what do u think better to train on?\ncolored pictures or black and white?
\n\nif u do black and white u loose information. \nbut in the other hand the test set not supposed to be like the train so maybe loosing information is good.
\n\ni might be test it later and tell u what i saw. \nbut if someone did it i will like to hear
", "messages": [ { "id": "276389", "postDate": "01/31/2018 11:50:06", "content": "what do u think better to train on?\ncolored pictures or black and white?
\n\nif u do black and white u loose information. \nbut in the other hand the test set not supposed to be like the train so maybe loosing information is good.
\n\ni might be test it later and tell u what i saw. \nbut if someone did it i will like to hear
", "rawMarkdown": "what do u think better to train on?\ncolored pictures or black and white?\n\nif u do black and white u loose information. \nbut in the other hand the test set not supposed to be like the train so maybe loosing information is good. \n\ni might be test it later and tell u what i saw. \nbut if someone did it i will like to hear", "votes": null }, { "id": "276510", "postDate": "01/31/2018 16:39:24", "content": "Just making the images black-and-white probably isn't enough. See the following kernel for additional ideas: https://www.kaggle.com/kmader/normalizing-brightfield-stained-and-fluorescence
", "rawMarkdown": "Just making the images black-and-white probably isn't enough. See the following kernel for additional ideas: https://www.kaggle.com/kmader/normalizing-brightfield-stained-and-fluorescence", "votes": null }, { "id": "276831", "postDate": "02/01/2018 16:05:56", "content": "I don’t know. However, I’m extremely interested in working on this! It might resolve a number of the possible generalization problems I and others have encountered (e.g. model robustness for both fluorescence and histological images) without any additional augmentation or architectural complexity. However, I’m weary for the obvious reasons (e.g. throwing out data).
", "rawMarkdown": "I don’t know. However, I’m extremely interested in working on this! It might resolve a number of the possible generalization problems I and others have encountered (e.g. model robustness for both fluorescence and histological images) without any additional augmentation or architectural complexity. However, I’m weary for the obvious reasons (e.g. throwing out data).", "votes": null }, { "id": "276857", "postDate": "02/01/2018 17:25:20", "content": "train separate models for the different cases. ensemble the results
", "rawMarkdown": "train separate models for the different cases. ensemble the results", "votes": null }, { "id": "276867", "postDate": "02/01/2018 17:51:20", "content": "It should also be possible to make the model learn this kind of transform, for example by starting out with one or more 1x1 convolutions.
", "rawMarkdown": "It should also be possible to make the model learn this kind of transform, for example by starting out with one or more 1x1 convolutions.", "votes": null }, { "id": "276874", "postDate": "02/01/2018 18:04:59", "content": "Yeah. I suppose you could get even fancier and train a VAE to learn some ideal augmentation.
", "rawMarkdown": "Yeah. I suppose you could get even fancier and train a VAE to learn some ideal augmentation.", "votes": null }, { "id": "276962", "postDate": "02/01/2018 22:40:06", "content": "i dont think so because the train set very small and the cnn will get good results using the color but maybe in the test it will fail
", "rawMarkdown": "i dont think so because the train set very small and the cnn will get good results using the color but maybe in the test it will fail", "votes": null }, { "id": "294772", "postDate": "03/12/2018 16:09:09", "content": "Hi @Human Analog, \nI'm not sure that 1x1 conv could help in this case. \nThe 2 main (alternative) approaches to deal with the 'different' stage-2 test data should be:
\n\nbase an information reduction (actually throwing out some info) on a priori knowledge. this works if you know that you are throwing out 'variance' (for example if you know that a b&w image of a nucleus is a good enough generalization).
make the network learn more general knowledge via heavy augmentation
Using 1x1 conv will make the network learn a good 'compact' (for example if your channels go from 3 to 1) representation; of course good for the training set, and not transferable to the stage-2 data.\nThinking better about it: if you run 1x1 conv (as a very first layer) on a heavily augmented set, you could learn a compact (which also implies regularized) representation. Have you tried it? I think I will if time will allow (and in this case I'll let you know).
", "rawMarkdown": "Hi @Human Analog, \nI'm not sure that 1x1 conv could help in this case. \nThe 2 main (alternative) approaches to deal with the 'different' stage-2 test data should be:\n\n1. base an information reduction (actually throwing out some info) on a priori knowledge. this works if you know that you are throwing out 'variance' (for example if you know that a b&w image of a nucleus is a good enough generalization). \n\n2. make the network learn more general knowledge via heavy augmentation\n\nUsing 1x1 conv will make the network learn a good 'compact' (for example if your channels go from 3 to 1) representation; of course good for the training set, and not transferable to the stage-2 data.\nThinking better about it: if you run 1x1 conv (as a very first layer) on a heavily augmented set, you could learn a compact (which also implies regularized) representation. Have you tried it? I think I will if time will allow (and in this case I'll let you know).", "votes": null }, { "id": "294780", "postDate": "03/12/2018 16:25:43", "content": "What I meant was really using a 1x1 convolution to go from 3 channels to many more, such as 10 channels, and then another 1x1 convolution back to 3 channels. This technique is described in the paper Systematic evaluation of CNN advances on the ImageNet, Mishkin et al. (2016).
", "rawMarkdown": "What I meant was really using a 1x1 convolution to go from 3 channels to many more, such as 10 channels, and then another 1x1 convolution back to 3 channels. This technique is described in the paper *Systematic evaluation of CNN advances on the ImageNet*, Mishkin et al. (2016).", "votes": null }, { "id": "294792", "postDate": "03/12/2018 16:42:43", "content": "Thanks, it is a very interesting paper!!! (for now, I've just read the par. 3.4 related to this topic). I would say (but I should test it) that in principle this can learn a good colourspace for the training set. But again, I'm intrigued about applying it to a heavily augmented (especially with colour-related transforms) set. As a regularization and (hopefully) generalization measure I would try 1x1x10+1x1x1 (instead of 1x1x10+1x1x3). Thanks again for sharing!!
", "rawMarkdown": "Thanks, it is a very interesting paper!!! (for now, I've just read the par. 3.4 related to this topic). I would say (but I should test it) that in principle this can learn a good colourspace for the training set. But again, I'm intrigued about applying it to a heavily augmented (especially with colour-related transforms) set. As a regularization and (hopefully) generalization measure I would try 1x1x10+1x1x1 (instead of 1x1x10+1x1x3). Thanks again for sharing!!", "votes": null }, { "id": "304734", "postDate": "03/28/2018 00:34:49", "content": "Have no idea what will be good for test2.
", "rawMarkdown": "Have no idea what will be good for test2.", "votes": null } ], "comments": [ { "id": 276510, "author_name": "humananalog", "author_url": "", "post_date": "01/31/2018 16:39:24", "content": "Just making the images black-and-white probably isn't enough. See the following kernel for additional ideas: https://www.kaggle.com/kmader/normalizing-brightfield-stained-and-fluorescence
", "votes": null, "replies": [] }, { "id": 276831, "author_name": "agoodman", "author_url": "", "post_date": "02/01/2018 16:05:56", "content": "I don’t know. However, I’m extremely interested in working on this! It might resolve a number of the possible generalization problems I and others have encountered (e.g. model robustness for both fluorescence and histological images) without any additional augmentation or architectural complexity. However, I’m weary for the obvious reasons (e.g. throwing out data).
", "votes": null, "replies": [] }, { "id": 276857, "author_name": "hengck23", "author_url": "", "post_date": "02/01/2018 17:25:20", "content": "train separate models for the different cases. ensemble the results
", "votes": null, "replies": [] }, { "id": 276867, "author_name": "humananalog", "author_url": "", "post_date": "02/01/2018 17:51:20", "content": "It should also be possible to make the model learn this kind of transform, for example by starting out with one or more 1x1 convolutions.
", "votes": null, "replies": [ { "id": 276874, "author_name": "agoodman", "author_url": "", "post_date": "02/01/2018 18:04:59", "content": "Yeah. I suppose you could get even fancier and train a VAE to learn some ideal augmentation.
", "votes": null, "replies": [] }, { "id": 276962, "author_name": "aviwolfson", "author_url": "", "post_date": "02/01/2018 22:40:06", "content": "i dont think so because the train set very small and the cnn will get good results using the color but maybe in the test it will fail
", "votes": null, "replies": [] }, { "id": 294772, "author_name": "andrearapuzzi", "author_url": "", "post_date": "03/12/2018 16:09:09", "content": "Hi @Human Analog, \nI'm not sure that 1x1 conv could help in this case. \nThe 2 main (alternative) approaches to deal with the 'different' stage-2 test data should be:
\n\nbase an information reduction (actually throwing out some info) on a priori knowledge. this works if you know that you are throwing out 'variance' (for example if you know that a b&w image of a nucleus is a good enough generalization).
make the network learn more general knowledge via heavy augmentation
Using 1x1 conv will make the network learn a good 'compact' (for example if your channels go from 3 to 1) representation; of course good for the training set, and not transferable to the stage-2 data.\nThinking better about it: if you run 1x1 conv (as a very first layer) on a heavily augmented set, you could learn a compact (which also implies regularized) representation. Have you tried it? I think I will if time will allow (and in this case I'll let you know).
", "votes": null, "replies": [] }, { "id": 294780, "author_name": "humananalog", "author_url": "", "post_date": "03/12/2018 16:25:43", "content": "What I meant was really using a 1x1 convolution to go from 3 channels to many more, such as 10 channels, and then another 1x1 convolution back to 3 channels. This technique is described in the paper Systematic evaluation of CNN advances on the ImageNet, Mishkin et al. (2016).
", "votes": null, "replies": [] }, { "id": 294792, "author_name": "andrearapuzzi", "author_url": "", "post_date": "03/12/2018 16:42:43", "content": "Thanks, it is a very interesting paper!!! (for now, I've just read the par. 3.4 related to this topic). I would say (but I should test it) that in principle this can learn a good colourspace for the training set. But again, I'm intrigued about applying it to a heavily augmented (especially with colour-related transforms) set. As a regularization and (hopefully) generalization measure I would try 1x1x10+1x1x1 (instead of 1x1x10+1x1x3). Thanks again for sharing!!
", "votes": null, "replies": [] } ] }, { "id": 304734, "author_name": "wayne999", "author_url": "", "post_date": "03/28/2018 00:34:49", "content": "Have no idea what will be good for test2.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "276389": "what do u think better to train on?\ncolored pictures or black and white?\n\nif u do black and white u loose information. \nbut in the other hand the test set not supposed to be like the train so maybe loosing information is good. \n\ni might be test it later and tell u what i saw. \nbut if someone did it i will like to hear", "276510": "Just making the images black-and-white probably isn't enough. See the following kernel for additional ideas: https://www.kaggle.com/kmader/normalizing-brightfield-stained-and-fluorescence", "276831": "I don’t know. However, I’m extremely interested in working on this! It might resolve a number of the possible generalization problems I and others have encountered (e.g. model robustness for both fluorescence and histological images) without any additional augmentation or architectural complexity. However, I’m weary for the obvious reasons (e.g. throwing out data).", "276857": "train separate models for the different cases. ensemble the results", "276867": "It should also be possible to make the model learn this kind of transform, for example by starting out with one or more 1x1 convolutions.", "276874": "Yeah. I suppose you could get even fancier and train a VAE to learn some ideal augmentation.", "276962": "i dont think so because the train set very small and the cnn will get good results using the color but maybe in the test it will fail", "294772": "Hi @Human Analog, \nI'm not sure that 1x1 conv could help in this case. \nThe 2 main (alternative) approaches to deal with the 'different' stage-2 test data should be:\n\n1. base an information reduction (actually throwing out some info) on a priori knowledge. this works if you know that you are throwing out 'variance' (for example if you know that a b&w image of a nucleus is a good enough generalization). \n\n2. make the network learn more general knowledge via heavy augmentation\n\nUsing 1x1 conv will make the network learn a good 'compact' (for example if your channels go from 3 to 1) representation; of course good for the training set, and not transferable to the stage-2 data.\nThinking better about it: if you run 1x1 conv (as a very first layer) on a heavily augmented set, you could learn a compact (which also implies regularized) representation. Have you tried it? I think I will if time will allow (and in this case I'll let you know).", "294780": "What I meant was really using a 1x1 convolution to go from 3 channels to many more, such as 10 channels, and then another 1x1 convolution back to 3 channels. This technique is described in the paper *Systematic evaluation of CNN advances on the ImageNet*, Mishkin et al. (2016).", "294792": "Thanks, it is a very interesting paper!!! (for now, I've just read the par. 3.4 related to this topic). I would say (but I should test it) that in principle this can learn a good colourspace for the training set. But again, I'm intrigued about applying it to a heavily augmented (especially with colour-related transforms) set. As a regularization and (hopefully) generalization measure I would try 1x1x10+1x1x1 (instead of 1x1x10+1x1x3). Thanks again for sharing!!", "304734": "Have no idea what will be good for test2." }, "source": "meta" }