![\"Imgur\"](\"https://i.imgur.com/JrkxmC8.png\")
\nhttps://medium.com/syncedreview/biggan-a-new-state-of-the-art-in-image-synthesis-cf2ec5694024BigGan is announced as the new state of the art in the field of generative models. \nA few searchers got amazing realistic pictures with this architecture:
\n\n\nhttps://medium.com/syncedreview/biggan-a-new-state-of-the-art-in-image-synthesis-cf2ec5694024
I have tried to apply BigGan to the competition in the kernel below but I did not manage to get a fair score.
\n\nhttps://www.kaggle.com/francoisdubois/biggan-aaronleong-64
\n\nI got stuck at 140.
\n\n![\"Imgur\"](\"https://i.imgur.com/4WdcEj2.png\")
![\"Imgur\"](\"https://i.imgur.com/fiwROXn.png\")
I think that this model can be very powerful and the problem here is probably due to not being able to run more steps on bigger batches. It took a week for the author of the paper to get their results on 4 GPUs.
\n\nI still wanted to share my kernel to see if anyone could make it work in this context on kernels\n or to hear from anyone who tried it too.
\n\nFor more information here is the paper: \nhttps://arxiv.org/pdf/1809.11096.pdf
", "messages": [ { "id": "583612", "postDate": "07/24/2019 17:48:07", "content": "BigGan is announced as the new state of the art in the field of generative models. \nA few searchers got amazing realistic pictures with this architecture:
\n\n\nhttps://medium.com/syncedreview/biggan-a-new-state-of-the-art-in-image-synthesis-cf2ec5694024
I have tried to apply BigGan to the competition in the kernel below but I did not manage to get a fair score.
\n\nhttps://www.kaggle.com/francoisdubois/biggan-aaronleong-64
\n\nI got stuck at 140.
\n\n
I think that this model can be very powerful and the problem here is probably due to not being able to run more steps on bigger batches. It took a week for the author of the paper to get their results on 4 GPUs.
\n\nI still wanted to share my kernel to see if anyone could make it work in this context on kernels\n or to hear from anyone who tried it too.
\n\nFor more information here is the paper: \nhttps://arxiv.org/pdf/1809.11096.pdf
", "rawMarkdown": "BigGan is announced as the new state of the art in the field of generative models. \nA few searchers got amazing realistic pictures with this architecture:\n\n\nhttps://medium.com/syncedreview/biggan-a-new-state-of-the-art-in-image-synthesis-cf2ec5694024\n\nI have tried to apply BigGan to the competition in the kernel below but I did not manage to get a fair score.\n\nhttps://www.kaggle.com/francoisdubois/biggan-aaronleong-64\n\nI got stuck at 140.\n\n\n\n\n\nI think that this model can be very powerful and the problem here is probably due to not being able to run more steps on bigger batches. It took a week for the author of the paper to get their results on 4 GPUs.\n\nI still wanted to share my kernel to see if anyone could make it work in this context on kernels\n or to hear from anyone who tried it too.\n\nFor more information here is the paper: \nhttps://arxiv.org/pdf/1809.11096.pdf", "votes": null }, { "id": "583623", "postDate": "07/24/2019 17:57:59", "content": "Thank you so much @francoisdubois .. I was not able to implement PROGAN .. But I am playing around with some of the functions like PixelNorm , modified conv and deconv layers etc ..
", "rawMarkdown": "Thank you so much @francoisdubois .. I was not able to implement PROGAN .. But I am playing around with some of the functions like PixelNorm , modified conv and deconv layers etc ..", "votes": null }, { "id": "583668", "postDate": "07/24/2019 19:25:15", "content": "I've been playing with StyleGAN today. Honestly, It's giving me a hard time to grasp it. But I read someone applied it for this competition and that's why I am inclined to do it too.
", "rawMarkdown": "I've been playing with StyleGAN today. Honestly, It's giving me a hard time to grasp it. But I read someone applied it for this competition and that's why I am inclined to do it too.", "votes": null }, { "id": "583852", "postDate": "07/25/2019 05:04:34", "content": "I also tried StyleGAN at the beginning of this competition kernel. But it's hard to make a decent result in 9 hours training. Reducing the number of channels, turning to conditional GAN just gave me around 90~100 LB. I think Dogdataset is more difficult to train by StyleGAN compared with CelebA. In my experiments, even the initial results at 8x8 or 16x16 generated by StyleGAN are not too meaningful.
", "rawMarkdown": "I also tried StyleGAN at the beginning of this competition [kernel](https://www.kaggle.com/quangbk/stylegan-dog-pytorch). But it's hard to make a decent result in 9 hours training. Reducing the number of channels, turning to conditional GAN just gave me around 90~100 LB. I think Dogdataset is more difficult to train by StyleGAN compared with CelebA. In my experiments, even the initial results at 8x8 or 16x16 generated by StyleGAN are not too meaningful.", "votes": null }, { "id": "583863", "postDate": "07/25/2019 05:22:46", "content": "So far I tried a few GANs and either I am making some big mistakes or the research papers mostly work with MNIST or CELEBA or similar images (front facing cats and dogs ) to log their findings by running the experiments for couple of days . Not sure if there is a implicit bias introduced in those method of experiments and makes those architectures very specific to those datasets . \nI feel the original DCGAN is more generalized and works for most of the unknown datasets with lesser number of tweaks .
", "rawMarkdown": "So far I tried a few GANs and either I am making some big mistakes or the research papers mostly work with MNIST or CELEBA or similar images (front facing cats and dogs ) to log their findings by running the experiments for couple of days . Not sure if there is a implicit bias introduced in those method of experiments and makes those architectures very specific to those datasets . \nI feel the original DCGAN is more generalized and works for most of the unknown datasets with lesser number of tweaks .", "votes": null }, { "id": "583880", "postDate": "07/25/2019 05:57:43", "content": "I tried to implement BigGan and I failed, this kernel will very much useful for me thanks for sharing @francoisdubois
", "rawMarkdown": "I tried to implement BigGan and I failed, this kernel will very much useful for me thanks for sharing @francoisdubois", "votes": null }, { "id": "584015", "postDate": "07/25/2019 10:06:42", "content": "I'm unfortunatly still far from getting fair results
", "rawMarkdown": "I'm unfortunatly still far from getting fair results", "votes": null }, { "id": "584019", "postDate": "07/25/2019 10:09:38", "content": "I am also getting my best results with DGGAN. Any other tries didn't end up well.
", "rawMarkdown": "I am also getting my best results with DGGAN. Any other tries didn't end up well.", "votes": null }, { "id": "584602", "postDate": "07/26/2019 08:01:41", "content": "I tried to run bigGan as well, and I failed due to out of GPU memory issue.
", "rawMarkdown": "I tried to run bigGan as well, and I failed due to out of GPU memory issue.", "votes": null }, { "id": "584613", "postDate": "07/26/2019 08:16:13", "content": "I had to reduce the batch size to make it work
", "rawMarkdown": "I had to reduce the batch size to make it work", "votes": null } ], "comments": [ { "id": 583623, "author_name": "phoenix9032", "author_url": "", "post_date": "07/24/2019 17:57:59", "content": "Thank you so much @francoisdubois .. I was not able to implement PROGAN .. But I am playing around with some of the functions like PixelNorm , modified conv and deconv layers etc ..
", "votes": null, "replies": [] }, { "id": 583668, "author_name": "timetraveller98", "author_url": "", "post_date": "07/24/2019 19:25:15", "content": "I've been playing with StyleGAN today. Honestly, It's giving me a hard time to grasp it. But I read someone applied it for this competition and that's why I am inclined to do it too.
", "votes": null, "replies": [ { "id": 583852, "author_name": "quangbk", "author_url": "", "post_date": "07/25/2019 05:04:34", "content": "I also tried StyleGAN at the beginning of this competition kernel. But it's hard to make a decent result in 9 hours training. Reducing the number of channels, turning to conditional GAN just gave me around 90~100 LB. I think Dogdataset is more difficult to train by StyleGAN compared with CelebA. In my experiments, even the initial results at 8x8 or 16x16 generated by StyleGAN are not too meaningful.
", "votes": null, "replies": [] } ] }, { "id": 583863, "author_name": "phoenix9032", "author_url": "", "post_date": "07/25/2019 05:22:46", "content": "So far I tried a few GANs and either I am making some big mistakes or the research papers mostly work with MNIST or CELEBA or similar images (front facing cats and dogs ) to log their findings by running the experiments for couple of days . Not sure if there is a implicit bias introduced in those method of experiments and makes those architectures very specific to those datasets . \nI feel the original DCGAN is more generalized and works for most of the unknown datasets with lesser number of tweaks .
", "votes": null, "replies": [ { "id": 584019, "author_name": "francoisdubois", "author_url": "", "post_date": "07/25/2019 10:09:38", "content": "I am also getting my best results with DGGAN. Any other tries didn't end up well.
", "votes": null, "replies": [] } ] }, { "id": 583880, "author_name": "joelhanson", "author_url": "", "post_date": "07/25/2019 05:57:43", "content": "I tried to implement BigGan and I failed, this kernel will very much useful for me thanks for sharing @francoisdubois
", "votes": null, "replies": [ { "id": 584015, "author_name": "francoisdubois", "author_url": "", "post_date": "07/25/2019 10:06:42", "content": "I'm unfortunatly still far from getting fair results
", "votes": null, "replies": [] } ] }, { "id": 584602, "author_name": "wtan0019", "author_url": "", "post_date": "07/26/2019 08:01:41", "content": "I tried to run bigGan as well, and I failed due to out of GPU memory issue.
", "votes": null, "replies": [ { "id": 584613, "author_name": "francoisdubois", "author_url": "", "post_date": "07/26/2019 08:16:13", "content": "I had to reduce the batch size to make it work
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "583612": "BigGan is announced as the new state of the art in the field of generative models. \nA few searchers got amazing realistic pictures with this architecture:\n\n\nhttps://medium.com/syncedreview/biggan-a-new-state-of-the-art-in-image-synthesis-cf2ec5694024\n\nI have tried to apply BigGan to the competition in the kernel below but I did not manage to get a fair score.\n\nhttps://www.kaggle.com/francoisdubois/biggan-aaronleong-64\n\nI got stuck at 140.\n\n\n\n\n\nI think that this model can be very powerful and the problem here is probably due to not being able to run more steps on bigger batches. It took a week for the author of the paper to get their results on 4 GPUs.\n\nI still wanted to share my kernel to see if anyone could make it work in this context on kernels\n or to hear from anyone who tried it too.\n\nFor more information here is the paper: \nhttps://arxiv.org/pdf/1809.11096.pdf", "583623": "Thank you so much @francoisdubois .. I was not able to implement PROGAN .. But I am playing around with some of the functions like PixelNorm , modified conv and deconv layers etc ..", "583668": "I've been playing with StyleGAN today. Honestly, It's giving me a hard time to grasp it. But I read someone applied it for this competition and that's why I am inclined to do it too.", "583852": "I also tried StyleGAN at the beginning of this competition [kernel](https://www.kaggle.com/quangbk/stylegan-dog-pytorch). But it's hard to make a decent result in 9 hours training. Reducing the number of channels, turning to conditional GAN just gave me around 90~100 LB. I think Dogdataset is more difficult to train by StyleGAN compared with CelebA. In my experiments, even the initial results at 8x8 or 16x16 generated by StyleGAN are not too meaningful.", "583863": "So far I tried a few GANs and either I am making some big mistakes or the research papers mostly work with MNIST or CELEBA or similar images (front facing cats and dogs ) to log their findings by running the experiments for couple of days . Not sure if there is a implicit bias introduced in those method of experiments and makes those architectures very specific to those datasets . \nI feel the original DCGAN is more generalized and works for most of the unknown datasets with lesser number of tweaks .", "583880": "I tried to implement BigGan and I failed, this kernel will very much useful for me thanks for sharing @francoisdubois", "584015": "I'm unfortunatly still far from getting fair results", "584019": "I am also getting my best results with DGGAN. Any other tries didn't end up well.", "584602": "I tried to run bigGan as well, and I failed due to out of GPU memory issue.", "584613": "I had to reduce the batch size to make it work" }, "source": "meta" }