Please share your generated dog pictures here!
", "messages": [ { "id": 563834, "postDate": "2019-06-28T18:14:13.283Z", "content": "Please share your generated dog pictures here!
", "rawMarkdown": "Please share your generated dog pictures here! ", "votes": 47 }, { "id": 577850, "postDate": "2019-07-17T06:16:59.420Z", "content": "My top dogs for now :D\nThere are a lot of noisy patterns in the images, but the score on public LB is pretty high
\n\n.png?generation=1563344100098383&alt=media)
Because of the confusions in the Leaderboard I am using a new metrics to calculate my position in Leaderboard and progression for past 5-6 days . The metric is Relativistic Dmitry Distance . I calculate my LB position based on your scores :D . Anyway , what is the FID for this dogs ?
", "rawMarkdown": "Because of the confusions in the Leaderboard I am using a new metrics to calculate my position in Leaderboard and progression for past 5-6 days . The metric is Relativistic Dmitry Distance . I calculate my LB position based on your scores :D . Anyway , what is the FID for this dogs ?", "votes": 3 }, { "id": 577906, "postDate": "2019-07-17T07:24:35.533Z", "content": "29.20708 on public LB :)\nI think it can be improved further if I find a way to smooth out these textures.
", "rawMarkdown": "29.20708 on public LB :)\nI think it can be improved further if I find a way to smooth out these textures.", "votes": 3 }, { "id": 577938, "postDate": "2019-07-17T07:59:36.943Z", "content": "Awesome! You are definitely wining this competition regardless of the final LB.
", "rawMarkdown": "Awesome! You are definitely wining this competition regardless of the final LB." }, { "id": 577949, "postDate": "2019-07-17T08:18:51.573Z", "content": "Holy cow those look good. Any tips you have for us? :)\nFor smoothing the images, using an upsampling + conv somewhere can help idk
", "rawMarkdown": "Holy cow those look good. Any tips you have for us? :)\nFor smoothing the images, using an upsampling + conv somewhere can help idk", "votes": 4 }, { "id": 577959, "postDate": "2019-07-17T08:30:25.237Z", "content": "Is it a fully generative GAN?
", "rawMarkdown": "Is it a fully generative GAN?" }, { "id": 577976, "postDate": "2019-07-17T08:44:53.480Z", "content": "@theoviel \nI use one of the recent GAN architectures and upsampling is already used there. I have another similar arch, which give me nicer textures, but it's worse at capturing the shapes.\nThe tip is to stabilize the generator weights after a certain point.
\n\n@francoisdubois \nyes, it's an ordinary GAN
", "rawMarkdown": "@theoviel \nI use one of the recent GAN architectures and upsampling is already used there. I have another similar arch, which give me nicer textures, but it's worse at capturing the shapes.\nThe tip is to stabilize the generator weights after a certain point.\n\n@francoisdubois \nyes, it's an ordinary GAN", "votes": 8 }, { "id": 577998, "postDate": "2019-07-17T09:05:29.357Z", "content": "@dvorobiev Thanks, I'll have a look at that
", "rawMarkdown": "@dvorobiev Thanks, I'll have a look at that" }, { "id": 578116, "postDate": "2019-07-17T11:08:05.723Z", "content": "Wow, incredible job Dmitry.
", "rawMarkdown": "Wow, incredible job Dmitry. ", "votes": 2 }, { "id": 578682, "postDate": "2019-07-18T02:50:00.673Z", "content": "Great job!
", "rawMarkdown": "Great job!", "votes": 1 }, { "id": 579162, "postDate": "2019-07-18T14:48:03.190Z", "content": "How many epoches?
", "rawMarkdown": "How many epoches?" }, { "id": 579653, "postDate": "2019-07-19T04:19:05.883Z", "content": "@grvgtm \n150
", "rawMarkdown": "@grvgtm \n150", "votes": 4 }, { "id": 580557, "postDate": "2019-07-20T11:00:59.890Z", "content": "@dvorobiev what is weight stabilize? is it adding Gaussian Noise or weight clipping?
", "rawMarkdown": "@dvorobiev what is weight stabilize? is it adding Gaussian Noise or weight clipping? " }, { "id": 580577, "postDate": "2019-07-20T12:03:32.103Z", "content": "@skyap79 \nNo, the idea was to add some sort of momentum. Like what Adam and other optimizers do to learning rate, but for weights.
", "rawMarkdown": "@skyap79 \nNo, the idea was to add some sort of momentum. Like what Adam and other optimizers do to learning rate, but for weights.", "votes": 4 }, { "id": 581796, "postDate": "2019-07-22T12:07:38.120Z", "content": "@dvorobiev Have you manually did some data cleaning for the training?
", "rawMarkdown": "@dvorobiev Have you manually did some data cleaning for the training? " }, { "id": 581838, "postDate": "2019-07-22T13:12:11.577Z", "content": "@rhonin0 \nNo, I don't use any manually cleaning or image selection, I am too lazy for it :) I used only basic preprocessing to generate these pictures. However recently I found that some automated image selection helps a bit.
", "rawMarkdown": "@rhonin0 \nNo, I don't use any manually cleaning or image selection, I am too lazy for it :) I used only basic preprocessing to generate these pictures. However recently I found that some automated image selection helps a bit.", "votes": 2 }, { "id": 581880, "postDate": "2019-07-22T13:46:21.740Z", "content": "Impressive results Dmitry!. Was stuck in scores around 60 for some days. But now, trying different architectures like you suggested.\nHowever, those generally are very deep for the time frame βΉοΈ
", "rawMarkdown": "Impressive results Dmitry!. Was stuck in scores around 60 for some days. But now, trying different architectures like you suggested.\nHowever, those generally are very deep for the time frame βΉοΈ " }, { "id": 581908, "postDate": "2019-07-22T14:21:11.953Z", "content": "Thanks!\nYeah, a lot of recent GANs aims at producing higher resolution images. Usually some parameter reduction or layer removal is necessary to make it work under 9 hours constraint.
", "rawMarkdown": "Thanks!\nYeah, a lot of recent GANs aims at producing higher resolution images. Usually some parameter reduction or layer removal is necessary to make it work under 9 hours constraint.", "votes": 2 }, { "id": 582397, "postDate": "2019-07-23T06:17:01.517Z", "content": "great job Dmitry :) I wonder what your generator and discriminator loss graph looks like..
", "rawMarkdown": "great job Dmitry :) I wonder what your generator and discriminator loss graph looks like.." }, { "id": 582402, "postDate": "2019-07-23T06:29:25.883Z", "content": "I haven't tried to plot it yet, but usually the generator loss increases over time, while the discriminator loss stays about the same.
", "rawMarkdown": "I haven't tried to plot it yet, but usually the generator loss increases over time, while the discriminator loss stays about the same." }, { "id": 582409, "postDate": "2019-07-23T06:38:09.467Z", "content": "I see. can't wait too see your solution :(
", "rawMarkdown": "I see. can't wait too see your solution :(", "votes": 1 }, { "id": 582598, "postDate": "2019-07-23T11:05:14.797Z", "content": "wow ! they looks so realistic π
", "rawMarkdown": "wow ! they looks so realistic π " } ] }, { "id": 593675, "postDate": "2019-08-07T01:07:41.110Z", "content": "My dog ββalso has a face in his buttocks. \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1626653%2F2b296fe94872660146d90b80684f2ad0%2F.PNG?generation=1565140095808450&alt=media\")
yeah, it seams so very much...
", "rawMarkdown": "yeah, it seams so very much...", "votes": 2 }, { "id": 597748, "postDate": "2019-08-12T18:05:39.523Z", "content": "lol π
", "rawMarkdown": "lol π " } ] }, { "id": 579144, "postDate": "2019-07-18T14:33:54.677Z", "content": "My generator unexpectedly generated a green dragon like Rayquaza...\n
", "rawMarkdown": "My generator unexpectedly generated a green dragon like Rayquaza...\n![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1755928%2F483614867bc0ebf7dc8d4b12a2574e9f%2FmessageImage_1563349604482.jpg?generation=1563460320173694&alt=media\")
![](\"https://cdn.bulbagarden.net/upload/e/e4/384Rayquaza.png\")
\n",
"votes": 13,
"replies": [
{
"id": 579155,
"postDate": "2019-07-18T14:40:23.137Z",
"content": "Hahaha :)
", "rawMarkdown": "Hahaha :)", "votes": 2 } ] }, { "id": 569049, "postDate": "2019-07-06T00:34:18.630Z", "content": "Here is my GAN results.\nSome of them look like cute puppies! However, aliens also seem to be generated ...
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1256949%2Faeb954563ebcc51236cad54ad3c556f2%2Fout.png?generation=1562372958556305&alt=media\")
Updated: with some tweaks, MiFID is now 39.42284 (not memorized)\nI also feel that the quality of dog images has been even better, but it doesn't reflect that much on the MiFID score....maybe MiFID is not a fair metric for visual quality of non-memorized dogs?
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2Fe14d56d785ce39b75497e91a432c8cee%2F20190814_v6_w1.0_lr_v4_1.png?generation=1565798065604999&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2F84997495dc2790b01439dab936ba6d27%2F20190814_v6_w1.0_lr_v4_5.png?generation=1565797354101553&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2F7c2980dcb70899eaf537cbd836bfd2be%2F20190814_v6_w1.0_lr_v4_3.png?generation=1565797357572020&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2Fff6797270db20a7c014310466239d03f%2F20190814_v6_w1.0_lr_v4_4.png?generation=1565797355419907&alt=media\")
Previous: Customized GAN, MiFID: 43.07151 (not memorized)\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2Fcb3d2b48ff33e25538c1c6f4e914c891%2F20190810_v6_w1.0_3.png?generation=1565579403664805&alt=media\")
Nice work. Looks great
", "rawMarkdown": "Nice work. Looks great", "votes": 1 }, { "id": 597471, "postDate": "2019-08-12T11:42:43.120Z", "content": "IMO these pics should get way better score. \nAre you using some brightness adjusments?
", "rawMarkdown": "IMO these pics should get way better score. \nAre you using some brightness adjusments?", "votes": 3 }, { "id": 597475, "postDate": "2019-08-12T11:47:16.877Z", "content": "@cdeotte Thanks a lot for your inspiring kernels and posts! They really helped shorten the learning path of GANs.
", "rawMarkdown": "@cdeotte Thanks a lot for your inspiring kernels and posts! They really helped shorten the learning path of GANs.", "votes": 1 }, { "id": 597554, "postDate": "2019-08-12T13:50:54.903Z", "content": "@dvorobiev Thanks, I only used random horizontal flip with a low prob., no improvement with other transformations yet.
", "rawMarkdown": "@dvorobiev Thanks, I only used random horizontal flip with a low prob., no improvement with other transformations yet." }, { "id": 597559, "postDate": "2019-08-12T13:54:35.200Z", "content": "But why do your images' colors appear muted? The colors are not as bright as other people's GANs. I think Dmitry's point is that if you can fix the color of your outputted images your score would increase.
", "rawMarkdown": "But why do your images' colors appear muted? The colors are not as bright as other people's GANs. I think Dmitry's point is that if you can fix the color of your outputted images your score would increase.", "votes": 2 }, { "id": 597560, "postDate": "2019-08-12T13:55:24.810Z", "content": "amazing job @markpeng :) hope you can keep improving :)
", "rawMarkdown": "amazing job @markpeng :) hope you can keep improving :)", "votes": 2 }, { "id": 597574, "postDate": "2019-08-12T14:24:09.653Z", "content": "@cdeotte Good point, will take a look, I also noticed that some of the generated images are a bit dark. Thanks.
", "rawMarkdown": "@cdeotte Good point, will take a look, I also noticed that some of the generated images are a bit dark. Thanks." }, { "id": 597575, "postDate": "2019-08-12T14:25:07.180Z", "content": "@cdeotte \nyeah, better colors might give a better results. Maybe something wrong with de-normalization...
", "rawMarkdown": "@cdeotte \nyeah, better colors might give a better results. Maybe something wrong with de-normalization...", "votes": 2 }, { "id": 597613, "postDate": "2019-08-12T15:05:45.220Z", "content": "Thanks a lot! @dvorobiev @cdeotte, you are right, there is a bug in de-normalization, I will make a new submission to see how it goes after the fix.\nFrom local test MiFID only improved a bit to 43.01698, but the dogs are more colorful now.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2F21c177a4dd200ac8dd0cf3e839ced971%2F20190810_v6_w0.3_fixed_1.png?generation=1565622315309985&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F201774%2F3dfe256b544b9c75ce200b633f4bf010%2F20190810_v6_w1.0_fixed_2.png?generation=1565622317057179&alt=media\")
your images are looking much better than your score
", "rawMarkdown": "your images are looking much better than your score" }, { "id": 599176, "postDate": "2019-08-14T15:47:15.490Z", "content": "@kranthi9 Feel the same, wonder if MiFID would be a fair metric, it seems to give more advantages to dog diversity (cgan-like and acgan-like) than quality (Just updated the latest examples).
", "rawMarkdown": "@kranthi9 Feel the same, wonder if MiFID would be a fair metric, it seems to give more advantages to dog diversity (cgan-like and acgan-like) than quality (Just updated the latest examples).", "votes": 2 }, { "id": 599455, "postDate": "2019-08-15T01:45:50.780Z", "content": "Yes, diversity is more important than quality. You can see it in my Big GAN kernel here. Setting the random seed's truncation=0.5 generates higher quality images but lower LB. (The latent walk examples use t=0.5). Setting truncation=1.0 generates lower quality images, more diversity, and better LB. (Bottom images use t=1.0).
@cdeotte I also set t=1.0 for truncation trick in my final models. But local MiFID showed that 0.6 is the best. I noticed some variance between public LB and local score, the same notebook didn't finish with the same number of iterations in several runs (using fixed seed and deterministic mode of PyTorch). There seems to have some resource contention in the Kernel environment (I suppose P100 GPU should be dedicated per Kernel). So I finally chose 1.0 as its score is more consistent.
Btw, my GAN is a mini version of StyleGAN with progressive growing from 8x8 to 64x64, learning rate is incremental and fade-in phase size is higher in lower resolution. The mapping network is 8 128-dim FCs. Tried to extend it with conditional labels (CGAN-like) and auxiliary classifier (ACGAN-like), but the generated images could not converge and mostly out of control. For some reason, spectral normalization slowed StyleGAN discriminator down a lot. Also experimented other architectures but none of them has done better than StyleGAN (not yet tried BigGAN).
Will share more after private LB announcement.
", "rawMarkdown": "@cdeotte I also set `t=1.0` for truncation trick in my final models. But local MiFID showed that `0.6` is the best. I noticed some variance between public LB and local score, the same notebook didn't finish with the same number of iterations in several runs (using fixed seed and deterministic mode of PyTorch). There seems to have some resource contention in the Kernel environment (I suppose P100 GPU should be dedicated per Kernel). So I finally chose `1.0` as its score is more consistent.\n\nBtw, my GAN is a mini version of `StyleGAN` with progressive growing from `8x8` to `64x64`, learning rate is incremental and fade-in phase size is higher in lower resolution. The mapping network is 8 `128-dim` FCs. Tried to extend it with conditional labels (CGAN-like) and auxiliary classifier (ACGAN-like), but the generated images could not converge and mostly out of control. For some reason, spectral normalization slowed StyleGAN discriminator down a lot. Also experimented other architectures but none of them has done better than StyleGAN (not yet tried BigGAN).\n\nWill share more after private LB announcement.", "votes": 1 } ] }, { "id": 585472, "postDate": "2019-07-27T14:18:28.647Z", "content": "Making a dog generator, ending with a goat generator.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2990900%2F2b19a5c690773d9c65569ffcdea37d8d%2Fimage_00041.png?generation=1564236886938465&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2990900%2Fd694dc6353ec17575a9d4168b3258c0c%2Fgoat-compare.jpg?generation=1564236941319221&alt=media\")
My network started to kill the dogs that don't look good.
\n\n![\"bleeding](\"https://i.imgur.com/noSRQqi.png\")
Lol
", "rawMarkdown": "Lol " }, { "id": 591735, "postDate": "2019-08-04T07:04:03.627Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2236951%2Ff3fe75ac2a80368ba1bfb5111e950e04%2Fss.png?generation=1564902222703675&alt=media\")
\nsame here, GAN violence is spreading.
My cute cerberus\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1846168%2Ffd87700c081d6c86b585db63a19d12e6%2F2019-07-31%2018.29.31.png?generation=1564582550926789&alt=media\")
thats awesome clarity :) 22+ is one of the best MFiD with generalized GAN that I have heard from anyone in this comp.
", "rawMarkdown": "thats awesome clarity :) 22+ is one of the best MFiD with generalized GAN that I have heard from anyone in this comp.", "votes": 1 }, { "id": 591903, "postDate": "2019-08-04T12:56:23.893Z", "content": "I usually have something like this:
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2972861%2Fd7e818ffe3071fc0c6f1b8e257713036%2F1.PNG?generation=1564923278067727&alt=media\")
totally agree lol
", "rawMarkdown": "totally agree lol" }, { "id": 597179, "postDate": "2019-08-12T01:10:21.393Z", "content": "They might be an origin of kerberos π€
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F916454%2Ff34ca01eb71f1b59086c079d8c3a2059%2Fn02085620_3488.jpg?generation=1565572110799481&alt=media\")
Not a two-headed dog, but just two dogs taking a photo together.
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F694899%2Ff04b773c1c7dc4345fcdb6946a350003%2F9964.png?generation=1565794394726667&alt=media\")
my best score images generated by GAN\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1565399%2Fb816ed94f888e41a3139374af0e88074%2Fgenerated_images.png?generation=1561948835449815&alt=media\")
thank you so much @tikutiku , amazing images!
", "rawMarkdown": "thank you so much @tikutiku , amazing images!" } ] }, { "id": 574333, "postDate": "2019-07-13T16:32:23.407Z", "content": "Images from my Dog Breed CGAN here\n![\"image\"](\"http://playagricola.com/Kaggle/cgan171319.png\")
\n![\"image\"](\"http://playagricola.com/Kaggle/cgan271319.png\")
\n![\"image\"](\"http://playagricola.com/Kaggle/cgan371319.png\")
Amazing work Chris, I have played with CGANs quite a bit lately but did not get anything close to your results. I'll get through your code and hopefully find where I did a mistake !
", "rawMarkdown": "Amazing work Chris, I have played with CGANs quite a bit lately but did not get anything close to your results. I'll get through your code and hopefully find where I did a mistake !", "votes": 1 }, { "id": 574404, "postDate": "2019-07-13T19:16:17.930Z", "content": "Thanks Theo. It took me a while to get it to work too. CGANs have proven to be more difficult than basic GANs. I had to try different architectures, different regularization, and different D/G learning rates. With basic GANs it feels like everything just works. Right now my CGAN is scoring LB 100 but I feel it has potential to do much better.
", "rawMarkdown": "Thanks Theo. It took me a while to get it to work too. CGANs have proven to be more difficult than basic GANs. I had to try different architectures, different regularization, and different D/G learning rates. With basic GANs it feels like everything just works. Right now my CGAN is scoring LB 100 but I feel it has potential to do much better.", "votes": 1 }, { "id": 574412, "postDate": "2019-07-13T19:31:59.633Z", "content": "They are very tricky & painful to configure.\nA good usage of dog labels is indeed the key to beat the 60~ threshold I believe. Well at least I hope so.
", "rawMarkdown": "They are very tricky & painful to configure.\nA good usage of dog labels is indeed the key to beat the 60~ threshold I believe. Well at least I hope so.", "votes": 1 } ] }, { "id": 579184, "postDate": "2019-07-18T15:31:52.713Z", "content": "Our dogs reached 62.77, @galeone and me are working to improve our model in order to get better scores.\n(some dogs are still a bit scary: HP Lovecraft would have used them for some of his tales XD)
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F19298%2F7e59daea0a5ce09cb0f85f97a7b5a3ba%2F__results___14_15.png?generation=1563463848137093&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F19298%2F2a96a647d692b2ad292d020382f53376%2F__results___14_23.png?generation=1563463892816296&alt=media\")
Here is some of my results generated by vanilla GAN. They look very ugly. But the LB is 44.05636. I don't know what to say......\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F853445%2F5e544d82df5c2d40f05afcdb573ca4bc%2F_20190715011056.png?generation=1563124274987084&alt=media\")
Well, they look very much like dogs :) .. So the classifier picks it up properly and low LB score . But why are they little zoomed and patchy in color ?
", "rawMarkdown": "Well, they look very much like dogs :) .. So the classifier picks it up properly and low LB score . But why are they little zoomed and patchy in color ? " }, { "id": 574920, "postDate": "2019-07-14T17:35:27.850Z", "content": "@phoenix9032 I crop using the box annotations. I'm sure I have rightly denormed the generated samples, so the color should be ok. But I really think they are so ugly.
", "rawMarkdown": "@phoenix9032 I crop using the box annotations. I'm sure I have rightly denormed the generated samples, so the color should be ok. But I really think they are so ugly.", "votes": 1 }, { "id": 574996, "postDate": "2019-07-14T20:15:36.987Z", "content": "I had some weird colors as well recently.\nIf you are on PyTorch, it can be because you set your model on eval mode, which changes the behaviour of the batch size. Otherwise, make sure your batch size is big enough. Also, make sure your pictures are scaled between 0 and 1, but I believe you did this already because unscaled pictures score a lot worse.
\n\nEDIT : These tricks are for image generation, rescaling is done after generating images
", "rawMarkdown": "I had some weird colors as well recently.\nIf you are on PyTorch, it can be because you set your model on eval mode, which changes the behaviour of the batch size. Otherwise, make sure your batch size is big enough. Also, make sure your pictures are scaled between 0 and 1, but I believe you did this already because unscaled pictures score a lot worse.\n\nEDIT : These tricks are for image generation, rescaling is done after generating images", "votes": 2 }, { "id": 575059, "postDate": "2019-07-15T02:06:35.463Z", "content": "You mean batchsize while training the network right ?
", "rawMarkdown": "You mean batchsize while training the network right ? " }, { "id": 575254, "postDate": "2019-07-15T08:09:33.547Z", "content": "Nope when generating pictures.
", "rawMarkdown": "Nope when generating pictures." }, { "id": 575294, "postDate": "2019-07-15T08:53:32.207Z", "content": "> make sure your pictures are scaled between 0 and 1
\n\nI thought we are using tanh so it's better to scale images between [-1 1]\nwhy between [0,1]?
I meant for computing the metric, not mifid, as chris mentionned in a separated thread :)
", "rawMarkdown": "I meant for computing the metric, not mifid, as chris mentionned in a separated thread :) ", "votes": 1 }, { "id": 575355, "postDate": "2019-07-15T10:34:52.083Z", "content": "I got 68 while i tried MFID in my kernel privately .. Got so excited and run the same next time and submitted .. got 84 MFID . Such a heartbreak :(
", "rawMarkdown": "I got 68 while i tried MFID in my kernel privately .. Got so excited and run the same next time and submitted .. got 84 MFID . Such a heartbreak :(" } ] }, { "id": 574761, "postDate": "2019-07-14T12:51:59.670Z", "content": "Unfortunately I missed that \"kernel only\" and trained my customized styleGAN on my RTX 2080Ti, so I can't submit the 10k images for getting scores, anyway here are some results:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3458746%2F1f94b7e63d7d25b5601e62cc5278b969%2Ffakes004763.jpg?generation=1563108696681839&alt=media\")
Wow, they look great. Have you computed their MiFID? (with code here)
", "rawMarkdown": "Wow, they look great. Have you computed their MiFID? (with code [here][1])\n\n[1]: https://www.kaggle.com/wendykan/demo-mifid-metric-for-dog-image-generation-comp", "votes": 1 }, { "id": 574771, "postDate": "2019-07-14T13:17:09.717Z", "content": "Thanks a lot for the code, here is my result:\ndistance= 0.1981568680329205\nFID_public: 49.61624316188124 distance_public: 1 multiplied_public: 49.616243161880746
", "rawMarkdown": "Thanks a lot for the code, here is my result:\ndistance= 0.1981568680329205\nFID_public: 49.61624316188124 distance_public: 1 multiplied_public: 49.616243161880746\n", "votes": 2 }, { "id": 574788, "postDate": "2019-07-14T13:42:34.833Z", "content": "Great score. Are you cropping to the bounding boxes or training with the full images?
", "rawMarkdown": "Great score. Are you cropping to the bounding boxes or training with the full images?" }, { "id": 574800, "postDate": "2019-07-14T13:54:14.850Z", "content": "I also tried styleGAN but didn't achieve any meaningful results.\nHow much images do you use per each resolution phase?
", "rawMarkdown": "I also tried styleGAN but didn't achieve any meaningful results.\nHow much images do you use per each resolution phase?" }, { "id": 574829, "postDate": "2019-07-14T14:39:19.703Z", "content": "Nice doggos, how long did the training took though?
", "rawMarkdown": "Nice doggos, how long did the training took though?" }, { "id": 574833, "postDate": "2019-07-14T14:44:29.233Z", "content": "Just used all the data as it is. Did something similar not long ago with watches, that's why I basically just reused my existing code.
\n\n500.000 for each resolution. Training took ~20h on a RTX 2080 Ti.
", "rawMarkdown": "Just used all the data as it is. Did something similar not long ago with [watches](http://digital-thinking.de/watchgan-advancing-generated-watch-images-with-stylegans/), that's why I basically just reused my existing code.\n\n500.000 for each resolution. Training took ~20h on a RTX 2080 Ti.\n", "votes": 2 }, { "id": 574850, "postDate": "2019-07-14T15:07:09.660Z", "content": "What do you feed the StyleGAN to create an image? My understanding was you need to feed a real image in. If so, this wouldnβt be acceptable according to the rules.
\n\nOf course you could always use an output from a memoriser gan as input ..sarcasm..
", "rawMarkdown": "What do you feed the StyleGAN to create an image? My understanding was you need to feed a real image in. If so, this wouldnβt be acceptable according to the rules. \n\nOf course you could always use an output from a memoriser gan as input ..**sarcasm..**", "votes": 1 }, { "id": 574875, "postDate": "2019-07-14T15:55:17.173Z", "content": "You feed a vector, so no real image involved.
", "rawMarkdown": "You feed a vector, so no real image involved.", "votes": 1 } ] }, { "id": 566435, "postDate": "2019-07-02T06:57:36.790Z", "content": "Another update:\n![\"gan](\"https://i.imgur.com/JemlWb5.png\")
I have the feeling that they look a bit more like dogs now, but that could just be me.
", "rawMarkdown": "Another update:\n\n\nI have the feeling that they look a bit more like dogs now, but that could just be me. ", "votes": 7, "replies": [ { "id": 566447, "postDate": "2019-07-02T07:13:48.677Z", "content": "Wow, those look very good. I'd be interested in running kNN on some of those to see what the closest training image is. Are those completely synthesized or is your network reproducing training images? For example, bottom row fourth from right, dog looking at us.
", "rawMarkdown": "Wow, those look very good. I'd be interested in running kNN on some of those to see what the closest training image is. Are those completely synthesized or is your network reproducing training images? For example, bottom row fourth from right, dog looking at us." }, { "id": 566576, "postDate": "2019-07-02T10:39:31.697Z", "content": "I think they are \"completly\" synthesized -- if you could ever say that about a GAN.
\n\nThat's the output of the MiFID calculation (to be fair, from one model earlier):\n\n100%|ββββββββββ| 201/201 [00:36<00:00, 5.55it/s]\n100%|ββββββββββ| 412/412 [01:09<00:00, 4.27it/s]\nm1,m2 shape= ((2048,), (2048,)) s1,s2= ((2048, 2048), (2048, 2048))\nstarting calculating FID\ncovmean.shape= (2048, 2048)\ndone with FID, starting distance calculation\nd.shape= (10000, 20579)\nnp.min(d, axis=1).shape= (10000,)\ndistance= 0.2254516669552895\nFID_public: 83.36472131619365 distance_public: 1 multiplied_public: 83.36472131619281\n
To my understanding distance_public is the measurement for memorizing abilities.\nAs this is my first ever GAN experience, feel free to correct me where i am wrong.
Checking the nearest neighbors in train is definetly a thing i wan't to try out when i find some spare time. \nIf you want I can make the output public so you can check it.
", "rawMarkdown": "I think they are \"completly\" synthesized -- if you could ever say that about a GAN. \n\nThat's the output of the MiFID calculation (to be fair, from one model earlier):\n```\n100%|ββββββββββ| 201/201 [00:36<00:00, 5.55it/s]\n100%|ββββββββββ| 412/412 [01:09<00:00, 4.27it/s]\nm1,m2 shape= ((2048,), (2048,)) s1,s2= ((2048, 2048), (2048, 2048))\nstarting calculating FID\ncovmean.shape= (2048, 2048)\ndone with FID, starting distance calculation\nd.shape= (10000, 20579)\nnp.min(d, axis=1).shape= (10000,)\ndistance= 0.2254516669552895\nFID_public: 83.36472131619365 distance_public: 1 multiplied_public: 83.36472131619281\n```\n\nTo my understanding `distance_public` is the measurement for memorizing abilities.\nAs this is my first ever GAN experience, feel free to correct me where i am wrong. \n\nChecking the nearest neighbors in train is definetly a thing i wan't to try out when i find some spare time. \nIf you want I can make the output public so you can check it. " }, { "id": 566594, "postDate": "2019-07-02T10:57:20.550Z", "content": "The cosine memorization distance is distance= 0.2254516669552895. Since that is above threshold = 0.1, it gets set to 1. But that distance is easy to fool :-) If you rescale and crop images, then it thinks your reformatted image has a large distance from original images.
This is my concern. Someone can design a GAN which simply scales and crops original images. My experiments show that it is easy to score under LB 10 (and distance stays over threshold=0.1).
\n\nThe reason i asked about yours is not to undermine your work but rather praise it. If that dog's face in the bottom row doesn't exist on any training images, then I am super impressed with GANs.
", "rawMarkdown": "The cosine memorization distance is `distance= 0.2254516669552895`. Since that is above `threshold = 0.1`, it gets set to 1. But that distance is easy to fool :-) If you rescale and crop images, then it thinks your reformatted image has a large distance from original images.\n\nThis is my concern. Someone can design a GAN which simply scales and crops original images. My experiments show that it is easy to score under LB 10 (and distance stays over threshold=0.1).\n\nThe reason i asked about yours is not to undermine your work but rather praise it. If that dog's face in the bottom row doesn't exist on any training images, then I am super impressed with GANs.", "votes": 1 }, { "id": 566615, "postDate": "2019-07-02T11:12:34.363Z", "content": "Regarding the memorization: This article was a good read about it and the reason I said:
\n\n\n", "rawMarkdown": "Regarding the memorization: [This](https://towardsdatascience.com/do-gans-really-model-the-true-data-distribution-or-are-they-just-cleverly-fooling-us-d08df69f25eb) article was a good read about it and the reason I said:\n> \"completly\"` synthesized -- if you could ever say that about a GAN.", "votes": 1 }, { "id": 566678, "postDate": "2019-07-02T12:58:30.680Z", "content": "\"completly\"` synthesized -- if you could ever say that about a GAN.
\n
Yes, that article is exactly what I'm talking about. It's so hard to tell whether a GAN is simply cropping, interpolating, and concatenating training images or whether it learned to draw dogs from scratch.
\n\nThe tests that they mention in that article are what I wanted to do with your GAN. I wanted to search for similar images (between output and train) and I wanted to perform that test they describe. You should perform that test on your GAN.
\n\nTake two different random seeds (that you feed into your generative network). Then interpolate 48 seeds in between on the straight line between them. Finally have your generative network create the 50 images and see how your network transitions from one seed to another.
\n\nI can't wait until I get a GAN up and running. There are so many cool tests I want to try out.
", "rawMarkdown": "Yes, that article is exactly what I'm talking about. It's so hard to tell whether a GAN is simply cropping, interpolating, and concatenating training images or whether it learned to draw dogs from scratch.\n\nThe tests that they mention in that article are what I wanted to do with your GAN. I wanted to search for similar images (between output and train) and I wanted to perform that test they describe. You should perform that test on your GAN.\n\nTake two different random seeds (that you feed into your generative network). Then interpolate 48 seeds in between on the straight line between them. Finally have your generative network create the 50 images and see how your network transitions from one seed to another.\n\nI can't wait until I get a GAN up and running. There are so many cool tests I want to try out.", "votes": 1 }, { "id": 566685, "postDate": "2019-07-02T13:07:22.447Z", "content": "Here are the output images for whatever you want to try on them. \nSadly, the exact dog you mentioned is not in there, as this was an intermediate picture (not the last epoch).
\n\nRegarding the transition part, that's definetly on my TODO list. From what I have heard from you in the last competitions, you probably don't want to team-up, do you? ;)
", "rawMarkdown": "[Here](https://www.kaggle.com/ilu000/doggan-images) are the output images for whatever you want to try on them. \nSadly, the exact dog you mentioned is not in there, as this was an intermediate picture (not the last epoch).\n\nRegarding the transition part, that's definetly on my TODO list. From what I have heard from you in the last competitions, you probably don't want to team-up, do you? ;)", "votes": 1 }, { "id": 567436, "postDate": "2019-07-03T14:16:04.020Z", "content": "I posted a kernel here exploring the idea of memorization versus generalization. I believe your net is generalizing. In my kernel, I show a test that you can perform on your GAN. You can make a sequence of images from one latent space seed to another and watch how your GAN generalizes.
", "rawMarkdown": "I posted a kernel [here][1] exploring the idea of memorization versus generalization. I believe your net is generalizing. In my kernel, I show a test that you can perform on your GAN. You can make a sequence of images from one latent space seed to another and watch how your GAN generalizes.\n\n[1]: https://www.kaggle.com/cdeotte/supervised-generative-dog-net" } ] }, { "id": 572144, "postDate": "2019-07-10T14:30:10.367Z", "content": "It's funny how nice this GAN can capture small dogs shapes and details and how awful it for large breeds :)
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2972861%2Fb848b79dc9f04fc71d5180c8d7762b79%2Fdawgs.png?generation=1562768884810943&alt=media\")
These images get 45.39 locally
", "rawMarkdown": "It's funny how nice this GAN can capture small dogs shapes and details and how awful it for large breeds :)\n\n\n\nThese images get 45.39 locally", "votes": 8, "replies": [ { "id": 572149, "postDate": "2019-07-10T14:36:37.100Z", "content": "Great job! Your dog faces are very good but the bodies are not so good. I wonder why that is. If you're using a conditional GAN perhaps it's learning the different classes by getting the faces correct but ignoring learning the bodies.
", "rawMarkdown": "Great job! Your dog faces are very good but the bodies are not so good. I wonder why that is. If you're using a conditional GAN perhaps it's learning the different classes by getting the faces correct but ignoring learning the bodies." }, { "id": 572153, "postDate": "2019-07-10T14:40:13.317Z", "content": "Yeah, this one is using labels, and each image corresponds to a different class
\n\nMaybe I should pick only easy classes :D
", "rawMarkdown": "Yeah, this one is using labels, and each image corresponds to a different class\n\nMaybe I should pick only easy classes :D", "votes": 1 }, { "id": 572188, "postDate": "2019-07-10T15:43:06.877Z", "content": "You can also change the \"classes\". The \"classes\" are just attributes that you want your GAN to understand. For example you could separate all dogs into the 4 groups: no hair, short hair, straight hair, and curly hair. Or you could do 3 groups: dog facing left, dog facing center, dog facing right.
\n\nYou can also input both \"classes\" at once with Keras call Model(inputs=[seed,hair,facing], outputs=x). Then if you want a \"short hair\" dog \"facing left\", you call generator(Z, \"short hair\", \"facing left\").
If your \"classes\" are more general it won't memorize faces and ignore bodies.
", "rawMarkdown": "You can also change the \"classes\". The \"classes\" are just attributes that you want your GAN to understand. For example you could separate all dogs into the 4 groups: no hair, short hair, straight hair, and curly hair. Or you could do 3 groups: dog facing left, dog facing center, dog facing right.\n\nYou can also input both \"classes\" at once with Keras call `Model(inputs=[seed,hair,facing], outputs=x)`. Then if you want a \"short hair\" dog \"facing left\", you call `generator(Z, \"short hair\", \"facing left\")`.\n\nIf your \"classes\" are more general it won't memorize faces and ignore bodies.", "votes": 3 }, { "id": 572215, "postDate": "2019-07-10T16:18:02.253Z", "content": "Thanks, maybe I will try it later :)
", "rawMarkdown": "Thanks, maybe I will try it later :)", "votes": 1 } ] }, { "id": 569828, "postDate": "2019-07-07T12:14:13.490Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2972861%2F2fb740217bb5887fffd3ee5472c23067%2Fdoggies.png?generation=1562501486073720&alt=media\")
My intermediate results\nStill training...
", "rawMarkdown": "\n\nMy intermediate results\nStill training...", "votes": 8, "replies": [ { "id": 569832, "postDate": "2019-07-07T12:18:00.413Z", "content": "They are so good ..with lots of variety !! A big thumbs up
", "rawMarkdown": "They are so good ..with lots of variety !! A big thumbs up " }, { "id": 569860, "postDate": "2019-07-07T13:16:59.933Z", "content": "Wow, impressive!
", "rawMarkdown": "Wow, impressive!" }, { "id": 570092, "postDate": "2019-07-07T19:04:55.007Z", "content": "Thanks!\nI am a bit sad there are still a lot of wishy washy images. I got 55.55 locally and 53.77655 on public LB\nI guess there is a lot of room for improvement...
", "rawMarkdown": "Thanks!\nI am a bit sad there are still a lot of wishy washy images. I got 55.55 locally and 53.77655 on public LB\nI guess there is a lot of room for improvement...", "votes": 1 }, { "id": 570094, "postDate": "2019-07-07T19:06:39.523Z", "content": "That's a great score. I think the best conventional GANs in this comp are currently scoring around 45.
", "rawMarkdown": "That's a great score. I think the best conventional GANs in this comp are currently scoring around 45." }, { "id": 570143, "postDate": "2019-07-07T21:31:24.040Z", "content": "Very impressive results.
\n\nI am still struggeling with non converging or slowly training models, but your results give me hope that it's possible.
", "rawMarkdown": "Very impressive results. \n\nI am still struggeling with non converging or slowly training models, but your results give me hope that it's possible. " }, { "id": 570456, "postDate": "2019-07-08T10:14:09.283Z", "content": "That's a good guess @cdeotte, 42 is the score to beat now, congratz @dvorobiev !
", "rawMarkdown": "That's a good guess @cdeotte, 42 is the score to beat now, congratz @dvorobiev !", "votes": 1 }, { "id": 570479, "postDate": "2019-07-08T10:59:32.487Z", "content": "Thanks :) \nSolution for 42 is a lot heavier than previous. I almost ran out of time. I guess, it will be hard to improve with current arch, unless I find a way to cut off some computations
", "rawMarkdown": "Thanks :) \nSolution for 42 is a lot heavier than previous. I almost ran out of time. I guess, it will be hard to improve with current arch, unless I find a way to cut off some computations", "votes": 2 } ] }, { "id": 564357, "postDate": "2019-06-29T10:07:32.660Z", "content": "Looks scary\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2003084%2F4724c217b9bdcb7a4ec0969a9b1badef%2F__results___17_0.png?generation=1561802828241411&alt=media\")
very scary !!!,
", "rawMarkdown": "very scary !!!," }, { "id": 564419, "postDate": "2019-06-29T11:50:23.627Z", "content": "Remind me this ;)
", "rawMarkdown": "Remind me [this](https://youtu.be/PXXoq3yr4eI?t=28) ;)", "votes": 7 } ] }, { "id": 563926, "postDate": "2019-06-28T19:31:29.243Z", "content": "The first attempt went wrong!\n![\"\"](\"https://i.imgur.com/WZnLyIN.png\")
A bit better\n![\"\"](\"https://i.imgur.com/FZreEob.png\")
Big GAN 256 scores LB 15 !! Kaggle kernel/notebook here.\n![\"image\"](\"http://playagricola.com/Kaggle/gan81419.png\")
Very nice, thanks for sharing it!
", "rawMarkdown": "Very nice, thanks for sharing it!", "votes": 1 }, { "id": 599458, "postDate": "2019-08-15T02:03:56.827Z", "content": "Awesome!!
", "rawMarkdown": "Awesome!!", "votes": 1 }, { "id": 599508, "postDate": "2019-08-15T04:11:12.777Z", "content": "nice work chris!
", "rawMarkdown": "nice work chris!", "votes": 1 } ] }, { "id": 578684, "postDate": "2019-07-18T02:50:59.403Z", "content": "66.9 LB
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F878270%2F4069576f75cf5d8545c57916d51e601e%2F.JPG?generation=1563418240253289&alt=media\")
I tried training only on one breed of Maltese Dog\nhere is the result after 30 minutes of training. And gif of training in progress.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2Fd61e20e33a27917e3ee8b817df3037d1%2Ftrain_00014450.png?generation=1562682396674729&alt=media\")
That's awesome!! Fantastic! You should try building a CGAN. When you train your Generator and Discriminator, give it a breed of dog. Then after your GAN is trained, you feed a random Z plus a random Breed (1 to 120) into your Generator.
", "rawMarkdown": "That's awesome!! Fantastic! You should try building a [CGAN][1]. When you train your Generator and Discriminator, give it a breed of dog. Then after your GAN is trained, you feed a random Z plus a random Breed (1 to 120) into your Generator.\n\n[1]: https://arxiv.org/abs/1411.1784", "votes": 1 }, { "id": 571376, "postDate": "2019-07-09T14:37:50.473Z", "content": "Thank you Chris! I'm currently giving breeds info to my D and G networks. It gave a significant boost when I implemented it. Now I'm thinking on adding classifier head to D network to force it to learn better features.
", "rawMarkdown": "Thank you Chris! I'm currently giving breeds info to my D and G networks. It gave a significant boost when I implemented it. Now I'm thinking on adding classifier head to D network to force it to learn better features.", "votes": 1 }, { "id": 571419, "postDate": "2019-07-09T15:19:50.607Z", "content": "One question , I have read that MFID uses variety of images also as parameter . In that case , if we generate one breed of dogs as output . Even with good image it can give lower MFID ?
", "rawMarkdown": "One question , I have read that MFID uses variety of images also as parameter . In that case , if we generate one breed of dogs as output . Even with good image it can give lower MFID ?" }, { "id": 571440, "postDate": "2019-07-09T15:37:54.177Z", "content": "I think so. Distribution of Inception latent features will be different for only one breed.
", "rawMarkdown": "I think so. Distribution of Inception latent features will be different for only one breed. " }, { "id": 571622, "postDate": "2019-07-09T21:05:46.930Z", "content": "\n\n\nOne question , I have read that MFID uses variety of images also as parameter . In that case , if we generate one breed of dogs as output . Even with good image it can give lower MFID ?
\n
Yes, Nirjhar. But learning breeds will give you better variety. You train the Generator by giving it a breed number (1-120) and a random seed Z. Then when you create your 10,000 dogs for Kaggle, you input a random Z and random breed number from 1-120. This way your 10,000 images are guarenteed to have approx equal distribution of all 120 breeds represented.
", "rawMarkdown": "> One question , I have read that MFID uses variety of images also as parameter . In that case , if we generate one breed of dogs as output . Even with good image it can give lower MFID ?\n\nYes, Nirjhar. But learning breeds will give you **better** variety. You train the Generator by giving it a breed number (1-120) and a random seed Z. Then when you create your 10,000 dogs for Kaggle, you input a random Z and random breed number from 1-120. This way your 10,000 images are guarenteed to have approx equal distribution of all 120 breeds represented.", "votes": 1 }, { "id": 571632, "postDate": "2019-07-09T21:34:34.083Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 571864, "postDate": "2019-07-10T06:59:04.940Z", "content": "\n\n\nYes, Nirjhar. But learning breeds will give you better variety. You train the Generator by giving it a breed number (1-120) and a random seed Z. Then when you create your 10,000 dogs for Kaggle, you input a random Z and random breed number from 1-120. This way your 10,000 images are guarenteed to have approx equal distribution of all 120 breeds represented.
\n
Yeah, I can confirm that. I feed all 120 labels uniformly when generating images. I also use it to view intermediate results during training, and since the label for each cell of the plot is constant, I can see a progress of my net on each breed :) Usually I visualize only the first 32 of them.
", "rawMarkdown": ">Yes, Nirjhar. But learning breeds will give you better variety. You train the Generator by giving it a breed number (1-120) and a random seed Z. Then when you create your 10,000 dogs for Kaggle, you input a random Z and random breed number from 1-120. This way your 10,000 images are guarenteed to have approx equal distribution of all 120 breeds represented.\n\nYeah, I can confirm that. I feed all 120 labels uniformly when generating images. I also use it to view intermediate results during training, and since the label for each cell of the plot is constant, I can see a progress of my net on each breed :) Usually I visualize only the first 32 of them.", "votes": 3 } ] }, { "id": 568886, "postDate": "2019-07-05T16:03:19.627Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F515b7b71f4f69dacbc4507f1e77e6094%2Ftrain_00067600.png?generation=1562342592096599&alt=media\")
Real GAN? I am deeply impressed If this is done in 9h Kernel time. Good job!\nWhat is the FID Score?
", "rawMarkdown": "Real GAN? I am deeply impressed If this is done in 9h Kernel time. Good job!\nWhat is the FID Score?" }, { "id": 568892, "postDate": "2019-07-05T16:10:39.947Z", "content": "It's real DCGAN after only ~2 hours of training on local machine without any external data, now working on moving it to the kernel. I haven't computed FID scores so far, working on task in general, not metric
", "rawMarkdown": "It's real DCGAN after only ~2 hours of training on local machine without any external data, now working on moving it to the kernel. I haven't computed FID scores so far, working on task in general, not metric", "votes": 2 }, { "id": 568894, "postDate": "2019-07-05T16:12:18.717Z", "content": "New leading score for real GANs incoming ;)\nWell done!
", "rawMarkdown": "New leading score for real GANs incoming ;)\nWell done!" }, { "id": 568896, "postDate": "2019-07-05T16:14:42.337Z", "content": "I already have a kernel with good looking dogs, but thanks to the bug I generated 10048 images instead of 10000 and fail to submit it))))
", "rawMarkdown": "I already have a kernel with good looking dogs, but thanks to the bug I generated 10048 images instead of 10000 and fail to submit it))))" }, { "id": 568903, "postDate": "2019-07-05T16:22:31.747Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F41f51b8f2f7ee0340d7f3072983e6438%2F0.0025806725025177_6532.png?generation=1562343743799497&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F734076228d3ec89480b4970c313c3606%2F0.001102358102798462_3720.png?generation=1562343744988485&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F00a446e02d869d6ceda6b7ea14366227%2F0.01647740602493286_691.png?generation=1562343747343305&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F7d94922e2db3901402d162800308563e%2F0.00148087739944458_5362.png?generation=1562343747519162&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2Fd7859dafcbf6eaf933bfd17d24cf2346%2F0.000816643238067627_5377.png?generation=1562343747473402&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F4ff7bf113f051d86d0835a774a30902d%2F0.0001100003719329834_9311.png?generation=1562343747506109&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F72055339157712666f90edc120371e90%2F0.00782039761543274_6637.png?generation=1562343747695834&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2Fe00dc7e18d1caa53cf762489bbfcb51d%2F0.00630989670753479_7778.png?generation=1562343747689118&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F111325fe28f19d0885f43f8856bb433d%2F0.00038909912109375_8779.png?generation=1562343747765554&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F0051993f3d29509d478ca198d2a8a8da%2F0.001667410135269165_1508.png?generation=1562343747678271&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F33538f997907088008d026d787cf5bd9%2F0.0791085958480835_9619.png?generation=1562343748256094&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2F7cad5ba07777cfcdb3367f0b42c39821%2F0.001471191644668579_1643.png?generation=1562343748970014&alt=media\")
Awesome job Kostya. They look great. From your images, it looks like your GAN first builds a face and then places it on a body. Nice design.
", "rawMarkdown": "Awesome job Kostya. They look great. From your images, it looks like your GAN first builds a face and then places it on a body. Nice design.", "votes": 2 }, { "id": 568962, "postDate": "2019-07-05T19:02:44.570Z", "content": "May I ask, if you use a selection of the \"best\" dogs to train your model?
", "rawMarkdown": "May I ask, if you use a selection of the \"best\" dogs to train your model? ", "votes": 1 }, { "id": 569012, "postDate": "2019-07-05T21:32:13.133Z", "content": "No, I train on all dogs available. And what do you mean by a selection of the \"best\" dogs?
", "rawMarkdown": "No, I train on all dogs available. And what do you mean by a selection of the \"best\" dogs?", "votes": 1 }, { "id": 569021, "postDate": "2019-07-05T21:49:10.527Z", "content": "Ok, I was wondering if it boosts model performance if we curate the images first. Like removing pictures with watermarks, low resolution, hardly visible dogs etc...
\n\nOr making subsamples of front-facing-dogs, dogs visible from the side ...
", "rawMarkdown": "Ok, I was wondering if it boosts model performance if we curate the images first. Like removing pictures with watermarks, low resolution, hardly visible dogs etc...\n\nOr making subsamples of front-facing-dogs, dogs visible from the side ...", "votes": 1 }, { "id": 572243, "postDate": "2019-07-10T17:14:20.317Z", "content": "i like how the last but one doggy has bunny teeth
", "rawMarkdown": "i like how the last but one doggy has bunny teeth", "votes": 2 } ] }, { "id": 566188, "postDate": "2019-07-01T21:58:43.857Z", "content": "After a few false starts, my first DCGAN at the end of its run. If you squint really hard, a few of them look dog-esque.![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F47954%2F75adb8290f9c856d78d9e6de9b9f8a24%2Famw5gster_DCGAN_sample.png?generation=1562018289077797&alt=media\")
Here are two of my GANs doing latents walk between dog A and dog B. Notice how the top GAN generalizes and invents a 3rd new dog as the middle point (in step 4 and 5). The bottom GAN does a basic pixel blend with awkward middle images.
\n\n![\"image\"](\"http://playagricola.com/Kaggle/latentc71219.png\")
For comparison, here's a Memorizer GAN's pixel blend of the same 2 images:\n![\"image\"](\"http://playagricola.com/Kaggle/blendc71219.png\")
few good images, but almost scary!!π
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F916454%2Fa21b054e51a804193e9745923ae48f19%2Fimage_latest.png?generation=1562830661069656&alt=media\")
![\"Dogs\"](\"https://i.imgur.com/jj34JSO.jpg\")
at least some of them look \"doglike\"
", "rawMarkdown": "\n\nat least some of them look \"doglike\"", "votes": 6 }, { "id": 599392, "postDate": "2019-08-14T22:56:07.010Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2062758%2F86f67d50425eea7ca0e38f01766c6362%2F__results___81_0.png?generation=1565823305865128&alt=media\")
woof
", "rawMarkdown": "\n\nwoof", "votes": 3 }, { "id": 598676, "postDate": "2019-08-13T21:36:00.080Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F133817%2F15e4eee8698f9daa81454b0e063a0d0f%2FScreenshot%20from%202019-08-14%2000-33-00.png?generation=1565732139308296&alt=media\")
\nThe dog we deserved π
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2236951%2F6d4dc27582a2c7ce2512a722477ab388%2Fss.png?generation=1565503251514167&alt=media\")
\nwas making a dog generator\ngot a patrick from spongebob generator
Some of the best dogs I got so far : \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2F29d93dd019eb16b8e2f7c24bca912c0d%2Fimage_06438.png?generation=1565181076304131&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2F4dcb6933c32633b90e94db7311fe9559%2F67766157_915909742076021_7404671193380290560_n.png?generation=1565181098313620&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2F85c332ecaf896535d103df8f14638535%2Fimage_09014.png?generation=1565181094616869&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2Fb2976537b3e11d02cc576bf1f4d52a07%2Fimage_07096.png?generation=1565181091683381&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2F5a1a419800bb121b135182f41bd69e76%2Fimage_04519.png?generation=1565181080637459&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2F87d6a7a3d47a516d6462ad4415fccaea%2F67479159_392929381351845_938845211994357760_n.png?generation=1565181080236162&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2Fa20579cb6a3c5ad2ae69067fda9f1eae%2Fimage_00022.png?generation=1565181078647737&alt=media\")
Still have a lot of messy images, so the LB score is not so great (~82)
", "rawMarkdown": "Some of the best dogs I got so far : \n  \n   \n\nStill have a lot of messy images, so the LB score is not so great (~82)", "votes": 3 }, { "id": 572996, "postDate": "2019-07-11T16:24:46.430Z", "content": "After 100 epochs [DCGAN]\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1614008%2F24a0e5d23af1ab843bc8ff01f461f997%2F100epoch.png?generation=1562862241800562&alt=media\")
After 800 epochs [DCGAN]\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1614008%2Fcd28e0999d633b26162e012154ad9a68%2F__results___13_0.png?generation=1562954423272836&alt=media\")
how is you G loss using 800 epoch?
", "rawMarkdown": "how is you G loss using 800 epoch?" }, { "id": 574360, "postDate": "2019-07-13T17:14:23.910Z", "content": "I think that G and D loss is not that much informative so I didn't track the G and D loss
After 100 epochs [RaLSGAN]\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1614008%2F9d3df28391574358714cfda060d96c87%2Fdownload.png?generation=1563042497285260&alt=media\")
Looks great. 100 epochs of RaLSGAN seems to be as good as 800 epochs of DCGAN.
", "rawMarkdown": "Looks great. 100 epochs of RaLSGAN seems to be as good as 800 epochs of DCGAN. ", "votes": 2 }, { "id": 574550, "postDate": "2019-07-14T04:23:51.010Z", "content": "Ya, that's because of your loss Function
", "rawMarkdown": "Ya, that's because of your [loss Function](https://www.kaggle.com/c/generative-dog-images/discussion/99485#latest-574501)", "votes": 3 }, { "id": 577507, "postDate": "2019-07-16T18:26:25.243Z", "rawMarkdown": "", "isDeleted": true }, { "id": 579601, "postDate": "2019-07-19T02:33:59.203Z", "content": "I have tried RaLSGAN but dont have good results as yours
", "rawMarkdown": "I have tried RaLSGAN but dont have good results as yours" } ] }, { "id": 572062, "postDate": "2019-07-10T12:04:57.337Z", "content": "animated collapse of one of my local runs
", "rawMarkdown": "animated collapse of one of my local runs", "votes": 3, "replies": [ { "id": 572080, "postDate": "2019-07-10T12:25:36.567Z", "content": "Cool. Is each image (of the 64) the same seed over time? It's interesting to watch the evolution of the images. I would have thought that the evolution would be smoother. There's a lot of flickering going on. I wonder if that is the GAN or the animated GIF compression algorithm. If I focus on one image, there is lot of changes occurring in the top left corner.
\n\nYou should save a sequence of one image (through all epochs) as an array of loss-less images. And display those. Then we can determine what artifacts in the movie are a result of GIF compression and what are caused by GAN.
", "rawMarkdown": "Cool. Is each image (of the 64) the same seed over time? It's interesting to watch the evolution of the images. I would have thought that the evolution would be smoother. There's a lot of flickering going on. I wonder if that is the GAN or the animated GIF compression algorithm. If I focus on one image, there is lot of changes occurring in the top left corner. \n\nYou should save a sequence of one image (through all epochs) as an array of loss-less images. And display those. Then we can determine what artifacts in the movie are a result of GIF compression and what are caused by GAN." }, { "id": 572091, "postDate": "2019-07-10T12:50:39.833Z", "content": "It's 4000 batches between consecutive images in the gif above. Here is the starting part of it with a step of 100 batches
", "rawMarkdown": "It's 4000 batches between consecutive images in the gif above. Here is the starting part of it with a step of 100 batches", "votes": 1 } ] }, { "id": 569605, "postDate": "2019-07-07T02:25:08.213Z", "content": "In my post Animated GANs - See the latent walk evolution I explain how to generate this animation using the kernel Animation - GAN Latent Walk
\n\nI'm looking forward to see your results animated too :)
\n\n\n\n
", "rawMarkdown": "In my post [Animated GANs - See the latent walk evolution](https://www.kaggle.com/c/generative-dog-images/discussion/98853#latest-569598) I explain how to generate this animation using the kernel [Animation - GAN Latent Walk](https://www.kaggle.com/jesucristo/animation-gan-latent-walk)\n\nI'm looking forward to see your results animated too :)\n\n### Some results:\n\n\n\n", "votes": 3 }, { "id": 567791, "postDate": "2019-07-04T02:35:39.310Z", "content": "I am very sorry I washed my dogs , now they looks like abstract art .Fast.ai WGAN\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F3d5559eb435d0e6bd1868d46e7fa19b6%2FCapture.PNG?generation=1562207732242307&alt=media\")
Looks like a famous fresco...
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2600958%2F7e432587fd89b24d3df676202020367f%2F.png?generation=1565167279030121&alt=media\")
![\"[Imgur\"](\"https://i.imgur.com/MUT7WKt.png\")
Are they looking like dogs or am I just too optimistic?
", "rawMarkdown": "\n\nAre they looking like dogs or am I just too optimistic? ", "votes": 3, "replies": [ { "id": 565514, "postDate": "2019-07-01T03:45:35.603Z", "content": "I guess too optimistic !
", "rawMarkdown": "I guess too optimistic !", "votes": 2 } ] }, { "id": 588189, "postDate": "2019-07-30T09:20:22.983Z", "content": "wow... is this melted panda...?\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1755928%2F59d5d7b2028aa9790cbae47acd5c1f35%2FScreen%20Shot%202019-07-30%20at%2018.15.12.png?generation=1564478397541378&alt=media\")
Our GAN successfully generated ghost-like skeleton doggos : \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2909950%2F01a546cb9cfefbe903d8f993110cab23%2Fskeleton-doggo.png?generation=1564145792191842&alt=media\")
New best dogs using the baseline Introducing DCGAN Dogs Images + tricks from All you need is GAN Hacks\n.png?generation=1562804966674757&alt=media)
LB 102 (see v39)
", "rawMarkdown": "New best dogs using the baseline [Introducing DCGAN Dogs Images](https://www.kaggle.com/jesucristo/introducing-dcgan-dogs-images) + tricks from [All you need is GAN Hacks](https://www.kaggle.com/c/generative-dog-images/discussion/98595#latest-570614)\n.png?generation=1562804966674757&alt=media)\n\n**LB 102** (see v39)", "votes": 4, "replies": [ { "id": 572457, "postDate": "2019-07-11T00:54:02Z", "content": "Nice. They look great
", "rawMarkdown": "Nice. They look great" }, { "id": 572778, "postDate": "2019-07-11T11:51:16.443Z", "content": "Nice images! How come that you only score 102?
", "rawMarkdown": "Nice images! How come that you only score 102?", "votes": 1 }, { "id": 572802, "postDate": "2019-07-11T12:19:38.023Z", "content": "I suspect it has something to do with either a) the way MFID calculated ,may be there are lots of washed images and many good images \nb) some problem in image normalization /denormalization \nI had higher fid value for some run where I found the sample dogs are looking more doggish than the kernel where I got lower fid score .
", "rawMarkdown": "I suspect it has something to do with either a) the way MFID calculated ,may be there are lots of washed images and many good images \nb) some problem in image normalization /denormalization \nI had higher fid value for some run where I found the sample dogs are looking more doggish than the kernel where I got lower fid score .", "votes": 1 }, { "id": 572808, "postDate": "2019-07-11T12:27:34.080Z", "content": "Yes, if for your own eyes G produces more doggish images you cant be sure it will be so doggish for Inception latent space.
", "rawMarkdown": "Yes, if for your own eyes G produces more doggish images you cant be sure it will be so doggish for Inception latent space." } ] }, { "id": 571245, "postDate": "2019-07-09T11:12:43.630Z", "content": "my best scoring kernel so far with 63![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1544891%2Fffe8ad3a7346ec60c9288765445313b4%2Fsample.png?generation=1562670802577793&alt=media\")
Impressive! Thanks for sharing.
", "rawMarkdown": "Impressive! Thanks for sharing. ", "votes": 1 }, { "id": 571372, "postDate": "2019-07-09T14:34:34.870Z", "content": "Thank you. Now I'm deciding what to do next, to try to optimize FID for this network, or switch and try one of the plenty fancy sota gans popping out every day(
", "rawMarkdown": "Thank you. Now I'm deciding what to do next, to try to optimize FID for this network, or switch and try one of the plenty fancy sota gans popping out every day(" } ] }, { "id": 571241, "postDate": "2019-07-09T11:04:33.280Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2185199%2Fdd51866f95974b62737f2459c4b182e9%2FCapture.PNG?generation=1562670268627478&alt=media\")
Some of my monsters. Score 77.
", "rawMarkdown": "\n\n\nSome of my monsters. Score 77.", "votes": 4 }, { "id": 569801, "postDate": "2019-07-07T11:20:44.367Z", "content": "Real GAN >\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2062758%2F57f8454a3319f1e7d5e0e208b24a5d1e%2F__results___68_0.png?generation=1562498380715515&alt=media\")
Wow Theo . What's your score ?
", "rawMarkdown": "Wow Theo . What's your score ?", "votes": 1 }, { "id": 569815, "postDate": "2019-07-07T11:55:05.343Z", "content": "Great job Theo!
", "rawMarkdown": "Great job Theo!", "votes": 1 }, { "id": 569955, "postDate": "2019-07-07T15:43:20.890Z", "content": "@phoenix9032 About 60 on my offline computations\n@cdeotte Thanks a lot Chris!
", "rawMarkdown": "@phoenix9032 About 60 on my offline computations\n@cdeotte Thanks a lot Chris!", "votes": 1 } ] }, { "id": 565470, "postDate": "2019-07-01T01:46:33.737Z", "content": "My favourite creation
\n\n", "rawMarkdown": "My favourite creation\n\n![\"Captura-de-pantalla-de-2019-07-01-03-44-18\"](\"https://i.ibb.co/bbq1xPX/Captura-de-pantalla-de-2019-07-01-03-44-18.png\")
",
"votes": 3
},
{
"id": 599242,
"postDate": "2019-08-14T17:59:57.207Z",
"content": "My best run:\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F204221%2F501549114a2d27e334e11c66e185ea1e%2Fres.png?generation=1565805337765858&alt=media\")
I am not sure why there is grey, but I think it is a result of overfitting because when I train a bit less the images are coloured normally.
", "rawMarkdown": "My best run:\n\n\nI am not sure why there is grey, but I think it is a result of overfitting because when I train a bit less the images are coloured normally.\n", "votes": 1 }, { "id": 599192, "postDate": "2019-08-14T16:35:07.880Z", "content": "I just have to share this one.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Ff023f7055f014ff21e17b70f095d6be9%2F53.png?generation=1565800494981902&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F413189%2F87d5487e73dc8684c8f519b91bcd005d%2FGan%20Output.png?generation=1565417206002663&alt=media\")
\nGAN dog looks like a goat
Some dogs generated with a race conditional GAN. Score is 77~. Not really close from my best models but at least it's working.
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2062758%2Fc7a89815be2a317dafeb1e13348f198f%2F__results___63_11.png?generation=1563179157173840&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2062758%2Fea4d7ee9aaf3a5ed650e84262d971f67%2F__results___62_0.png?generation=1563299139772985&alt=media\")
Better, scores 65~
", "rawMarkdown": "\n\nBetter, scores 65~", "votes": 3 }, { "id": 577466, "postDate": "2019-07-16T17:53:52.667Z", "content": "Wow, looks great. Is that a CGAN with one output neuron in Discriminator? Or are you making some sort of CGAN-classifier which has 120 softmax output?
", "rawMarkdown": "Wow, looks great. Is that a CGAN with one output neuron in Discriminator? Or are you making some sort of CGAN-classifier which has 120 softmax output?", "votes": 1 }, { "id": 577606, "postDate": "2019-07-16T20:32:33.397Z", "content": "CGAN with one output only, I did not have any success with ACGANs yet
", "rawMarkdown": "CGAN with one output only, I did not have any success with ACGANs yet", "votes": 1 } ] }, { "id": 570248, "postDate": "2019-07-08T03:52:10.450Z", "content": "TensorFlow/Keras DCGAN after 35 epochs (22000 batches):
\n\n![\"image\"](\"http://playagricola.com/Kaggle/dog2.png\")
The image is not visible for me . Not sure if its only my problem
", "rawMarkdown": "The image is not visible for me . Not sure if its only my problem" }, { "id": 570468, "postDate": "2019-07-08T10:39:27.977Z", "content": "I see them
", "rawMarkdown": "I see them" }, { "id": 570886, "postDate": "2019-07-08T22:34:54.617Z", "content": "Here's 250 epochs. It has MiFID score 110:\n![\"image\"](\"http://playagricola.com/Kaggle/dogs7819.png\")
My dogs are getting better! It's so exciting to watch dogs grow! Scores 65 now with 150 epochs. I changed architecture and cost functions for Discriminator and Generator:\n![\"image\"](\"http://playagricola.com/Kaggle/doga71019.png\")
65 in 150 epochs !! Thats magic for me :) Once the competition finishes would love to know those tricks .
", "rawMarkdown": "65 in 150 epochs !! Thats magic for me :) Once the competition finishes would love to know those tricks .", "votes": 1 } ] }, { "id": 599386, "postDate": "2019-08-14T22:43:47.383Z", "content": "![\"\"](\"https://raw.githubusercontent.com/balansky/pytorch_gan/master/samples/250000_img.png\")
Nice work. Good luck
", "rawMarkdown": "Nice work. Good luck", "votes": 1 }, { "id": 599390, "postDate": "2019-08-14T22:50:36.097Z", "content": "I can't wait to see your solution @cdeotte :) Hope is not BigGan so I could learn another new thing ;)
", "rawMarkdown": "I can't wait to see your solution @cdeotte :) Hope is not BigGan so I could learn another new thing ;)" } ] }, { "id": 571194, "postDate": "2019-07-09T09:34:44.770Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1176471%2Fe613015366a9dd6d0c8f5929ad9a035c%2F799649B6-326A-4D63-BF33-63CEE84B1399.png?generation=1562664860589494&alt=media\")
Not yet dogs , but i can see one interesting thing that the nose probably gets understood by the model first . Second could be eyes ? Are you noticing the same thing ? \nSo 20000+ front facing dogs could be generating better dogs ?
", "rawMarkdown": "Not yet dogs , but i can see one interesting thing that the nose probably gets understood by the model first . Second could be eyes ? Are you noticing the same thing ? \nSo 20000+ front facing dogs could be generating better dogs ?", "votes": 1 }, { "id": 571223, "postDate": "2019-07-09T10:26:47.983Z", "content": "Yes, agree. This is after 100 epochs. Trying to see if longer training might help. LB score for this is 111.
", "rawMarkdown": "Yes, agree. This is after 100 epochs. Trying to see if longer training might help. LB score for this is 111." }, { "id": 571339, "postDate": "2019-07-09T14:05:35.950Z", "rawMarkdown": "", "isDeleted": true }, { "id": 571343, "postDate": "2019-07-09T14:08:01.250Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1176471%2F78391007ece1e491bbfc5e4939dd43a6%2F__results___41_0.png?generation=1562681258897673&alt=media\")
After 250 epochs. Not much improvement.
", "rawMarkdown": "After 250 epochs. Not much improvement." }, { "id": 571363, "postDate": "2019-07-09T14:29:13.577Z", "content": "But you did have a MFID jump right ?
", "rawMarkdown": "But you did have a MFID jump right ?" }, { "id": 571410, "postDate": "2019-07-09T15:12:13.857Z", "content": "Yes, this has LB score of 93, from 111.
", "rawMarkdown": "Yes, this has LB score of 93, from 111." } ] }, { "id": 569025, "postDate": "2019-07-05T22:09:55.570Z", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F5cd1233b0188ca7b3314701c49c3b6e9%2FCapture.PNG?generation=1562364575882529&alt=media\")
I think i finally might be onto something here ...
", "rawMarkdown": "\n\nI think i finally might be onto something here ...", "votes": 2 }, { "id": 568990, "postDate": "2019-07-05T20:07:14.457Z", "content": "I like the dog in the top row, 3rd from right.\n![\"dogs\"](\"https://i.imgur.com/sjEfJh0.png\")
Sadly, I see an increase in my loss variances that I can't get rid off. Any tips? That might be the problem why the dogs quality doesn't increase much after 50k iterations for me.
\n![\"loss\"](\"https://i.imgur.com/2vFfx38.png\")
I made images walking in the latent space\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1565399%2Feb9fa1598beee75c243cc5ade8b9e8e7%2Fimages_latent.png?generation=1562336014492197&alt=media\")
\nHere, I picked up two good images manually (most left one and most right one)
Wow, that's fantastic. Super job Tom! That demonstrates that a generalizing GAN's latent walk is not a simple pixel blend. The eyes move and an ear turns into an eye.
", "rawMarkdown": "Wow, that's fantastic. Super job Tom! That demonstrates that a **generalizing** GAN's latent walk is not a simple pixel blend. The eyes move and an ear turns into an eye." } ] }, { "id": 566249, "postDate": "2019-07-02T00:53:41.237Z", "content": "Produced by adding variance to an autoencoder's latent layer in this kernel. However after reading the new rules clarification here, I don't believe these images are legal. But all is not lost, I'll save the decoder (as a generative network), add a discriminative network, and turn it into a GAN !
\n\n![\"image\"](\"http://playagricola.com/Kaggle/dogs7219.png\")
great idea! \nPS: third image of the last row hahahaha a human face dog π π
", "rawMarkdown": "great idea! \nPS: third image of the last row hahahaha a human face dog π π ", "votes": 4 }, { "id": 566948, "postDate": "2019-07-02T20:20:09.300Z", "content": "That's a VAE-GAN: https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a
", "rawMarkdown": "That's a VAE-GAN: https://towardsdatascience.com/what-the-heck-are-vae-gans-17b86023588a", "votes": 1 }, { "id": 567439, "postDate": "2019-07-03T14:22:23.187Z", "content": "I'm making progress. I turned my autoencoder's decoder into a supervised generative network here. It produced the dogs below. They look scary! Now I'm gonna to add a discriminative network to help clean up these images.
\n\n![\"image\"](\"http://playagricola.com/Kaggle/dogs7319.png\")
One more step. I turned my supervised generative net into a GAN posted here. Here is the first output of dogs. They score LB 90. If I adjust hyperparameters and architecture, it can score LB 7 but then it most likely becomes an invalid solution (because the output is too similar to training images).
\n\nThe next step is have the Discriminator teach the Generator to generalize better.\n![\"image\"](\"http://playagricola.com/Kaggle/dogs7619.png\")
a.k.a Choose your own LB score. π€
", "rawMarkdown": "a.k.a Choose your own LB score. π€ ", "votes": 2 } ] }, { "id": 565265, "postDate": "2019-06-30T17:18:34.890Z", "content": "Using \nIntroducing DCGAN Dogs Images\n![\"Captura-de-pantalla-de-2019-06-30-14-50-01\"](\"https://i.ibb.co/k2HGV7w/Captura-de-pantalla-de-2019-06-30-14-50-01.png\")
\n",
"votes": 1
},
{
"id": 599142,
"postDate": "2019-08-14T14:48:44.050Z",
"content": "A new breed: Melted-Cheese Dog.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F694899%2F4fe26f36caa87bc8d7f466f6ec902380%2F9984.png?generation=1565794019873389&alt=media\")
Some of my dogs. Front-facing dogs are working best, visually.\nSome images are still completely abstract.\nIn a lot of images, I have some sort of blue \"bleeding\", I do not yet understand why. I do not have different processes for the different color channels. Is it something that can happen? My GAN is a bit special/home made so maybe I'm missing a side effect from my code...
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Ff0e2b53d95b31c4ad6343bff0c8547ab%2F216.png?generation=1565702226849958&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2F39a7de3cbed4a631c630b7fbf4a81c9c%2F131.png?generation=1565702226591248&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2F3b06b36b59d7977899c15ccf07f6d8eb%2F115.png?generation=1565702226416177&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2F07d5a9f0ecb9382f4d937cee7ad70093%2F717.png?generation=1565702226274440&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2F517ba4d2cff90a2fba69381e2029c898%2F8494.png?generation=1565702225731400&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Fd95cf31795d6404db1678421cbc0d420%2F268.png?generation=1565702225728240&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2F214d31b7a8313f328bd992ba32d0b85d%2F8254.png?generation=1565702230432368&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2F92e7c6108a29cfffa6e323ab9cd335f4%2F92.png?generation=1565702229515255&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Fda25813d22a75048ac6c97cff47dada9%2F8173.png?generation=1565702228609079&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Fbf241c7d436057f408c74a7ba8a36600%2F318.png?generation=1565702227624948&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Fa76a6235c45beffdd07335dfdc7ebb80%2F455.png?generation=1565702227013806&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3550605%2Ffbcd7b8d6cbb5e4f08234aefa9734d87%2F8517.png?generation=1565702236105446&alt=media\")
If you are using python-opencv (cv2), converting color from BGR to RGB is recommended. Are you doing this?
", "rawMarkdown": "If you are using python-opencv (cv2), converting color from BGR to RGB is recommended. Are you doing this?", "votes": 1 }, { "id": 598695, "postDate": "2019-08-13T22:35:59.537Z", "content": "Thank you, yes I am doing this already. There is a good chance there is something stupid in my code, or maybe this is an interesting behavior... I will try to investigate further once I have ran my last GANs for today
", "rawMarkdown": "Thank you, yes I am doing this already. There is a good chance there is something stupid in my code, or maybe this is an interesting behavior... I will try to investigate further once I have ran my last GANs for today" }, { "id": 598718, "postDate": "2019-08-14T00:14:27.540Z", "content": "Well, it turns out it was this anyway... I did the conversion at the beginning of my pipeline but should be doin the reverse at the end probably, because I use OpenCV to write the result images π
", "rawMarkdown": "Well, it turns out it was this anyway... I did the conversion at the beginning of my pipeline but should be doin the reverse at the end probably, because I use OpenCV to write the result images π ", "votes": 1 } ] }, { "id": 595985, "postDate": "2019-08-10T00:06:03.910Z", "content": "hello guys...\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F800064%2Fed74796fc14f0de872b4abd950f0a696%2F.JPG?generation=1565395561433748&alt=media\")
preliminary result by DCGAN with ImageNet classification result\nnothing like dog yet...
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1014923%2F69c344df508e5447d00eb9c68981a323%2F20190808-dcgan-with-imagenet.jpg?generation=1565331036811684&alt=media\")
I removed Batch Normalization on the discriminator and it gave me 50 Shades of Grey images for a batch size of 50. There are some dogs hidden behind that , but why the Grey curtain ?\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2F81d0f3ef50d2093fcda3dfb593b1d3a5%2FnobatchNorm.png?generation=1563976256976486&alt=media\")
perhaps they are images with pixel range [0.4,0.6]. Check the max and min pixel values and scale it [0,1]
", "rawMarkdown": "perhaps they are images with pixel range [0.4,0.6]. Check the max and min pixel values and scale it [0,1]", "votes": 1 } ] }, { "id": 570911, "postDate": "2019-07-08T23:35:41.460Z", "content": "DCGAN with Complement between vectors
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1517250%2F2f22df7d615bc303573f939d8df7c7e1%2F__results___9_0.png?generation=1562628896340279&alt=media\")
@ibraheemmoosa Here, have some nightmare material xD
", "rawMarkdown": "@ibraheemmoosa Here, have some nightmare material xD" }, { "id": 588188, "postDate": "2019-07-30T09:16:55.177Z", "rawMarkdown": "", "isDeleted": true }, { "id": 584412, "postDate": "2019-07-25T22:28:01.420Z", "rawMarkdown": "", "votes": -5, "isDeleted": true, "replies": [ { "id": 584558, "postDate": "2019-07-26T06:45:09.890Z", "content": "It seems obvious that they are not generated but memorized.
", "rawMarkdown": "It seems obvious that they are not generated but memorized.", "votes": 11 } ] }, { "id": 582574, "postDate": "2019-07-23T10:48:34.337Z", "rawMarkdown": "", "votes": -1, "isDeleted": true }, { "id": 574981, "postDate": "2019-07-14T19:34:44.930Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 575110, "postDate": "2019-07-15T04:30:26.057Z", "content": "wow, 30 epochs?
", "rawMarkdown": "wow, 30 epochs?" }, { "id": 575374, "postDate": "2019-07-15T10:52:06.620Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 571216, "postDate": "2019-07-09T10:17:15.093Z", "rawMarkdown": "", "isDeleted": true }, { "id": 570845, "postDate": "2019-07-08T20:40:04.793Z", "rawMarkdown": "", "votes": 2, "isDeleted": true, "replies": [ { "id": 570863, "postDate": "2019-07-08T21:10:00.647Z", "content": "That looks great! What are their MiFID score?
", "rawMarkdown": "That looks great! What are their MiFID score?" }, { "id": 571238, "postDate": "2019-07-09T11:00:31.857Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] }, { "id": 570764, "postDate": "2019-07-08T18:36:23.873Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 569822, "postDate": "2019-07-07T12:07:07.313Z", "rawMarkdown": "", "isDeleted": true }, { "id": 566214, "postDate": "2019-07-01T22:55:04.493Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 564996, "postDate": "2019-06-30T09:37:57.420Z", "content": "*Thanks for sharing *
", "rawMarkdown": "**Thanks for sharing **" } ], "comments": [ { "id": 577850, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-17T06:16:59.420000", "content": "My top dogs for now :D\nThere are a lot of noisy patterns in the images, but the score on public LB is pretty high
\n\n.png?generation=1563344100098383&alt=media)
Because of the confusions in the Leaderboard I am using a new metrics to calculate my position in Leaderboard and progression for past 5-6 days . The metric is Relativistic Dmitry Distance . I calculate my LB position based on your scores :D . Anyway , what is the FID for this dogs ?
", "votes": 3, "replies": [] }, { "id": 577906, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-17T07:24:35.533000", "content": "29.20708 on public LB :)\nI think it can be improved further if I find a way to smooth out these textures.
", "votes": 3, "replies": [] }, { "id": 577938, "author_name": "Alvaro", "author_url": "", "post_date": "2019-07-17T07:59:36.943000", "content": "Awesome! You are definitely wining this competition regardless of the final LB.
", "votes": 0, "replies": [] }, { "id": 577949, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-17T08:18:51.573000", "content": "Holy cow those look good. Any tips you have for us? :)\nFor smoothing the images, using an upsampling + conv somewhere can help idk
", "votes": 4, "replies": [] }, { "id": 577959, "author_name": "François", "author_url": "", "post_date": "2019-07-17T08:30:25.237000", "content": "Is it a fully generative GAN?
", "votes": 0, "replies": [] }, { "id": 577976, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-17T08:44:53.480000", "content": "@theoviel \nI use one of the recent GAN architectures and upsampling is already used there. I have another similar arch, which give me nicer textures, but it's worse at capturing the shapes.\nThe tip is to stabilize the generator weights after a certain point.
\n\n@francoisdubois \nyes, it's an ordinary GAN
", "votes": 8, "replies": [] }, { "id": 577998, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-17T09:05:29.357000", "content": "@dvorobiev Thanks, I'll have a look at that
", "votes": 0, "replies": [] }, { "id": 578116, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-17T11:08:05.723000", "content": "Wow, incredible job Dmitry.
", "votes": 2, "replies": [] }, { "id": 578682, "author_name": "Jar", "author_url": "", "post_date": "2019-07-18T02:50:00.673000", "content": "Great job!
", "votes": 1, "replies": [] }, { "id": 579162, "author_name": "Gaurav Gautam", "author_url": "", "post_date": "2019-07-18T14:48:03.190000", "content": "How many epoches?
", "votes": 0, "replies": [] }, { "id": 579653, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-19T04:19:05.883000", "content": "@grvgtm \n150
", "votes": 4, "replies": [] }, { "id": 580557, "author_name": "SKYap", "author_url": "", "post_date": "2019-07-20T11:00:59.890000", "content": "@dvorobiev what is weight stabilize? is it adding Gaussian Noise or weight clipping?
", "votes": 0, "replies": [] }, { "id": 580577, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-20T12:03:32.103000", "content": "@skyap79 \nNo, the idea was to add some sort of momentum. Like what Adam and other optimizers do to learning rate, but for weights.
", "votes": 4, "replies": [] }, { "id": 581796, "author_name": "zeyu wang", "author_url": "", "post_date": "2019-07-22T12:07:38.120000", "content": "@dvorobiev Have you manually did some data cleaning for the training?
", "votes": 0, "replies": [] }, { "id": 581838, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-22T13:12:11.577000", "content": "@rhonin0 \nNo, I don't use any manually cleaning or image selection, I am too lazy for it :) I used only basic preprocessing to generate these pictures. However recently I found that some automated image selection helps a bit.
", "votes": 2, "replies": [] }, { "id": 581880, "author_name": "Dhruuv Agarwal", "author_url": "", "post_date": "2019-07-22T13:46:21.740000", "content": "Impressive results Dmitry!. Was stuck in scores around 60 for some days. But now, trying different architectures like you suggested.\nHowever, those generally are very deep for the time frame βΉοΈ
", "votes": 0, "replies": [] }, { "id": 581908, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-22T14:21:11.953000", "content": "Thanks!\nYeah, a lot of recent GANs aims at producing higher resolution images. Usually some parameter reduction or layer removal is necessary to make it work under 9 hours constraint.
", "votes": 2, "replies": [] }, { "id": 582397, "author_name": "saewonYang", "author_url": "", "post_date": "2019-07-23T06:17:01.517000", "content": "great job Dmitry :) I wonder what your generator and discriminator loss graph looks like..
", "votes": 0, "replies": [] }, { "id": 582402, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-23T06:29:25.883000", "content": "I haven't tried to plot it yet, but usually the generator loss increases over time, while the discriminator loss stays about the same.
", "votes": 0, "replies": [] }, { "id": 582409, "author_name": "saewonYang", "author_url": "", "post_date": "2019-07-23T06:38:09.467000", "content": "I see. can't wait too see your solution :(
", "votes": 1, "replies": [] }, { "id": 582598, "author_name": "Ashutosh Chandra", "author_url": "", "post_date": "2019-07-23T11:05:14.797000", "content": "wow ! they looks so realistic π
", "votes": 0, "replies": [] } ] }, { "id": 593675, "author_name": "hirune924", "author_url": "", "post_date": "2019-08-07T01:07:41.110000", "content": "My dog ββalso has a face in his buttocks. \n
yeah, it seams so very much...
", "votes": 2, "replies": [] }, { "id": 597748, "author_name": "Kranthi Kumar", "author_url": "", "post_date": "2019-08-12T18:05:39.523000", "content": "lol π
", "votes": 0, "replies": [] } ] }, { "id": 579144, "author_name": "cafeal", "author_url": "", "post_date": "2019-07-18T14:33:54.677000", "content": "My generator unexpectedly generated a green dragon like Rayquaza...\n
", "votes": 13, "replies": [ { "id": 579155, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-18T14:40:23.137000", "content": "Hahaha :)
", "votes": 2, "replies": [] } ] }, { "id": 569049, "author_name": "nekoder", "author_url": "", "post_date": "2019-07-06T00:34:18.630000", "content": "Here is my GAN results.\nSome of them look like cute puppies! However, aliens also seem to be generated ...
\n\n
Updated: with some tweaks, MiFID is now 39.42284 (not memorized)\nI also feel that the quality of dog images has been even better, but it doesn't reflect that much on the MiFID score....maybe MiFID is not a fair metric for visual quality of non-memorized dogs?
\n\n\n\n\n
Previous: Customized GAN, MiFID: 43.07151 (not memorized)\n
Nice work. Looks great
", "votes": 1, "replies": [] }, { "id": 597471, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-08-12T11:42:43.120000", "content": "IMO these pics should get way better score. \nAre you using some brightness adjusments?
", "votes": 3, "replies": [] }, { "id": 597475, "author_name": "Mark Peng", "author_url": "", "post_date": "2019-08-12T11:47:16.877000", "content": "@cdeotte Thanks a lot for your inspiring kernels and posts! They really helped shorten the learning path of GANs.
", "votes": 1, "replies": [] }, { "id": 597554, "author_name": "Mark Peng", "author_url": "", "post_date": "2019-08-12T13:50:54.903000", "content": "@dvorobiev Thanks, I only used random horizontal flip with a low prob., no improvement with other transformations yet.
", "votes": 0, "replies": [] }, { "id": 597559, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-08-12T13:54:35.200000", "content": "But why do your images' colors appear muted? The colors are not as bright as other people's GANs. I think Dmitry's point is that if you can fix the color of your outputted images your score would increase.
", "votes": 2, "replies": [] }, { "id": 597560, "author_name": "Nanashi", "author_url": "", "post_date": "2019-08-12T13:55:24.810000", "content": "amazing job @markpeng :) hope you can keep improving :)
", "votes": 2, "replies": [] }, { "id": 597574, "author_name": "Mark Peng", "author_url": "", "post_date": "2019-08-12T14:24:09.653000", "content": "@cdeotte Good point, will take a look, I also noticed that some of the generated images are a bit dark. Thanks.
", "votes": 0, "replies": [] }, { "id": 597575, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-08-12T14:25:07.180000", "content": "@cdeotte \nyeah, better colors might give a better results. Maybe something wrong with de-normalization...
", "votes": 2, "replies": [] }, { "id": 597613, "author_name": "Mark Peng", "author_url": "", "post_date": "2019-08-12T15:05:45.220000", "content": "Thanks a lot! @dvorobiev @cdeotte, you are right, there is a bug in de-normalization, I will make a new submission to see how it goes after the fix.\nFrom local test MiFID only improved a bit to 43.01698, but the dogs are more colorful now.\n\n
your images are looking much better than your score
", "votes": 0, "replies": [] }, { "id": 599176, "author_name": "Mark Peng", "author_url": "", "post_date": "2019-08-14T15:47:15.490000", "content": "@kranthi9 Feel the same, wonder if MiFID would be a fair metric, it seems to give more advantages to dog diversity (cgan-like and acgan-like) than quality (Just updated the latest examples).
", "votes": 2, "replies": [] }, { "id": 599455, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-08-15T01:45:50.780000", "content": "Yes, diversity is more important than quality. You can see it in my Big GAN kernel here. Setting the random seed's truncation=0.5 generates higher quality images but lower LB. (The latent walk examples use t=0.5). Setting truncation=1.0 generates lower quality images, more diversity, and better LB. (Bottom images use t=1.0).
@cdeotte I also set t=1.0 for truncation trick in my final models. But local MiFID showed that 0.6 is the best. I noticed some variance between public LB and local score, the same notebook didn't finish with the same number of iterations in several runs (using fixed seed and deterministic mode of PyTorch). There seems to have some resource contention in the Kernel environment (I suppose P100 GPU should be dedicated per Kernel). So I finally chose 1.0 as its score is more consistent.
Btw, my GAN is a mini version of StyleGAN with progressive growing from 8x8 to 64x64, learning rate is incremental and fade-in phase size is higher in lower resolution. The mapping network is 8 128-dim FCs. Tried to extend it with conditional labels (CGAN-like) and auxiliary classifier (ACGAN-like), but the generated images could not converge and mostly out of control. For some reason, spectral normalization slowed StyleGAN discriminator down a lot. Also experimented other architectures but none of them has done better than StyleGAN (not yet tried BigGAN).
Will share more after private LB announcement.
", "votes": 1, "replies": [] } ] }, { "id": 585472, "author_name": "Keuh20", "author_url": "", "post_date": "2019-07-27T14:18:28.647000", "content": "Making a dog generator, ending with a goat generator.\n
My network started to kill the dogs that don't look good.
\n\n
Lol
", "votes": 0, "replies": [] }, { "id": 591735, "author_name": "timetraveller", "author_url": "", "post_date": "2019-08-04T07:04:03.627000", "content": "\nsame here, GAN violence is spreading.
My cute cerberus\n
thats awesome clarity :) 22+ is one of the best MFiD with generalized GAN that I have heard from anyone in this comp.
", "votes": 1, "replies": [] }, { "id": 591903, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-08-04T12:56:23.893000", "content": "I usually have something like this:
\n\n
totally agree lol
", "votes": 0, "replies": [] }, { "id": 597179, "author_name": "tkato", "author_url": "", "post_date": "2019-08-12T01:10:21.393000", "content": "They might be an origin of kerberos π€
\n\n
Not a two-headed dog, but just two dogs taking a photo together.
\n\n
my best score images generated by GAN\n
thank you so much @tikutiku , amazing images!
", "votes": 0, "replies": [] } ] }, { "id": 574333, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-13T16:32:23.407000", "content": "Images from my Dog Breed CGAN here\n\n\n
Amazing work Chris, I have played with CGANs quite a bit lately but did not get anything close to your results. I'll get through your code and hopefully find where I did a mistake !
", "votes": 1, "replies": [] }, { "id": 574404, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-13T19:16:17.930000", "content": "Thanks Theo. It took me a while to get it to work too. CGANs have proven to be more difficult than basic GANs. I had to try different architectures, different regularization, and different D/G learning rates. With basic GANs it feels like everything just works. Right now my CGAN is scoring LB 100 but I feel it has potential to do much better.
", "votes": 1, "replies": [] }, { "id": 574412, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-13T19:31:59.633000", "content": "They are very tricky & painful to configure.\nA good usage of dog labels is indeed the key to beat the 60~ threshold I believe. Well at least I hope so.
", "votes": 1, "replies": [] } ] }, { "id": 579184, "author_name": "Luca Massaron", "author_url": "", "post_date": "2019-07-18T15:31:52.713000", "content": "Our dogs reached 62.77, @galeone and me are working to improve our model in order to get better scores.\n(some dogs are still a bit scary: HP Lovecraft would have used them for some of his tales XD)
\n\n\n
Here is some of my results generated by vanilla GAN. They look very ugly. But the LB is 44.05636. I don't know what to say......\n
Well, they look very much like dogs :) .. So the classifier picks it up properly and low LB score . But why are they little zoomed and patchy in color ?
", "votes": 0, "replies": [] }, { "id": 574920, "author_name": "baigou", "author_url": "", "post_date": "2019-07-14T17:35:27.850000", "content": "@phoenix9032 I crop using the box annotations. I'm sure I have rightly denormed the generated samples, so the color should be ok. But I really think they are so ugly.
", "votes": 1, "replies": [] }, { "id": 574996, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-14T20:15:36.987000", "content": "I had some weird colors as well recently.\nIf you are on PyTorch, it can be because you set your model on eval mode, which changes the behaviour of the batch size. Otherwise, make sure your batch size is big enough. Also, make sure your pictures are scaled between 0 and 1, but I believe you did this already because unscaled pictures score a lot worse.
\n\nEDIT : These tricks are for image generation, rescaling is done after generating images
", "votes": 2, "replies": [] }, { "id": 575059, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-15T02:06:35.463000", "content": "You mean batchsize while training the network right ?
", "votes": 0, "replies": [] }, { "id": 575254, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-15T08:09:33.547000", "content": "Nope when generating pictures.
", "votes": 0, "replies": [] }, { "id": 575294, "author_name": "Hitesh Gorana", "author_url": "", "post_date": "2019-07-15T08:53:32.207000", "content": "> make sure your pictures are scaled between 0 and 1
\n\nI thought we are using tanh so it's better to scale images between [-1 1]\nwhy between [0,1]?
I meant for computing the metric, not mifid, as chris mentionned in a separated thread :)
", "votes": 1, "replies": [] }, { "id": 575355, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-15T10:34:52.083000", "content": "I got 68 while i tried MFID in my kernel privately .. Got so excited and run the same next time and submitted .. got 84 MFID . Such a heartbreak :(
", "votes": 0, "replies": [] } ] }, { "id": 574761, "author_name": "Christian", "author_url": "", "post_date": "2019-07-14T12:51:59.670000", "content": "Unfortunately I missed that \"kernel only\" and trained my customized styleGAN on my RTX 2080Ti, so I can't submit the 10k images for getting scores, anyway here are some results:\n
Wow, they look great. Have you computed their MiFID? (with code here)
", "votes": 1, "replies": [] }, { "id": 574771, "author_name": "Christian", "author_url": "", "post_date": "2019-07-14T13:17:09.717000", "content": "Thanks a lot for the code, here is my result:\ndistance= 0.1981568680329205\nFID_public: 49.61624316188124 distance_public: 1 multiplied_public: 49.616243161880746
", "votes": 2, "replies": [] }, { "id": 574788, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-14T13:42:34.833000", "content": "Great score. Are you cropping to the bounding boxes or training with the full images?
", "votes": 0, "replies": [] }, { "id": 574800, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-14T13:54:14.850000", "content": "I also tried styleGAN but didn't achieve any meaningful results.\nHow much images do you use per each resolution phase?
", "votes": 0, "replies": [] }, { "id": 574829, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-14T14:39:19.703000", "content": "Nice doggos, how long did the training took though?
", "votes": 0, "replies": [] }, { "id": 574833, "author_name": "Christian", "author_url": "", "post_date": "2019-07-14T14:44:29.233000", "content": "Just used all the data as it is. Did something similar not long ago with watches, that's why I basically just reused my existing code.
\n\n500.000 for each resolution. Training took ~20h on a RTX 2080 Ti.
", "votes": 2, "replies": [] }, { "id": 574850, "author_name": "robga", "author_url": "", "post_date": "2019-07-14T15:07:09.660000", "content": "What do you feed the StyleGAN to create an image? My understanding was you need to feed a real image in. If so, this wouldnβt be acceptable according to the rules.
\n\nOf course you could always use an output from a memoriser gan as input ..sarcasm..
", "votes": 1, "replies": [] }, { "id": 574875, "author_name": "Christian", "author_url": "", "post_date": "2019-07-14T15:55:17.173000", "content": "You feed a vector, so no real image involved.
", "votes": 1, "replies": [] } ] }, { "id": 566435, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-02T06:57:36.790000", "content": "Another update:\n
I have the feeling that they look a bit more like dogs now, but that could just be me.
", "votes": 7, "replies": [ { "id": 566447, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-02T07:13:48.677000", "content": "Wow, those look very good. I'd be interested in running kNN on some of those to see what the closest training image is. Are those completely synthesized or is your network reproducing training images? For example, bottom row fourth from right, dog looking at us.
", "votes": 0, "replies": [] }, { "id": 566576, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-02T10:39:31.697000", "content": "I think they are \"completly\" synthesized -- if you could ever say that about a GAN.
\n\nThat's the output of the MiFID calculation (to be fair, from one model earlier):\n\n100%|ββββββββββ| 201/201 [00:36<00:00, 5.55it/s]\n100%|ββββββββββ| 412/412 [01:09<00:00, 4.27it/s]\nm1,m2 shape= ((2048,), (2048,)) s1,s2= ((2048, 2048), (2048, 2048))\nstarting calculating FID\ncovmean.shape= (2048, 2048)\ndone with FID, starting distance calculation\nd.shape= (10000, 20579)\nnp.min(d, axis=1).shape= (10000,)\ndistance= 0.2254516669552895\nFID_public: 83.36472131619365 distance_public: 1 multiplied_public: 83.36472131619281\n
To my understanding distance_public is the measurement for memorizing abilities.\nAs this is my first ever GAN experience, feel free to correct me where i am wrong.
Checking the nearest neighbors in train is definetly a thing i wan't to try out when i find some spare time. \nIf you want I can make the output public so you can check it.
", "votes": 0, "replies": [] }, { "id": 566594, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-02T10:57:20.550000", "content": "The cosine memorization distance is distance= 0.2254516669552895. Since that is above threshold = 0.1, it gets set to 1. But that distance is easy to fool :-) If you rescale and crop images, then it thinks your reformatted image has a large distance from original images.
This is my concern. Someone can design a GAN which simply scales and crops original images. My experiments show that it is easy to score under LB 10 (and distance stays over threshold=0.1).
\n\nThe reason i asked about yours is not to undermine your work but rather praise it. If that dog's face in the bottom row doesn't exist on any training images, then I am super impressed with GANs.
", "votes": 1, "replies": [] }, { "id": 566615, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-02T11:12:34.363000", "content": "Regarding the memorization: This article was a good read about it and the reason I said:
\n\n\n", "votes": 1, "replies": [] }, { "id": 566678, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-02T12:58:30.680000", "content": "\"completly\"` synthesized -- if you could ever say that about a GAN.
\n
Yes, that article is exactly what I'm talking about. It's so hard to tell whether a GAN is simply cropping, interpolating, and concatenating training images or whether it learned to draw dogs from scratch.
\n\nThe tests that they mention in that article are what I wanted to do with your GAN. I wanted to search for similar images (between output and train) and I wanted to perform that test they describe. You should perform that test on your GAN.
\n\nTake two different random seeds (that you feed into your generative network). Then interpolate 48 seeds in between on the straight line between them. Finally have your generative network create the 50 images and see how your network transitions from one seed to another.
\n\nI can't wait until I get a GAN up and running. There are so many cool tests I want to try out.
", "votes": 1, "replies": [] }, { "id": 566685, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-02T13:07:22.447000", "content": "Here are the output images for whatever you want to try on them. \nSadly, the exact dog you mentioned is not in there, as this was an intermediate picture (not the last epoch).
\n\nRegarding the transition part, that's definetly on my TODO list. From what I have heard from you in the last competitions, you probably don't want to team-up, do you? ;)
", "votes": 1, "replies": [] }, { "id": 567436, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-03T14:16:04.020000", "content": "I posted a kernel here exploring the idea of memorization versus generalization. I believe your net is generalizing. In my kernel, I show a test that you can perform on your GAN. You can make a sequence of images from one latent space seed to another and watch how your GAN generalizes.
", "votes": 0, "replies": [] } ] }, { "id": 572144, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-10T14:30:10.367000", "content": "It's funny how nice this GAN can capture small dogs shapes and details and how awful it for large breeds :)
\n\n
These images get 45.39 locally
", "votes": 8, "replies": [ { "id": 572149, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-10T14:36:37.100000", "content": "Great job! Your dog faces are very good but the bodies are not so good. I wonder why that is. If you're using a conditional GAN perhaps it's learning the different classes by getting the faces correct but ignoring learning the bodies.
", "votes": 0, "replies": [] }, { "id": 572153, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-10T14:40:13.317000", "content": "Yeah, this one is using labels, and each image corresponds to a different class
\n\nMaybe I should pick only easy classes :D
", "votes": 1, "replies": [] }, { "id": 572188, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-10T15:43:06.877000", "content": "You can also change the \"classes\". The \"classes\" are just attributes that you want your GAN to understand. For example you could separate all dogs into the 4 groups: no hair, short hair, straight hair, and curly hair. Or you could do 3 groups: dog facing left, dog facing center, dog facing right.
\n\nYou can also input both \"classes\" at once with Keras call Model(inputs=[seed,hair,facing], outputs=x). Then if you want a \"short hair\" dog \"facing left\", you call generator(Z, \"short hair\", \"facing left\").
If your \"classes\" are more general it won't memorize faces and ignore bodies.
", "votes": 3, "replies": [] }, { "id": 572215, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-10T16:18:02.253000", "content": "Thanks, maybe I will try it later :)
", "votes": 1, "replies": [] } ] }, { "id": 569828, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-07T12:14:13.490000", "content": "
My intermediate results\nStill training...
", "votes": 8, "replies": [ { "id": 569832, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-07T12:18:00.413000", "content": "They are so good ..with lots of variety !! A big thumbs up
", "votes": 0, "replies": [] }, { "id": 569860, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-07T13:16:59.933000", "content": "Wow, impressive!
", "votes": 0, "replies": [] }, { "id": 570092, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-07T19:04:55.007000", "content": "Thanks!\nI am a bit sad there are still a lot of wishy washy images. I got 55.55 locally and 53.77655 on public LB\nI guess there is a lot of room for improvement...
", "votes": 1, "replies": [] }, { "id": 570094, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-07T19:06:39.523000", "content": "That's a great score. I think the best conventional GANs in this comp are currently scoring around 45.
", "votes": 0, "replies": [] }, { "id": 570143, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-07T21:31:24.040000", "content": "Very impressive results.
\n\nI am still struggeling with non converging or slowly training models, but your results give me hope that it's possible.
", "votes": 0, "replies": [] }, { "id": 570456, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-07-08T10:14:09.283000", "content": "That's a good guess @cdeotte, 42 is the score to beat now, congratz @dvorobiev !
", "votes": 1, "replies": [] }, { "id": 570479, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-08T10:59:32.487000", "content": "Thanks :) \nSolution for 42 is a lot heavier than previous. I almost ran out of time. I guess, it will be hard to improve with current arch, unless I find a way to cut off some computations
", "votes": 2, "replies": [] } ] }, { "id": 564357, "author_name": "Vladislav Bakhteev", "author_url": "", "post_date": "2019-06-29T10:07:32.660000", "content": "Looks scary\n
very scary !!!,
", "votes": 0, "replies": [] }, { "id": 564419, "author_name": "Serigne ", "author_url": "", "post_date": "2019-06-29T11:50:23.627000", "content": "Remind me this ;)
", "votes": 7, "replies": [] } ] }, { "id": 563926, "author_name": "Andrey Lukyanenko", "author_url": "", "post_date": "2019-06-28T19:31:29.243000", "content": "The first attempt went wrong!\n
A bit better\n
Big GAN 256 scores LB 15 !! Kaggle kernel/notebook here.\n
Very nice, thanks for sharing it!
", "votes": 1, "replies": [] }, { "id": 599458, "author_name": "Adiachen", "author_url": "", "post_date": "2019-08-15T02:03:56.827000", "content": "Awesome!!
", "votes": 1, "replies": [] }, { "id": 599508, "author_name": "saewonYang", "author_url": "", "post_date": "2019-08-15T04:11:12.777000", "content": "nice work chris!
", "votes": 1, "replies": [] } ] }, { "id": 578684, "author_name": "Jar", "author_url": "", "post_date": "2019-07-18T02:50:59.403000", "content": "66.9 LB
\n\n
I tried training only on one breed of Maltese Dog\nhere is the result after 30 minutes of training. And gif of training in progress.\n
That's awesome!! Fantastic! You should try building a CGAN. When you train your Generator and Discriminator, give it a breed of dog. Then after your GAN is trained, you feed a random Z plus a random Breed (1 to 120) into your Generator.
", "votes": 1, "replies": [] }, { "id": 571376, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-09T14:37:50.473000", "content": "Thank you Chris! I'm currently giving breeds info to my D and G networks. It gave a significant boost when I implemented it. Now I'm thinking on adding classifier head to D network to force it to learn better features.
", "votes": 1, "replies": [] }, { "id": 571419, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-09T15:19:50.607000", "content": "One question , I have read that MFID uses variety of images also as parameter . In that case , if we generate one breed of dogs as output . Even with good image it can give lower MFID ?
", "votes": 0, "replies": [] }, { "id": 571440, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-09T15:37:54.177000", "content": "I think so. Distribution of Inception latent features will be different for only one breed.
", "votes": 0, "replies": [] }, { "id": 571622, "author_name": "Chris Deotte", "author_url": "", "post_date": "2019-07-09T21:05:46.930000", "content": "\n\n\nOne question , I have read that MFID uses variety of images also as parameter . In that case , if we generate one breed of dogs as output . Even with good image it can give lower MFID ?
\n
Yes, Nirjhar. But learning breeds will give you better variety. You train the Generator by giving it a breed number (1-120) and a random seed Z. Then when you create your 10,000 dogs for Kaggle, you input a random Z and random breed number from 1-120. This way your 10,000 images are guarenteed to have approx equal distribution of all 120 breeds represented.
", "votes": 1, "replies": [] }, { "id": 571632, "author_name": "", "author_url": "", "post_date": "2019-07-09T21:34:34.083000", "content": "", "votes": 2, "replies": [] }, { "id": 571864, "author_name": "Dmitry Vorobiev", "author_url": "", "post_date": "2019-07-10T06:59:04.940000", "content": "\n\n\nYes, Nirjhar. But learning breeds will give you better variety. You train the Generator by giving it a breed number (1-120) and a random seed Z. Then when you create your 10,000 dogs for Kaggle, you input a random Z and random breed number from 1-120. This way your 10,000 images are guarenteed to have approx equal distribution of all 120 breeds represented.
\n
Yeah, I can confirm that. I feed all 120 labels uniformly when generating images. I also use it to view intermediate results during training, and since the label for each cell of the plot is constant, I can see a progress of my net on each breed :) Usually I visualize only the first 32 of them.
", "votes": 3, "replies": [] } ] }, { "id": 568886, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-05T16:03:19.627000", "content": "
Real GAN? I am deeply impressed If this is done in 9h Kernel time. Good job!\nWhat is the FID Score?
", "votes": 0, "replies": [] }, { "id": 568892, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-05T16:10:39.947000", "content": "It's real DCGAN after only ~2 hours of training on local machine without any external data, now working on moving it to the kernel. I haven't computed FID scores so far, working on task in general, not metric
", "votes": 2, "replies": [] }, { "id": 568894, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-05T16:12:18.717000", "content": "New leading score for real GANs incoming ;)\nWell done!
", "votes": 0, "replies": [] }, { "id": 568896, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-05T16:14:42.337000", "content": "I already have a kernel with good looking dogs, but thanks to the bug I generated 10048 images instead of 10000 and fail to submit it))))
", "votes": 0, "replies": [] }, { "id": 568903, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-05T16:22:31.747000", "content": "\n\n\n\n\n\n\n\n\n\n\n
Awesome job Kostya. They look great. From your images, it looks like your GAN first builds a face and then places it on a body. Nice design.
", "votes": 2, "replies": [] }, { "id": 568962, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-05T19:02:44.570000", "content": "May I ask, if you use a selection of the \"best\" dogs to train your model?
", "votes": 1, "replies": [] }, { "id": 569012, "author_name": "Kostya Atarik", "author_url": "", "post_date": "2019-07-05T21:32:13.133000", "content": "No, I train on all dogs available. And what do you mean by a selection of the \"best\" dogs?
", "votes": 1, "replies": [] }, { "id": 569021, "author_name": "Pascal Pfeiffer", "author_url": "", "post_date": "2019-07-05T21:49:10.527000", "content": "Ok, I was wondering if it boosts model performance if we curate the images first. Like removing pictures with watermarks, low resolution, hardly visible dogs etc...
\n\nOr making subsamples of front-facing-dogs, dogs visible from the side ...
", "votes": 1, "replies": [] }, { "id": 572243, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-07-10T17:14:20.317000", "content": "i like how the last but one doggy has bunny teeth
", "votes": 2, "replies": [] } ] }, { "id": 566188, "author_name": "Amw5g", "author_url": "", "post_date": "2019-07-01T21:58:43.857000", "content": "After a few false starts, my first DCGAN at the end of its run. If you squint really hard, a few of them look dog-esque.
Here are two of my GANs doing latents walk between dog A and dog B. Notice how the top GAN generalizes and invents a 3rd new dog as the middle point (in step 4 and 5). The bottom GAN does a basic pixel blend with awkward middle images.
\n\n
For comparison, here's a Memorizer GAN's pixel blend of the same 2 images:\n
few good images, but almost scary!!π
\n\n
at least some of them look \"doglike\"
", "votes": 6, "replies": [] }, { "id": 599392, "author_name": "Theo Viel", "author_url": "", "post_date": "2019-08-14T22:56:07.010000", "content": "
woof
", "votes": 3, "replies": [] }, { "id": 598676, "author_name": "Vitaly Bondar", "author_url": "", "post_date": "2019-08-13T21:36:00.080000", "content": "\nThe dog we deserved π
\nwas making a dog generator\ngot a patrick from spongebob generator
\n",
"569049": "Here is my GAN results.\nSome of them look like cute puppies! However, aliens also seem to be generated ...\n\n",
"597226": "Updated: with some tweaks, MiFID is now 39.42284 (not memorized)\nI also feel that the quality of dog images has been even better, but it doesn't reflect that much on the MiFID score....maybe MiFID is not a fair metric for visual quality of non-memorized dogs?\n\n\n\n\n\n\nPrevious: Customized GAN, MiFID: 43.07151 (not memorized)\n\n\n",
"585472": "Making a dog generator, ending with a goat generator.\n \n",
"584788": "My network started to kill the dogs that don't look good.\n\n",
"589144": "My cute cerberus\n",
"565489": "my best score images generated by GAN\n",
"574333": "Images from my Dog Breed CGAN [here][1]\n\n\n\n\n[1]: https://www.kaggle.com/cdeotte/dog-breed-cgan",
"579184": "Our dogs reached 62.77, @galeone and me are working to improve our model in order to get better scores.\n(some dogs are still a bit scary: HP Lovecraft would have used them for some of his tales XD)\n\n\n\n",
"574903": "Here is some of my results generated by vanilla GAN. They look very ugly. But the LB is 44.05636. I don't know what to say......\n\n",
"574761": "Unfortunately I missed that \"kernel only\" and trained my customized styleGAN on my RTX 2080Ti, so I can't submit the 10k images for getting scores, anyway here are some results:\n\n",
"566435": "Another update:\n\n\nI have the feeling that they look a bit more like dogs now, but that could just be me. ",
"572144": "It's funny how nice this GAN can capture small dogs shapes and details and how awful it for large breeds :)\n\n\n\nThese images get 45.39 locally",
"569828": "\n\nMy intermediate results\nStill training...",
"564357": "Looks scary\n\n",
"563926": "The first attempt went wrong!\n",
"599415": "Big GAN 256 scores LB 15 !! Kaggle kernel/notebook [here][1].\n\n\n[1]: https://www.kaggle.com/cdeotte/big-gan-256-lb-15",
"578684": "66.9 LB\n\n\n",
"571365": "I tried training only on one breed of [Maltese Dog](http://vision.stanford.edu/aditya86/ImageNetDogs/n02085936.html)\nhere is the result after 30 minutes of training. And gif of training in progress.\n\n",
"568886": "\n",
"566188": "After a few false starts, my first DCGAN at the end of its run. If you squint really hard, a few of them look dog-esque.\n",
"573897": "Here are two of my GANs doing latents walk between dog A and dog B. Notice how the top GAN generalizes and invents a 3rd new dog as the middle point (in step 4 and 5). The bottom GAN does a basic pixel blend with awkward middle images.\n## Generalizing GAN\n\n## Memorizing GAN\nFor comparison, here's a Memorizer GAN's pixel blend of the same 2 images:\n",
"572632": "few good images, but almost scary!!π \n\n\n",
"565772": "\n\nat least some of them look \"doglike\"",
"599392": "\n\nwoof",
"598676": "\nThe dog we deserved π ",
"596684": "\nwas making a dog generator\ngot a patrick from spongebob generator",
"594009": "Some of the best dogs I got so far : \n  \n   \n\nStill have a lot of messy images, so the LB score is not so great (~82)",
"572996": "After 100 epochs [**DCGAN**]\n\n",
"572062": "animated collapse of one of my local runs",
"569605": "In my post [Animated GANs - See the latent walk evolution](https://www.kaggle.com/c/generative-dog-images/discussion/98853#latest-569598) I explain how to generate this animation using the kernel [Animation - GAN Latent Walk](https://www.kaggle.com/jesucristo/animation-gan-latent-walk)\n\nI'm looking forward to see your results animated too :)\n\n### Some results:\n\n\n\n",
"567791": "I am very sorry I washed my dogs , now they looks like abstract art .Fast.ai WGAN\n\n",
"593900": "Looks like a famous fresco...\n\n",
"564564": "\n\nAre they looking like dogs or am I just too optimistic? ",
"588189": "wow... is this melted panda...?\n\n",
"584768": "Our GAN successfully generated ghost-like skeleton doggos : \n\n",
"572445": "New best dogs using the baseline [Introducing DCGAN Dogs Images](https://www.kaggle.com/jesucristo/introducing-dcgan-dogs-images) + tricks from [All you need is GAN Hacks](https://www.kaggle.com/c/generative-dog-images/discussion/98595#latest-570614)\n.png?generation=1562804966674757&alt=media)\n\n**LB 102** (see v39)",
"571245": "my best scoring kernel so far with 63\n",
"571241": "\n\n\nSome of my monsters. Score 77.",
"569801": "Real GAN >\n\n",
"565470": "My favourite creation\n\n",
"599242": "My best run:\n\n\nI am not sure why there is grey, but I think it is a result of overfitting because when I train a bit less the images are coloured normally.\n",
"599192": "I just have to share this one.\n\n",
"596074": "\nGAN dog looks like a goat",
"575270": "Some dogs generated with a race conditional GAN. Score is 77~. Not really close from my best models but at least it's working.\n\n\n",
"570248": "TensorFlow/Keras DCGAN after 35 epochs (22000 batches):\n\n",
"599386": "### GOOD LUCK!\n\n",
"571194": "\n",
"569025": "\n\nI think i finally might be onto something here ...",
"568990": "I like the dog in the top row, 3rd from right.\n\n\nSadly, I see an increase in my loss variances that I can't get rid off. Any tips? That might be the problem why the dogs quality doesn't increase much after 50k iterations for me. \n",
"568837": "I made images walking in the latent space\n\nHere, I picked up two good images manually (most left one and most right one)",
"566249": "Produced by adding variance to an autoencoder's latent layer in [this kernel][1]. However after reading the new rules clarification [here][2], I don't believe these images are legal. But all is not lost, I'll save the decoder (as a generative network), add a discriminative network, and turn it into a GAN ! \n\n\n\n[1]: https://www.kaggle.com/cdeotte/dog-autoencoder\n[2]: https://www.kaggle.com/c/generative-dog-images/discussion/98183",
"565265": "Using **[\nIntroducing DCGAN Dogs Images](https://www.kaggle.com/jesucristo/introducing-dcgan-dogs-images)**\n\n",
"599142": "A new breed: Melted-Cheese Dog.\n\n",
"598326": "Some of my dogs. Front-facing dogs are working best, visually.\nSome images are still completely abstract.\nIn a lot of images, I have some sort of blue \"bleeding\", I do not yet understand why. I do not have different processes for the different color channels. Is it something that can happen? My GAN is a bit special/home made so maybe I'm missing a side effect from my code...\n\n\n\n\n\n",
"595985": "hello guys...\n\n",
"595360": "preliminary result by DCGAN with ImageNet classification result\nnothing like dog yet...\n\n\n",
"583451": "I removed Batch Normalization on the discriminator and it gave me 50 Shades of Grey images for a batch size of 50. There are some dogs hidden behind that , but why the Grey curtain ?\n\n",
"570911": "DCGAN with Complement between vectors\n\n\n",
"568840": "@ibraheemmoosa Here, have some nightmare material xD",
"588188": "",
"584412": "",
"582574": "",
"574981": "",
"571216": "",
"570845": "",
"570764": "",
"569822": "",
"566214": "",
"564996": "**Thanks for sharing **"
}
}![https://i.ibb.co/y5n6MV2/2.gif](\"https://i.ibb.co/y5n6MV2/2.gif\"){kind=link}
![https://i.ibb.co/pnv1RdL/1.gif](\"https://i.ibb.co/pnv1RdL/1.gif\"){kind=link}
![https://i.ibb.co/5RNkdK1/2-2.gif](\"https://i.ibb.co/5RNkdK1/2-2.gif\"){kind=link}