it took me a week to figureout the problem of loss function. does it case for everyone? :(
", "rawMarkdown": "Yeah, I do. Sometimes the network doesnt learn. it took me a week to figureout the problem of loss function. does it case for everyone? :(", "votes": null }, { "id": "1101343", "postDate": "12/03/2020 20:22:14", "content": "@bmanikan Try to overfit one batch of images on your model first, if that does not work with a state of the art model like EfficientNet or ResNet then you have problems with your configuration (loss, how you train, no use of gradients etc.)
", "rawMarkdown": "bmanikan Try to overfit one batch of images on your model first, if that does not work with a state of the art model like EfficientNet or ResNet then you have problems with your configuration (loss, how you train, no use of gradients etc.)", "votes": null }, { "id": "1101459", "postDate": "12/03/2020 23:22:16", "content": "If your model is spending a lot of epochs to reach an accuracy of ~0.6, you are wasting your time and resources. This score can be reached by predicting all the data with the most frequent class.
", "rawMarkdown": "If your model is spending a lot of epochs to reach an accuracy of ~0.6, you are wasting your time and resources. This score can be reached by predicting all the data with the most frequent class.", "votes": null }, { "id": "1101591", "postDate": "12/04/2020 03:53:08", "content": "@piantic Thanks for pointing some steps, As you said \"3. Try one by one.\" - > Im doing like for each modification i will run the entire train loop to see whether it benefits in that. The problem is every time we train the model for 20 or 30 epochs which takes minimum an hour considering the size of data.
\n@aliabdin1 This is new, I will try that out first. Thanks man!
", "rawMarkdown": "piantic Thanks for pointing some steps, As you said \"3. Try one by one.\" - > Im doing like for each modification i will run the entire train loop to see whether it benefits in that. The problem is every time we train the model for 20 or 30 epochs which takes minimum an hour considering the size of data. \n\n@aliabdin1 This is new, I will try that out first. Thanks man!", "votes": null }, { "id": "1101593", "postDate": "12/04/2020 03:55:03", "content": "It starts at 0.65 and not going beyond 0.8. If we try some steps the loss and accuracy stays static.
", "rawMarkdown": "It starts at 0.65 and not going beyond 0.8. If we try some steps the loss and accuracy stays static.", "votes": null }, { "id": "1101767", "postDate": "12/04/2020 08:20:14", "content": "Monitor not only train and validation loss but the gap between them. I mean if you modified your model and both train and valid loss became better but gap between them became larger, maybe this model is not that good.
", "rawMarkdown": "Monitor not only train and validation loss but the gap between them. I mean if you modified your model and both train and valid loss became better but gap between them became larger, maybe this model is not that good.", "votes": null }, { "id": "1101779", "postDate": "12/04/2020 08:45:23", "content": "Are you using a LR-Scheduler? What is your batch-size? You could be stuck in a local minima.
", "rawMarkdown": "Are you using a LR-Scheduler? What is your batch-size? You could be stuck in a local minima.", "votes": null } ], "comments": [ { "id": 1101198, "author_name": "mrinath", "author_url": "", "post_date": "12/03/2020 17:51:13", "content": "The most important thing that every kaggler say is \" Always believe your CV \"; so preparing a good validation scheme is very important.
", "votes": null, "replies": [] }, { "id": 1101255, "author_name": "piantic", "author_url": "", "post_date": "12/03/2020 18:50:33", "content": "Try everything.
", "votes": null, "replies": [ { "id": 1101331, "author_name": "bmanikan", "author_url": "", "post_date": "12/03/2020 20:07:07", "content": "Yeah, I do. Sometimes the network doesnt learn. it took me a week to figureout the problem of loss function. does it case for everyone? :(
", "votes": null, "replies": [] }, { "id": 1101343, "author_name": "piantic", "author_url": "", "post_date": "12/03/2020 20:22:14", "content": "@bmanikan Try to overfit one batch of images on your model first, if that does not work with a state of the art model like EfficientNet or ResNet then you have problems with your configuration (loss, how you train, no use of gradients etc.)
", "votes": null, "replies": [] }, { "id": 1101591, "author_name": "bmanikan", "author_url": "", "post_date": "12/04/2020 03:53:08", "content": "@piantic Thanks for pointing some steps, As you said \"3. Try one by one.\" - > Im doing like for each modification i will run the entire train loop to see whether it benefits in that. The problem is every time we train the model for 20 or 30 epochs which takes minimum an hour considering the size of data.
\n@aliabdin1 This is new, I will try that out first. Thanks man!
", "votes": null, "replies": [] } ] }, { "id": 1101459, "author_name": "tolgadincer", "author_url": "", "post_date": "12/03/2020 23:22:16", "content": "If your model is spending a lot of epochs to reach an accuracy of ~0.6, you are wasting your time and resources. This score can be reached by predicting all the data with the most frequent class.
", "votes": null, "replies": [ { "id": 1101593, "author_name": "bmanikan", "author_url": "", "post_date": "12/04/2020 03:55:03", "content": "It starts at 0.65 and not going beyond 0.8. If we try some steps the loss and accuracy stays static.
", "votes": null, "replies": [] }, { "id": 1101779, "author_name": "aliabdin1", "author_url": "", "post_date": "12/04/2020 08:45:23", "content": "Are you using a LR-Scheduler? What is your batch-size? You could be stuck in a local minima.
", "votes": null, "replies": [] } ] }, { "id": 1101767, "author_name": "dunklerwald", "author_url": "", "post_date": "12/04/2020 08:20:14", "content": "Monitor not only train and validation loss but the gap between them. I mean if you modified your model and both train and valid loss became better but gap between them became larger, maybe this model is not that good.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1101176": "Putdown some Good to know facts regarding the common techniques for the starters in the community.", "1101198": "The most important thing that every kaggler say is \" Always believe your CV \"; so preparing a good validation scheme is very important.", "1101255": "Try everything.", "1101331": "Yeah, I do. Sometimes the network doesnt learn. it took me a week to figureout the problem of loss function. does it case for everyone? :(", "1101343": "1. Use good baseline.\n - There are many good notebooks.\n2. Start small model.\n - You can see result very quickly.\n3. Try one by one.\n - Check your model predictions.\n4. Find useful things to improve your model score.\n5. See datasets.\n\n@bmanikan", "1101371": "bmanikan Try to overfit one batch of images on your model first, if that does not work with a state of the art model like EfficientNet or ResNet then you have problems with your configuration (loss, how you train, no use of gradients etc.)", "1101459": "If your model is spending a lot of epochs to reach an accuracy of ~0.6, you are wasting your time and resources. This score can be reached by predicting all the data with the most frequent class.", "1101591": "piantic Thanks for pointing some steps, As you said \"3. Try one by one.\" - > Im doing like for each modification i will run the entire train loop to see whether it benefits in that. The problem is every time we train the model for 20 or 30 epochs which takes minimum an hour considering the size of data. \n\n@aliabdin1 This is new, I will try that out first. Thanks man!", "1101593": "It starts at 0.65 and not going beyond 0.8. If we try some steps the loss and accuracy stays static.", "1101767": "Monitor not only train and validation loss but the gap between them. I mean if you modified your model and both train and valid loss became better but gap between them became larger, maybe this model is not that good.", "1101779": "Are you using a LR-Scheduler? What is your batch-size? You could be stuck in a local minima." }, "source": "meta" }