![\"enter](\"https://i.imgflip.com/2cnaz2.jpg\")
anyone?
", "messages": [ { "id": "342262", "postDate": "06/13/2018 07:40:31", "content": "anyone?
", "rawMarkdown": "anyone?", "votes": null }, { "id": "342270", "postDate": "06/13/2018 07:49:16", "content": "They help - at least for us. Not the simple ones like blurriness or average color but \"deeper\" ones.
", "rawMarkdown": "They help - at least for us. Not the simple ones like blurriness or average color but \"deeper\" ones.", "votes": null }, { "id": "342274", "postDate": "06/13/2018 08:05:09", "content": "Yep, I've got a boost in both XGB and LGB with simpler image features.
", "rawMarkdown": "Yep, I've got a boost in both XGB and LGB with simpler image features.", "votes": null }, { "id": "342372", "postDate": "06/13/2018 11:57:31", "content": "helping a lot
", "rawMarkdown": "helping a lot", "votes": null }, { "id": "342469", "postDate": "06/13/2018 15:06:50", "content": "I've got at least a 0.0012 boost from image features.
", "rawMarkdown": "I've got at least a 0.0012 boost from image features.", "votes": null }, { "id": "342982", "postDate": "06/14/2018 13:31:19", "content": "Interesting, I'm a newbie and I have the contradictory case. Simple ones like blurriness and average color give me small boost.
\n\nOn the other hand, I found advanced ones (like probability predicted by Resnet) make my model worse, even though they have a correlation with deal probability higher than some important features indicated by my model.
", "rawMarkdown": "Interesting, I'm a newbie and I have the contradictory case. Simple ones like blurriness and average color give me small boost. \n\nOn the other hand, I found advanced ones (like probability predicted by Resnet) make my model worse, even though they have a correlation with deal probability higher than some important features indicated by my model.", "votes": null }, { "id": "343977", "postDate": "06/16/2018 16:27:32", "content": "May ask you about what kinds of model you use get such improvements in NN or LGBM？And how you get them，use pretrained model or opencv
", "rawMarkdown": "May ask you about what kinds of model you use get such improvements in NN or LGBM？And how you get them，use pretrained model or opencv", "votes": null }, { "id": "344059", "postDate": "06/16/2018 21:03:41", "content": "As already said ..simple image features may be helpful for LGBM but not for NN.
\n\nSome deeper images features are definitely helpful. for NN
", "rawMarkdown": "As already said ..simple image features may be helpful for LGBM but not for NN. \n\nSome deeper images features are definitely helpful. for NN", "votes": null }, { "id": "344870", "postDate": "06/18/2018 20:22:47", "content": "How are missing images in test set handled?
", "rawMarkdown": "How are missing images in test set handled?", "votes": null }, { "id": "345529", "postDate": "06/20/2018 02:15:28", "content": "The answer is yes as others have said. Image features gave me about 0.0012 improvement on my LGBM model and a 0.0003 with Catboost which uses fewer features. I have not run my XGB model with image features yet but expect it to make a difference there as well.
", "rawMarkdown": "The answer is yes as others have said. Image features gave me about 0.0012 improvement on my LGBM model and a 0.0003 with Catboost which uses fewer features. I have not run my XGB model with image features yet but expect it to make a difference there as well.", "votes": null }, { "id": "346326", "postDate": "06/21/2018 14:11:40", "content": "If you have features like \"blurriness\" with values between, let's say 0 and 100, just use a -1 for rows with missing images.
", "rawMarkdown": "If you have features like \"blurriness\" with values between, let's say 0 and 100, just use a -1 for rows with missing images.", "votes": null }, { "id": "346376", "postDate": "06/21/2018 15:32:50", "content": "I also have a similar case where Resnet etc. Class and Probability make my LGB model worse. Not sure why.
", "rawMarkdown": "I also have a similar case where Resnet etc. Class and Probability make my LGB model worse. Not sure why.", "votes": null }, { "id": "346378", "postDate": "06/21/2018 15:34:37", "content": "By features is it meta features ? Pretrained keras predictions or some dense activation values at the end of one of those pretrained models ? 0.0012 is quite a lot !
", "rawMarkdown": "By features is it meta features ? Pretrained keras predictions or some dense activation values at the end of one of those pretrained models ? 0.0012 is quite a lot !", "votes": null }, { "id": "346466", "postDate": "06/21/2018 18:20:19", "content": "the basic ones like blur definitely help. we have also come up with similar/improved image features based on https://www.kaggle.com/shivamb/ideas-for-image-features-and-image-quality. but with my limited experience I failed to develop an pure NN based feature extraction method.
", "rawMarkdown": "the basic ones like blur definitely help. we have also come up with similar/improved image features based on https://www.kaggle.com/shivamb/ideas-for-image-features-and-image-quality. but with my limited experience I failed to develop an pure NN based feature extraction method.", "votes": null }, { "id": "346478", "postDate": "06/21/2018 18:49:38", "content": "Both and both
", "rawMarkdown": "Both and both", "votes": null }, { "id": "346479", "postDate": "06/21/2018 18:50:07", "content": "Though the model's reactions to some of the image features can be a bit strange...
\n\n
Thanks @Arnaud Roussel. Yes it is a very good improvement. I have those features all along but had an error in my code which I found after all the +ve feedback about image features I have read here and in other threads. Once I fixed that I got the benefit.
\n\nI asked the same question you are asking me and all people were willing to say is simple image features or nothing complicated. I will be honest and say, I will only share that information after the end of the competition if anyone wants to know.
\n\nHappy Kaggling!
", "rawMarkdown": "Thanks @Arnaud Roussel. Yes it is a very good improvement. I have those features all along but had an error in my code which I found after all the +ve feedback about image features I have read here and in other threads. Once I fixed that I got the benefit.\n\nI asked the same question you are asking me and all people were willing to say is simple image features or nothing complicated. I will be honest and say, I will only share that information after the end of the competition if anyone wants to know.\n\nHappy Kaggling!", "votes": null }, { "id": "346623", "postDate": "06/22/2018 04:11:50", "content": "Found a way to make the predicted imagenet labels working. Guess your answer inspired me to look again :) Got a 0.0008 improvement in a first draft on a test GBM model for features. Need now to retrain a full GBM.
\n\nThanks !
", "rawMarkdown": "Found a way to make the predicted imagenet labels working. Guess your answer inspired me to look again :) Got a 0.0008 improvement in a first draft on a test GBM model for features. Need now to retrain a full GBM.\n\nThanks !", "votes": null }, { "id": "346637", "postDate": "06/22/2018 04:37:50", "content": "Thanks @Arnaud Roussel, I am glad you found a way. An upvote would be nice if you got inspired by my comment :-)
\n\nI also got inspired by the kind of non-answer answers I got to look hard at my code and eventually found the bug.
\n\nGood luck.
", "rawMarkdown": "Thanks @Arnaud Roussel, I am glad you found a way. An upvote would be nice if you got inspired by my comment :-)\n\nI also got inspired by the kind of non-answer answers I got to look hard at my code and eventually found the bug.\n\nGood luck.", "votes": null }, { "id": "347241", "postDate": "06/23/2018 17:32:56", "content": "Yes, it improved score for me:
\n\nThey help - at least for us. Not the simple ones like blurriness or average color but \"deeper\" ones.
", "votes": null, "replies": [ { "id": 342982, "author_name": "naivelamb", "author_url": "", "post_date": "06/14/2018 13:31:19", "content": "Interesting, I'm a newbie and I have the contradictory case. Simple ones like blurriness and average color give me small boost.
\n\nOn the other hand, I found advanced ones (like probability predicted by Resnet) make my model worse, even though they have a correlation with deal probability higher than some important features indicated by my model.
", "votes": null, "replies": [] }, { "id": 346376, "author_name": "arroqc", "author_url": "", "post_date": "06/21/2018 15:32:50", "content": "I also have a similar case where Resnet etc. Class and Probability make my LGB model worse. Not sure why.
", "votes": null, "replies": [] } ] }, { "id": 342274, "author_name": "nuhsikander", "author_url": "", "post_date": "06/13/2018 08:05:09", "content": "Yep, I've got a boost in both XGB and LGB with simpler image features.
", "votes": null, "replies": [] }, { "id": 342372, "author_name": "liuhdsgoal", "author_url": "", "post_date": "06/13/2018 11:57:31", "content": "helping a lot
", "votes": null, "replies": [] }, { "id": 342469, "author_name": "peterhurford", "author_url": "", "post_date": "06/13/2018 15:06:50", "content": "I've got at least a 0.0012 boost from image features.
", "votes": null, "replies": [ { "id": 343977, "author_name": "sheboke93", "author_url": "", "post_date": "06/16/2018 16:27:32", "content": "May ask you about what kinds of model you use get such improvements in NN or LGBM？And how you get them，use pretrained model or opencv
", "votes": null, "replies": [] }, { "id": 346478, "author_name": "peterhurford", "author_url": "", "post_date": "06/21/2018 18:49:38", "content": "Both and both
", "votes": null, "replies": [] } ] }, { "id": 344059, "author_name": "serigne", "author_url": "", "post_date": "06/16/2018 21:03:41", "content": "As already said ..simple image features may be helpful for LGBM but not for NN.
\n\nSome deeper images features are definitely helpful. for NN
", "votes": null, "replies": [] }, { "id": 344870, "author_name": "rajeshbhat", "author_url": "", "post_date": "06/18/2018 20:22:47", "content": "How are missing images in test set handled?
", "votes": null, "replies": [ { "id": 346326, "author_name": "frankherfert", "author_url": "", "post_date": "06/21/2018 14:11:40", "content": "If you have features like \"blurriness\" with values between, let's say 0 and 100, just use a -1 for rows with missing images.
", "votes": null, "replies": [] } ] }, { "id": 345529, "author_name": "sheriytm", "author_url": "", "post_date": "06/20/2018 02:15:28", "content": "The answer is yes as others have said. Image features gave me about 0.0012 improvement on my LGBM model and a 0.0003 with Catboost which uses fewer features. I have not run my XGB model with image features yet but expect it to make a difference there as well.
", "votes": null, "replies": [ { "id": 346378, "author_name": "arroqc", "author_url": "", "post_date": "06/21/2018 15:34:37", "content": "By features is it meta features ? Pretrained keras predictions or some dense activation values at the end of one of those pretrained models ? 0.0012 is quite a lot !
", "votes": null, "replies": [] }, { "id": 346527, "author_name": "sheriytm", "author_url": "", "post_date": "06/21/2018 22:08:10", "content": "Thanks @Arnaud Roussel. Yes it is a very good improvement. I have those features all along but had an error in my code which I found after all the +ve feedback about image features I have read here and in other threads. Once I fixed that I got the benefit.
\n\nI asked the same question you are asking me and all people were willing to say is simple image features or nothing complicated. I will be honest and say, I will only share that information after the end of the competition if anyone wants to know.
\n\nHappy Kaggling!
", "votes": null, "replies": [] }, { "id": 346623, "author_name": "arroqc", "author_url": "", "post_date": "06/22/2018 04:11:50", "content": "Found a way to make the predicted imagenet labels working. Guess your answer inspired me to look again :) Got a 0.0008 improvement in a first draft on a test GBM model for features. Need now to retrain a full GBM.
\n\nThanks !
", "votes": null, "replies": [] }, { "id": 346637, "author_name": "sheriytm", "author_url": "", "post_date": "06/22/2018 04:37:50", "content": "Thanks @Arnaud Roussel, I am glad you found a way. An upvote would be nice if you got inspired by my comment :-)
\n\nI also got inspired by the kind of non-answer answers I got to look hard at my code and eventually found the bug.
\n\nGood luck.
", "votes": null, "replies": [] } ] }, { "id": 346466, "author_name": "yl1202", "author_url": "", "post_date": "06/21/2018 18:20:19", "content": "the basic ones like blur definitely help. we have also come up with similar/improved image features based on https://www.kaggle.com/shivamb/ideas-for-image-features-and-image-quality. but with my limited experience I failed to develop an pure NN based feature extraction method.
", "votes": null, "replies": [] }, { "id": 346479, "author_name": "peterhurford", "author_url": "", "post_date": "06/21/2018 18:50:07", "content": "Though the model's reactions to some of the image features can be a bit strange...
\n\n
Yes, it improved score for me:
\n\n