![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2069843%2F6044b68dc9fe77cd71c54eabaf26eb8f%2Fdogs2.JPG?generation=1564312775721322&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2069843%2F147e2d0dfcb1e60c5bcfc471c6dd35c0%2Fdogs3.JPG?generation=1564313154469645&alt=media\")
I cannot understand why do these spots occur in the generated images.\nI also wanted to know, whether it will affect the score since, except those spots the images of dogs as a whole seem fine.
\n\nOn the other hand, I noticed that if the loss of generator reduces extremely like- 1e - 8, these spots start covering a major part of the images and become very prominent, which is quite a contradiction, since the quality should have improved, that's why I wanted to figure out the reason.
", "messages": [ { "id": "585941", "postDate": "07/28/2019 11:23:00", "content": "
I cannot understand why do these spots occur in the generated images.\nI also wanted to know, whether it will affect the score since, except those spots the images of dogs as a whole seem fine.
\n\nOn the other hand, I noticed that if the loss of generator reduces extremely like- 1e - 8, these spots start covering a major part of the images and become very prominent, which is quite a contradiction, since the quality should have improved, that's why I wanted to figure out the reason.
", "rawMarkdown": "\n\n\n\n\nI cannot understand why do these spots occur in the generated images.\nI also wanted to know, whether it will affect the score since, except those spots the images of dogs as a whole seem fine.\n\nOn the other hand, I noticed that if the loss of generator reduces extremely like- 1e - 8, these spots start covering a major part of the images and become very prominent, which is quite a contradiction, since the quality should have improved, that's why I wanted to figure out the reason.", "votes": null }, { "id": "585978", "postDate": "07/28/2019 12:33:22", "content": "I usually get these spots when I have overtrained (a lot of epochs) in memorizing GANs. \nTry reducing the number of epochs.
", "rawMarkdown": "I usually get these spots when I have overtrained (a lot of epochs) in memorizing GANs. \nTry reducing the number of epochs.", "votes": null }, { "id": "588595", "postDate": "07/30/2019 19:22:11", "content": "yeah, it seems probable. Thanks
", "rawMarkdown": "yeah, it seems probable. Thanks", "votes": null }, { "id": "589480", "postDate": "08/01/2019 00:28:51", "content": "They may also occur if your using an output activation of linear and training to match pixel values of 0-255 with loss MSE. If the activation outputs more than 255 or less than 0, it produces a spot.
", "rawMarkdown": "They may also occur if your using an output activation of linear and training to match pixel values of 0-255 with loss MSE. If the activation outputs more than 255 or less than 0, it produces a spot.", "votes": null } ], "comments": [ { "id": 585978, "author_name": "timetraveller98", "author_url": "", "post_date": "07/28/2019 12:33:22", "content": "I usually get these spots when I have overtrained (a lot of epochs) in memorizing GANs. \nTry reducing the number of epochs.
", "votes": null, "replies": [ { "id": 588595, "author_name": "tenzinbhotia", "author_url": "", "post_date": "07/30/2019 19:22:11", "content": "yeah, it seems probable. Thanks
", "votes": null, "replies": [] } ] }, { "id": 589480, "author_name": "cdeotte", "author_url": "", "post_date": "08/01/2019 00:28:51", "content": "They may also occur if your using an output activation of linear and training to match pixel values of 0-255 with loss MSE. If the activation outputs more than 255 or less than 0, it produces a spot.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "585941": "\n\n\n\n\nI cannot understand why do these spots occur in the generated images.\nI also wanted to know, whether it will affect the score since, except those spots the images of dogs as a whole seem fine.\n\nOn the other hand, I noticed that if the loss of generator reduces extremely like- 1e - 8, these spots start covering a major part of the images and become very prominent, which is quite a contradiction, since the quality should have improved, that's why I wanted to figure out the reason.", "585978": "I usually get these spots when I have overtrained (a lot of epochs) in memorizing GANs. \nTry reducing the number of epochs.", "588595": "yeah, it seems probable. Thanks", "589480": "They may also occur if your using an output activation of linear and training to match pixel values of 0-255 with loss MSE. If the activation outputs more than 255 or less than 0, it produces a spot." }, "source": "meta" }