![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F936560%2F98fc7abd3d8f5e98e776cc4a3d3386c9%2F__results___14_179.png?generation=1562684407430056&alt=media\")
This is my GAN training Loss (Discriminator and Generator) lets share it all in order to have references of how is the training going :D
", "messages": [ { "id": "571402", "postDate": "07/09/2019 15:01:22", "content": "
This is my GAN training Loss (Discriminator and Generator) lets share it all in order to have references of how is the training going :D
", "rawMarkdown": "\n\nThis is my GAN training Loss (Discriminator and Generator) lets share it all in order to have references of how is the training going :D", "votes": null }, { "id": "571679", "postDate": "07/10/2019 01:09:32", "content": "G loss is hard... but I think that's excessive (bad) what GAN are you using?
", "rawMarkdown": "G loss is hard... but I think that's excessive (bad) what GAN are you using?", "votes": null }, { "id": "571772", "postDate": "07/10/2019 04:44:17", "content": "Te adjunto un ejemplo mio :)\nI share with you one example LB = 100...
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2779868%2F816fc82d80299b9ba09095612aa36fd0%2Ftrain_losses.png?generation=1562733798351851&alt=media\")
Wow that seems really nice, but the Generator loss should decrease along the epochs intead of increase right?
", "rawMarkdown": "Wow that seems really nice, but the Generator loss should decrease along the epochs intead of increase right?", "votes": null }, { "id": "571854", "postDate": "07/10/2019 06:46:35", "content": "This is not necessary, because every time the discriminator gets better, it becomes harder for the generator to deceive him. Both losses can fluctuate but for discriminator loss it's crucial to stay above 0 from what I've heard.
", "rawMarkdown": "This is not necessary, because every time the discriminator gets better, it becomes harder for the generator to deceive him. Both losses can fluctuate but for discriminator loss it's crucial to stay above 0 from what I've heard.", "votes": null }, { "id": "571871", "postDate": "07/10/2019 07:11:18", "content": "Exactly π it's from a simple kernel that I'm going to publish today (later)
", "rawMarkdown": "Exactly π it's from a simple kernel that I'm going to publish today (later)", "votes": null }, { "id": "571943", "postDate": "07/10/2019 08:58:59", "content": "Ok i understand, maybe my optimizers are not correctly tunned, which lr did you use for Gen/Disc?
", "rawMarkdown": "Ok i understand, maybe my optimizers are not correctly tunned, which lr did you use for Gen/Disc?", "votes": null }, { "id": "572213", "postDate": "07/10/2019 16:10:20", "content": "Between 5e-5 and 5e-4,but as I said it depends
", "rawMarkdown": "Between 5e-5 and 5e-4,but as I said it depends", "votes": null }, { "id": "572224", "postDate": "07/10/2019 16:30:34", "content": "Ok i will wait for your kerne :), i did some changes in my code and they are working well i will share my new training loss when i finished
", "rawMarkdown": "Ok i will wait for your kerne :), i did some changes in my code and they are working well i will share my new training loss when i finished", "votes": null }, { "id": "572690", "postDate": "07/11/2019 09:18:33", "content": "βοΈ δ½ εΎεΈ
", "rawMarkdown": "βοΈ δ½ εΎεΈ ", "votes": null }, { "id": "575291", "postDate": "07/15/2019 08:45:24", "content": "Nice
", "rawMarkdown": "Nice", "votes": null }, { "id": "577994", "postDate": "07/17/2019 09:00:08", "content": "I have a desire to expand into Fourier series\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2003084%2Fb6f5b97cd201b220b51f6e40bd163465%2F2019-07-17%2011.58.10.jpg?generation=1563353917428975&alt=media\")
Wow it looks really nice, my losses now are over 1.5 in Generator and 0.5 in Discriminator, how did you reach those good results?
", "rawMarkdown": "Wow it looks really nice, my losses now are over 1.5 in Generator and 0.5 in Discriminator, how did you reach those good results?", "votes": null }, { "id": "578178", "postDate": "07/17/2019 12:44:21", "content": "In fact, there is no good result. dogs are generated below. Just wanted to show beautiful loss graph :)\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2003084%2F179b1dc993b178133072910dcdefc978%2F2019-07-17%2015.37.34.jpg?generation=1563367091116317&alt=media\")
π π π π π
", "rawMarkdown": "π π π π π", "votes": null }, { "id": "3168442", "postDate": "04/02/2025 12:44:07", "content": "I am training my GAN for a synthetic multi-modal 2d gaussian dataset. I have kept the Discriminator losses to be negative of the relevant terms in the value function (so that the terms are maximized) and the Generator loss to be same as the relevant terms in the value function (so that Generator loss is minimized). The negative-Disc loss is getting minimized (as expected) but the Generator loss increases over epochs and then converges. However, the output from the trained model is pretty good (in the sense that visually, generator distribution is close to real-data distribution). Am I missing something or is such a behavior expected.
\nP.S. : I am relatively new to the topic.
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9882198%2F94cf4f71b742d7b4b3f58862e894090f%2FScreenshot%202025-04-02%20180815.png?generation=1743597836553453&alt=media\")
I am training my GAN for a synthetic multi-modal 2d gaussian dataset. I have kept the Discriminator losses to be negative of the relevant terms in the value function (so that the terms are maximized) and the Generator loss to be same as the relevant terms in the value function (so that Generator loss is minimized). The negative-Disc loss is getting minimized (as expected) but the Generator loss increases over epochs and then converges. However, the output from the trained model is pretty good (in the sense that visually, generator distribution is close to real-data distribution). Am I missing something or is such a behavior expected.
\nP.S. : I am relatively new to the topic.
G loss is hard... but I think that's excessive (bad) what GAN are you using?
", "votes": null, "replies": [] }, { "id": 571772, "author_name": "jesucristo", "author_url": "", "post_date": "07/10/2019 04:44:17", "content": "Te adjunto un ejemplo mio :)\nI share with you one example LB = 100...
\n\n
Wow that seems really nice, but the Generator loss should decrease along the epochs intead of increase right?
", "votes": null, "replies": [] }, { "id": 571854, "author_name": "dvorobiev", "author_url": "", "post_date": "07/10/2019 06:46:35", "content": "This is not necessary, because every time the discriminator gets better, it becomes harder for the generator to deceive him. Both losses can fluctuate but for discriminator loss it's crucial to stay above 0 from what I've heard.
", "votes": null, "replies": [] }, { "id": 571871, "author_name": "jesucristo", "author_url": "", "post_date": "07/10/2019 07:11:18", "content": "Exactly π it's from a simple kernel that I'm going to publish today (later)
", "votes": null, "replies": [] }, { "id": 571943, "author_name": "enric1296", "author_url": "", "post_date": "07/10/2019 08:58:59", "content": "Ok i understand, maybe my optimizers are not correctly tunned, which lr did you use for Gen/Disc?
", "votes": null, "replies": [] }, { "id": 572213, "author_name": "jesucristo", "author_url": "", "post_date": "07/10/2019 16:10:20", "content": "Between 5e-5 and 5e-4,but as I said it depends
", "votes": null, "replies": [] }, { "id": 572224, "author_name": "enric1296", "author_url": "", "post_date": "07/10/2019 16:30:34", "content": "Ok i will wait for your kerne :), i did some changes in my code and they are working well i will share my new training loss when i finished
", "votes": null, "replies": [] } ] }, { "id": 572690, "author_name": "saluzi", "author_url": "", "post_date": "07/11/2019 09:18:33", "content": "βοΈ δ½ εΎεΈ
", "votes": null, "replies": [] }, { "id": 575291, "author_name": "", "author_url": "", "post_date": "07/15/2019 08:45:24", "content": "Nice
", "votes": null, "replies": [] }, { "id": 577994, "author_name": "speedwagon", "author_url": "", "post_date": "07/17/2019 09:00:08", "content": "I have a desire to expand into Fourier series\n
Wow it looks really nice, my losses now are over 1.5 in Generator and 0.5 in Discriminator, how did you reach those good results?
", "votes": null, "replies": [] }, { "id": 578178, "author_name": "speedwagon", "author_url": "", "post_date": "07/17/2019 12:44:21", "content": "In fact, there is no good result. dogs are generated below. Just wanted to show beautiful loss graph :)\n
π π π π π
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "571402": "\n\nThis is my GAN training Loss (Discriminator and Generator) lets share it all in order to have references of how is the training going :D", "571679": "G loss is hard... but I think that's excessive (bad) what GAN are you using?", "571772": "Te adjunto un ejemplo mio :)\nI share with you one example LB = 100...\n\n", "571832": "Wow that seems really nice, but the Generator loss should decrease along the epochs intead of increase right?", "571854": "This is not necessary, because every time the discriminator gets better, it becomes harder for the generator to deceive him. Both losses can fluctuate but for discriminator loss it's crucial to stay above 0 from what I've heard.", "571871": "Exactly π it's from a simple kernel that I'm going to publish today (later)", "571943": "Ok i understand, maybe my optimizers are not correctly tunned, which lr did you use for Gen/Disc?", "572213": "Between 5e-5 and 5e-4,but as I said it depends", "572224": "Ok i will wait for your kerne :), i did some changes in my code and they are working well i will share my new training loss when i finished", "572690": "βοΈ δ½ εΎεΈ ", "575291": "Nice", "577994": "I have a desire to expand into Fourier series\n", "578163": "Wow it looks really nice, my losses now are over 1.5 in Generator and 0.5 in Discriminator, how did you reach those good results?", "578178": "In fact, there is no good result. dogs are generated below. Just wanted to show beautiful loss graph :)\n", "578780": "π π π π π", "3168442": "I am training my GAN for a synthetic multi-modal 2d gaussian dataset. I have kept the Discriminator losses to be negative of the relevant terms in the value function (so that the terms are maximized) and the Generator loss to be same as the relevant terms in the value function (so that Generator loss is minimized). The negative-Disc loss is getting minimized (as expected) but the Generator loss increases over epochs and then converges. However, the output from the trained model is pretty good (in the sense that visually, generator distribution is close to real-data distribution). Am I missing something or is such a behavior expected.\nP.S. : I am relatively new to the topic.\n\n" }, "source": "meta" }