![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F90bb6472cbaacc2b61110fa4cda45b58%2F__results___11_0.png?generation=1563874006996360&alt=media\")
I have upsampled my input images with equal number of original images (20579 ) . I had used 20579 images + 20579 horizontally flipped images. It should have learned better right ?\nBut not only it messed up my GAN (same GAN performed well with original set of images) but also it started creating something like two faced dogs . Does that mean its memorizing ? How else it will try to create same dog with two faces attached ? Or am i seeing things ?\n
I have upsampled my input images with equal number of original images (20579 ) . I had used 20579 images + 20579 horizontally flipped images. It should have learned better right ?\nBut not only it messed up my GAN (same GAN performed well with original set of images) but also it started creating something like two faced dogs . Does that mean its memorizing ? How else it will try to create same dog with two faces attached ? Or am i seeing things ?\n
Yeah, sometimes i got this too :) The generator definitely remembers a lot of spatial patterns including preprocessing-specific ones
", "rawMarkdown": "Yeah, sometimes i got this too :) The generator definitely remembers a lot of spatial patterns including preprocessing-specific ones", "votes": null }, { "id": "582539", "postDate": "07/23/2019 09:50:40", "content": "I had a similar experience. When I was using random rotate with black bars, the generator was producing “rotated images” with the same black bars. \nI was thinking to try preprocessing with all dogs facing the same direction and no flip.
", "rawMarkdown": "I had a similar experience. When I was using random rotate with black bars, the generator was producing “rotated images” with the same black bars. \nI was thinking to try preprocessing with all dogs facing the same direction and no flip.", "votes": null }, { "id": "582566", "postDate": "07/23/2019 10:30:40", "content": "I have been wondering since I started participating in this competition: What transformations will help my GAN?\nI am still unsure. Rotation has always reduced my score. Horizontal flip ... has almost always increased my score although I don't keep both versions of the image like you (normal and flipped ones) I only keep one.
", "rawMarkdown": "I have been wondering since I started participating in this competition: What transformations will help my GAN?\nI am still unsure. Rotation has always reduced my score. Horizontal flip ... has almost always increased my score although I don't keep both versions of the image like you (normal and flipped ones) I only keep one.", "votes": null }, { "id": "582734", "postDate": "07/23/2019 14:08:59", "content": "I didn't find any improvement in the score after my image data augmentation. Now does that mean it tries to memorize the images?
", "rawMarkdown": "I didn't find any improvement in the score after my image data augmentation. Now does that mean it tries to memorize the images?", "votes": null } ], "comments": [ { "id": 582529, "author_name": "dvorobiev", "author_url": "", "post_date": "07/23/2019 09:32:32", "content": "Yeah, sometimes i got this too :) The generator definitely remembers a lot of spatial patterns including preprocessing-specific ones
", "votes": null, "replies": [] }, { "id": 582539, "author_name": "robga", "author_url": "", "post_date": "07/23/2019 09:50:40", "content": "I had a similar experience. When I was using random rotate with black bars, the generator was producing “rotated images” with the same black bars. \nI was thinking to try preprocessing with all dogs facing the same direction and no flip.
", "votes": null, "replies": [] }, { "id": 582566, "author_name": "timetraveller98", "author_url": "", "post_date": "07/23/2019 10:30:40", "content": "I have been wondering since I started participating in this competition: What transformations will help my GAN?\nI am still unsure. Rotation has always reduced my score. Horizontal flip ... has almost always increased my score although I don't keep both versions of the image like you (normal and flipped ones) I only keep one.
", "votes": null, "replies": [] }, { "id": 582734, "author_name": "sriharipramod", "author_url": "", "post_date": "07/23/2019 14:08:59", "content": "I didn't find any improvement in the score after my image data augmentation. Now does that mean it tries to memorize the images?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "582525": "I have upsampled my input images with equal number of original images (20579 ) . I had used 20579 images + 20579 horizontally flipped images. It should have learned better right ?\nBut not only it messed up my GAN (same GAN performed well with original set of images) but also it started creating something like two faced dogs . Does that mean its memorizing ? How else it will try to create same dog with two faces attached ? Or am i seeing things ?\n", "582529": "Yeah, sometimes i got this too :) The generator definitely remembers a lot of spatial patterns including preprocessing-specific ones", "582539": "I had a similar experience. When I was using random rotate with black bars, the generator was producing “rotated images” with the same black bars. \nI was thinking to try preprocessing with all dogs facing the same direction and no flip.", "582566": "I have been wondering since I started participating in this competition: What transformations will help my GAN?\nI am still unsure. Rotation has always reduced my score. Horizontal flip ... has almost always increased my score although I don't keep both versions of the image like you (normal and flipped ones) I only keep one.", "582734": "I didn't find any improvement in the score after my image data augmentation. Now does that mean it tries to memorize the images?" }, "source": "meta" }