![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3982638%2Fafc88db124a5069d35654a5af222ad21%2FScreenshot%202020-07-29%20at%203.15.51%20PM.png?generation=1596016300535534&alt=media\")
My team is trying to use GAN to generate more melanoma images, but before proceeding further, i want to know if it is allowed to use GAN in this competition.
\n\nHere is a glimpse of few images generated using GAN:\n
@jwebermsk please reply.
", "messages": [ { "id": 950257, "postDate": "2020-07-29T09:53:42.363Z", "content": "My team is trying to use GAN to generate more melanoma images, but before proceeding further, i want to know if it is allowed to use GAN in this competition.
\n\nHere is a glimpse of few images generated using GAN:\n
@jwebermsk please reply.
", "rawMarkdown": "My team is trying to use GAN to generate more melanoma images, but before proceeding further, i want to know if it is allowed to use GAN in this competition. \n\nHere is a glimpse of few images generated using GAN:\n\n\n@jwebermsk please reply.", "votes": 11 }, { "id": 950298, "postDate": "2020-07-29T10:24:05.157Z", "content": "@rohitsingh9990 As long as you are generating them from the competition dataset or approved external datasets, you can use the generated images in the competition.
\n\nPS: There is no requirement about sharing generated images. You only need to post the origin datasource links in the External data thread.
", "rawMarkdown": "@rohitsingh9990 As long as you are **generating** them from the competition dataset or approved external datasets, you can use the generated images in the competition.\n\nPS: There is no requirement about sharing **generated** images. You only need to post the origin datasource links in the External data thread.", "votes": 3, "replies": [ { "id": 950303, "postDate": "2020-07-29T10:27:11.587Z", "content": "@sirishks thanks, you are right. But i will wait for organizers to confirm this.
", "rawMarkdown": "@sirishks thanks, you are right. But i will wait for organizers to confirm this." } ] }, { "id": 950946, "postDate": "2020-07-29T18:42:15.407Z", "content": "In the external data thread, Kaggle writes
\n\n\n\n\nYou may use generated images, as long as they are not generated from the test set, as that effectively constitutes hand-labeling.
\n
Can we generate images without labels for use with unsupervised learning? I'm confused how images without labels can be considered \"hand-labeled\"??
", "rawMarkdown": "In the external data thread, Kaggle writes\n> You may use generated images, as long as they are not generated from the test set, as that effectively constitutes hand-labeling.\n\nCan we generate images **without** labels for use with unsupervised learning? I'm confused how images **without** labels can be considered \"hand-labeled\"??", "votes": 4, "replies": [ { "id": 952905, "postDate": "2020-07-31T09:58:57.503Z", "content": "interesting, I want to remind that Porto Seguro competition was won by making autoencoder for combined train-test sets
", "rawMarkdown": "interesting, I want to remind that Porto Seguro competition was won by making autoencoder for combined train-test sets" } ] }, { "id": 950860, "postDate": "2020-07-29T17:16:15.280Z", "content": "Yes, generated images are OK to use, as long as they are generated from the training set or other external data and not the test set.
", "rawMarkdown": "Yes, generated images are OK to use, as long as they are generated from the training set or other external data and **not the test set**.", "votes": 4, "replies": [ { "id": 950892, "postDate": "2020-07-29T17:36:25.600Z", "content": "Thanks .
", "rawMarkdown": "Thanks . " }, { "id": 950930, "postDate": "2020-07-29T18:25:24.133Z", "content": "> Yes, generated images are OK to use, as long as they are generated from the training set or other external data and not the test set.
\n\nThis is a confusing ruling @juliaelliott . Is it allowed to pseudo label the test data and use it to train on?
\n\nIf so, what is the difference between using pseudo labeled test data with subsequent data augmentation (rotation, zoom, cutout, etc) versus generating new train data from pseudo labeled data?
", "rawMarkdown": "> Yes, generated images are OK to use, as long as they are generated from the training set or other external data and not the test set.\n\nThis is a confusing ruling @juliaelliott . Is it allowed to pseudo label the test data and use it to train on? \n\nIf so, what is the difference between using pseudo labeled test data with subsequent data augmentation (rotation, zoom, cutout, etc) versus generating new train data from pseudo labeled data?", "votes": 5 }, { "id": 950940, "postDate": "2020-07-29T18:30:35.360Z", "content": "@cdeotte yes , this rule is pretty confusing about pseudolabling . It is not really handlabeling . It is also possible to generate near test image data (not exact data ) via GAN to reduce train /test image dissimilarity if any . I believe you already had a similar if not same point updated in this discussion if I am right . \nhttps://www.kaggle.com/c/flower-classification-with-tpus/discussion/151476
", "rawMarkdown": "@cdeotte yes , this rule is pretty confusing about pseudolabling . It is not really handlabeling . It is also possible to generate near test image data (not exact data ) via GAN to reduce train /test image dissimilarity if any . I believe you already had a similar if not same point updated in this discussion if I am right . \nhttps://www.kaggle.com/c/flower-classification-with-tpus/discussion/151476" }, { "id": 950952, "postDate": "2020-07-29T18:53:40Z", "content": "Yes, the same confusion is happening here. There are 4 unanswered questions:
\n\nSorry for the confusion. The bottom line is that you cannot hand-label the test set or cleverly-disguise hand-labeling of the test set behind pseudolabeling, generative or other augmentation techniques. What I'm trying to specify is that you cannot use generative images or other methods in an attempt to circumvent the prohibition against hand-labeling the test set. So yes, pseudolabeling and generative images are permitted, but hand-labeling of the test set is not.
", "rawMarkdown": "Sorry for the confusion. The bottom line is that you cannot hand-label the test set or cleverly-disguise hand-labeling of the test set behind pseudolabeling, generative or other augmentation techniques. What I'm trying to specify is that you cannot use generative images or other methods in an attempt to circumvent the prohibition against hand-labeling the test set. So yes, pseudolabeling and generative images are permitted, but hand-labeling of the test set is not.", "votes": 4 }, { "id": 950987, "postDate": "2020-07-29T19:22:41.157Z", "content": "@juliaelliott Good answer Julia. I understand your point. We cannot hand select test images that we think are malignant, then apply a GAN to them, then label the results malignant, then train. That is effectively hand labeling test images and is disallowed.
\n\nNor apply a GAN to all test images then hand label the results. Again we would effectively be hand labeling test data. Because that is like rotating all test images 90 degrees. Then hand labeling those saying \"these aren't test images anymore\" because we would effectively be hand labeling test images.
\n\netc etc etc
", "rawMarkdown": "@juliaelliott Good answer Julia. I understand your point. We cannot hand select test images that we think are malignant, then apply a GAN to them, then label the results malignant, then train. That is effectively hand labeling test images and is disallowed. \n\nNor apply a GAN to all test images then hand label the results. Again we would effectively be hand labeling test data. Because that is like rotating all test images 90 degrees. Then hand labeling those saying \"these aren't test images anymore\" because we would effectively be hand labeling test images.\n\netc etc etc", "votes": 6 }, { "id": 951020, "postDate": "2020-07-29T20:06:29.073Z", "content": "@cdeotte Exactly! :) Appreciate your pushing me to clarify.
", "rawMarkdown": "@cdeotte Exactly! :) Appreciate your pushing me to clarify.", "votes": 5 } ] }, { "id": 972998, "postDate": "2020-08-17T03:05:14.413Z", "content": "I have also tried similar approach but did not work out. I will be happy to discuss about usability of GAN generated images after the competition ends :)
\nHere is how the artificially generated images look like.
Good work. Are you also generating non-melanoma images using GAN? If we generate only the melanoma images, the model might overfit on the generated images and become less general.
", "rawMarkdown": "Good work. Are you also generating non-melanoma images using GAN? If we generate only the melanoma images, the model might overfit on the generated images and become less general." }, { "id": 952829, "postDate": "2020-07-31T08:55:39.293Z", "content": "Hey @rohitsingh9990 , what kind of GAN are you using. I am generating positive only images with DCGAN but images are not as clear as yours. I ran for 1000 epochs, BS 64. Can you elaborate a little on your GAN setup?
\n\nThanks
", "rawMarkdown": "Hey @rohitsingh9990 , what kind of GAN are you using. I am generating positive only images with DCGAN but images are not as clear as yours. I ran for 1000 epochs, BS 64. Can you elaborate a little on your GAN setup? \n\nThanks", "replies": [ { "id": 952882, "postDate": "2020-07-31T09:34:45.303Z", "content": "@pheadrus I have used (SPGAN) for above image creation. Sorry i can't reveal more than this before the competition ends.
", "rawMarkdown": "@pheadrus I have used (SPGAN) for above image creation. Sorry i can't reveal more than this before the competition ends.", "votes": 1 }, { "id": 952984, "postDate": "2020-07-31T11:58:00.640Z", "content": "Fair enough @rohitsingh9990 . Thanks for the pointer.
", "rawMarkdown": "Fair enough @rohitsingh9990 . Thanks for the pointer. " } ] }, { "id": 951398, "postDate": "2020-07-30T05:58:42.973Z", "content": "I think, it should be allowed because GAN is a kind of augmentation technique. In competition rule, I did not find any algorithm related restrictions or limitations. Although it is a huge work. But very nice idea for this competition.
", "rawMarkdown": "I think, it should be allowed because GAN is a kind of augmentation technique. In competition rule, I did not find any algorithm related restrictions or limitations. Although it is a huge work. But very nice idea for this competition." }, { "id": 950844, "postDate": "2020-07-29T17:03:32.427Z", "content": "tagging @juliaelliott .
\n\nWe are generating the GAN images in the approved dataset as per already posted common competition and external data .
", "rawMarkdown": "tagging @juliaelliott .\n\nWe are generating the GAN images in the approved dataset as per already posted common competition and external data ." }, { "id": 950284, "postDate": "2020-07-29T10:13:16.773Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 950289, "postDate": "2020-07-29T10:18:51.077Z", "content": "I am not sure about that, that's why i am asking organizers regarding this.
", "rawMarkdown": "I am not sure about that, that's why i am asking organizers regarding this.", "votes": 1 }, { "id": 950297, "postDate": "2020-07-29T10:23:40.340Z", "rawMarkdown": "", "isDeleted": true }, { "id": 950312, "postDate": "2020-07-29T10:32:03.790Z", "rawMarkdown": "", "isDeleted": true }, { "id": 950326, "postDate": "2020-07-29T10:40:19.787Z", "content": "@synked thanks for your advice, i have just posted it into the external data forum as well.
", "rawMarkdown": "@synked thanks for your advice, i have just posted it into the external data forum as well." } ] } ], "comments": [ { "id": 950298, "author_name": "Sirish Somanchi", "author_url": "", "post_date": "2020-07-29T10:24:05.157000", "content": "@rohitsingh9990 As long as you are generating them from the competition dataset or approved external datasets, you can use the generated images in the competition.
\n\nPS: There is no requirement about sharing generated images. You only need to post the origin datasource links in the External data thread.
", "votes": 3, "replies": [ { "id": 950303, "author_name": "Dracarys", "author_url": "", "post_date": "2020-07-29T10:27:11.587000", "content": "@sirishks thanks, you are right. But i will wait for organizers to confirm this.
", "votes": 0, "replies": [] } ] }, { "id": 950946, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-07-29T18:42:15.407000", "content": "In the external data thread, Kaggle writes
\n\n\n\n\nYou may use generated images, as long as they are not generated from the test set, as that effectively constitutes hand-labeling.
\n
Can we generate images without labels for use with unsupervised learning? I'm confused how images without labels can be considered \"hand-labeled\"??
", "votes": 4, "replies": [ { "id": 952905, "author_name": "Jacek Poplawski", "author_url": "", "post_date": "2020-07-31T09:58:57.503000", "content": "interesting, I want to remind that Porto Seguro competition was won by making autoencoder for combined train-test sets
", "votes": 0, "replies": [] } ] }, { "id": 950860, "author_name": "Julia Elliott", "author_url": "", "post_date": "2020-07-29T17:16:15.280000", "content": "Yes, generated images are OK to use, as long as they are generated from the training set or other external data and not the test set.
", "votes": 4, "replies": [ { "id": 950892, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2020-07-29T17:36:25.600000", "content": "Thanks .
", "votes": 0, "replies": [] }, { "id": 950930, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-07-29T18:25:24.133000", "content": "> Yes, generated images are OK to use, as long as they are generated from the training set or other external data and not the test set.
\n\nThis is a confusing ruling @juliaelliott . Is it allowed to pseudo label the test data and use it to train on?
\n\nIf so, what is the difference between using pseudo labeled test data with subsequent data augmentation (rotation, zoom, cutout, etc) versus generating new train data from pseudo labeled data?
", "votes": 5, "replies": [] }, { "id": 950940, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2020-07-29T18:30:35.360000", "content": "@cdeotte yes , this rule is pretty confusing about pseudolabling . It is not really handlabeling . It is also possible to generate near test image data (not exact data ) via GAN to reduce train /test image dissimilarity if any . I believe you already had a similar if not same point updated in this discussion if I am right . \nhttps://www.kaggle.com/c/flower-classification-with-tpus/discussion/151476
", "votes": 0, "replies": [] }, { "id": 950952, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-07-29T18:53:40", "content": "Yes, the same confusion is happening here. There are 4 unanswered questions:
\n\nSorry for the confusion. The bottom line is that you cannot hand-label the test set or cleverly-disguise hand-labeling of the test set behind pseudolabeling, generative or other augmentation techniques. What I'm trying to specify is that you cannot use generative images or other methods in an attempt to circumvent the prohibition against hand-labeling the test set. So yes, pseudolabeling and generative images are permitted, but hand-labeling of the test set is not.
", "votes": 4, "replies": [] }, { "id": 950987, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-07-29T19:22:41.157000", "content": "@juliaelliott Good answer Julia. I understand your point. We cannot hand select test images that we think are malignant, then apply a GAN to them, then label the results malignant, then train. That is effectively hand labeling test images and is disallowed.
\n\nNor apply a GAN to all test images then hand label the results. Again we would effectively be hand labeling test data. Because that is like rotating all test images 90 degrees. Then hand labeling those saying \"these aren't test images anymore\" because we would effectively be hand labeling test images.
\n\netc etc etc
", "votes": 6, "replies": [] }, { "id": 951020, "author_name": "Julia Elliott", "author_url": "", "post_date": "2020-07-29T20:06:29.073000", "content": "@cdeotte Exactly! :) Appreciate your pushing me to clarify.
", "votes": 5, "replies": [] } ] }, { "id": 972998, "author_name": "Ganbold", "author_url": "", "post_date": "2020-08-17T03:05:14.413000", "content": "I have also tried similar approach but did not work out. I will be happy to discuss about usability of GAN generated images after the competition ends :)
\nHere is how the artificially generated images look like.
Good work. Are you also generating non-melanoma images using GAN? If we generate only the melanoma images, the model might overfit on the generated images and become less general.
", "votes": 0, "replies": [] }, { "id": 952829, "author_name": "Phaedrus", "author_url": "", "post_date": "2020-07-31T08:55:39.293000", "content": "Hey @rohitsingh9990 , what kind of GAN are you using. I am generating positive only images with DCGAN but images are not as clear as yours. I ran for 1000 epochs, BS 64. Can you elaborate a little on your GAN setup?
\n\nThanks
", "votes": 0, "replies": [ { "id": 952882, "author_name": "Dracarys", "author_url": "", "post_date": "2020-07-31T09:34:45.303000", "content": "@pheadrus I have used (SPGAN) for above image creation. Sorry i can't reveal more than this before the competition ends.
", "votes": 1, "replies": [] }, { "id": 952984, "author_name": "Phaedrus", "author_url": "", "post_date": "2020-07-31T11:58:00.640000", "content": "Fair enough @rohitsingh9990 . Thanks for the pointer.
", "votes": 0, "replies": [] } ] }, { "id": 951398, "author_name": "Partha Pratim Banik", "author_url": "", "post_date": "2020-07-30T05:58:42.973000", "content": "I think, it should be allowed because GAN is a kind of augmentation technique. In competition rule, I did not find any algorithm related restrictions or limitations. Although it is a huge work. But very nice idea for this competition.
", "votes": 0, "replies": [] }, { "id": 950844, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2020-07-29T17:03:32.427000", "content": "tagging @juliaelliott .
\n\nWe are generating the GAN images in the approved dataset as per already posted common competition and external data .
", "votes": 0, "replies": [] }, { "id": 950284, "author_name": "", "author_url": "", "post_date": "2020-07-29T10:13:16.773000", "content": "", "votes": 1, "replies": [ { "id": 950289, "author_name": "Dracarys", "author_url": "", "post_date": "2020-07-29T10:18:51.077000", "content": "I am not sure about that, that's why i am asking organizers regarding this.
", "votes": 1, "replies": [] }, { "id": 950297, "author_name": "", "author_url": "", "post_date": "2020-07-29T10:23:40.340000", "content": "", "votes": 0, "replies": [] }, { "id": 950312, "author_name": "", "author_url": "", "post_date": "2020-07-29T10:32:03.790000", "content": "", "votes": 0, "replies": [] }, { "id": 950326, "author_name": "Dracarys", "author_url": "", "post_date": "2020-07-29T10:40:19.787000", "content": "@synked thanks for your advice, i have just posted it into the external data forum as well.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "950257": "My team is trying to use GAN to generate more melanoma images, but before proceeding further, i want to know if it is allowed to use GAN in this competition. \n\nHere is a glimpse of few images generated using GAN:\n\n\n@jwebermsk please reply.", "950298": "@rohitsingh9990 As long as you are **generating** them from the competition dataset or approved external datasets, you can use the generated images in the competition.\n\nPS: There is no requirement about sharing **generated** images. You only need to post the origin datasource links in the External data thread.", "950946": "In the external data thread, Kaggle writes\n> You may use generated images, as long as they are not generated from the test set, as that effectively constitutes hand-labeling.\n\nCan we generate images **without** labels for use with unsupervised learning? I'm confused how images **without** labels can be considered \"hand-labeled\"??", "950860": "Yes, generated images are OK to use, as long as they are generated from the training set or other external data and **not the test set**.", "972998": "I have also tried similar approach but did not work out. I will be happy to discuss about usability of GAN generated images after the competition ends :) \n[Here](https://www.kaggle.com/chimdee/generatedmalignantimages) is how the artificially generated images look like.", "958507": "Good work. Are you also generating non-melanoma images using GAN? If we generate only the melanoma images, the model might overfit on the generated images and become less general.", "952829": "Hey @rohitsingh9990 , what kind of GAN are you using. I am generating positive only images with DCGAN but images are not as clear as yours. I ran for 1000 epochs, BS 64. Can you elaborate a little on your GAN setup? \n\nThanks", "951398": "I think, it should be allowed because GAN is a kind of augmentation technique. In competition rule, I did not find any algorithm related restrictions or limitations. Although it is a huge work. But very nice idea for this competition.", "950844": "tagging @juliaelliott .\n\nWe are generating the GAN images in the approved dataset as per already posted common competition and external data .", "950284": "" } }