![\"image\"](\"http://playagricola.com/Kaggle/dogs681519.png\")
Hi there!
\n\nThe notebook for my solution is available here : https://www.kaggle.com/alvaroma/dcgan-lb-30-65
\n\nKeypoints:
\n\nGood luck for all participants!
", "messages": [ { "id": "600294", "postDate": "08/15/2019 23:44:17", "content": "Hi there!
\n\nThe notebook for my solution is available here : https://www.kaggle.com/alvaroma/dcgan-lb-30-65
\n\nKeypoints:
\n\nGood luck for all participants!
", "rawMarkdown": "Hi there!\n\nThe notebook for my solution is available here : [https://www.kaggle.com/alvaroma/dcgan-lb-30-65](https://www.kaggle.com/alvaroma/dcgan-lb-30-65)\n\nKeypoints:\n\n- Pytorch DCGan\n- NZ 64, Batch 32, LRG 0.0004, LRD 0.0008\n- Cropped images, selecting only those with ratio 1:50\n- Transformations: Flip Horizontal, Random Crop\n- Early stop at best FID\n- Select best random images\n\nGood luck for all participants!", "votes": null }, { "id": "600306", "postDate": "08/16/2019 00:32:39", "content": "Wow, simple and effective. Congratulations. Great job.
\n\n(Do you mean to say, select only with ratio less than 1 : 1.5 or less?)
Thanks Chris. The idea was selecting \"squared\" images to better fit in 64x64, so W / H < 1.50 (or H / W if H > W). I tried with different ratios (1:25, 1:50, 1:75) and 1:50 seems to perform better.
", "rawMarkdown": "Thanks Chris. The idea was selecting \"squared\" images to better fit in 64x64, so W / H < 1.50 (or H / W if H > W). I tried with different ratios (1:25, 1:50, 1:75) and 1:50 seems to perform better.", "votes": null }, { "id": "600361", "postDate": "08/16/2019 02:55:31", "content": "That's smart. I think selecting quality training images and using data augmentation (random crops etc) improves LB a lot. I also see that you don't use dogs if the resolution of the bounding box is less than 64x64.
", "rawMarkdown": "That's smart. I think selecting quality training images and using data augmentation (random crops etc) improves LB a lot. I also see that you don't use dogs if the resolution of the bounding box is less than 64x64.", "votes": null }, { "id": "600396", "postDate": "08/16/2019 03:58:50", "content": "Yes, you are right. I forgot to mention that.
", "rawMarkdown": "Yes, you are right. I forgot to mention that.", "votes": null }, { "id": "600523", "postDate": "08/16/2019 08:27:50", "content": "Simple, Nice and clean code. Good luck with stage II.
", "rawMarkdown": "Simple, Nice and clean code. Good luck with stage II.", "votes": null }, { "id": "600601", "postDate": "08/16/2019 10:42:26", "content": "@alvaroma Good work! But as pointed out by @bestfitting, kaggle will use another datasets for private LB, so the MiFID of it might be different if the model was only trained to learn from selective squared dog images?
", "rawMarkdown": "alvaroma Good work! But as pointed out by @bestfitting, kaggle will use another datasets for private LB, so the MiFID of it might be different if the model was only trained to learn from selective squared dog images?", "votes": null }, { "id": "600866", "postDate": "08/16/2019 17:05:31", "content": "For those that are interesting, here are samples of Alvaro's dogs. They look great!\n
This one is very interesting. I wasn't able to get good results with my DCGAN implementation. Good to know that such a high result could be achieved with relatively simple architectures and not only BigGANs.
", "rawMarkdown": "This one is very interesting. I wasn't able to get good results with my DCGAN implementation. Good to know that such a high result could be achieved with relatively simple architectures and not only BigGANs.", "votes": null } ], "comments": [ { "id": 600306, "author_name": "cdeotte", "author_url": "", "post_date": "08/16/2019 00:32:39", "content": "Wow, simple and effective. Congratulations. Great job.
\n\n(Do you mean to say, select only with ratio less than 1 : 1.5 or less?)
Thanks Chris. The idea was selecting \"squared\" images to better fit in 64x64, so W / H < 1.50 (or H / W if H > W). I tried with different ratios (1:25, 1:50, 1:75) and 1:50 seems to perform better.
", "votes": null, "replies": [] }, { "id": 600361, "author_name": "cdeotte", "author_url": "", "post_date": "08/16/2019 02:55:31", "content": "That's smart. I think selecting quality training images and using data augmentation (random crops etc) improves LB a lot. I also see that you don't use dogs if the resolution of the bounding box is less than 64x64.
", "votes": null, "replies": [] }, { "id": 600396, "author_name": "alvaroma", "author_url": "", "post_date": "08/16/2019 03:58:50", "content": "Yes, you are right. I forgot to mention that.
", "votes": null, "replies": [] }, { "id": 600601, "author_name": "markpeng", "author_url": "", "post_date": "08/16/2019 10:42:26", "content": "@alvaroma Good work! But as pointed out by @bestfitting, kaggle will use another datasets for private LB, so the MiFID of it might be different if the model was only trained to learn from selective squared dog images?
", "votes": null, "replies": [] } ] }, { "id": 600523, "author_name": "jmourad100", "author_url": "", "post_date": "08/16/2019 08:27:50", "content": "Simple, Nice and clean code. Good luck with stage II.
", "votes": null, "replies": [] }, { "id": 600866, "author_name": "cdeotte", "author_url": "", "post_date": "08/16/2019 17:05:31", "content": "For those that are interesting, here are samples of Alvaro's dogs. They look great!\n
This one is very interesting. I wasn't able to get good results with my DCGAN implementation. Good to know that such a high result could be achieved with relatively simple architectures and not only BigGANs.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "600294": "Hi there!\n\nThe notebook for my solution is available here : [https://www.kaggle.com/alvaroma/dcgan-lb-30-65](https://www.kaggle.com/alvaroma/dcgan-lb-30-65)\n\nKeypoints:\n\n- Pytorch DCGan\n- NZ 64, Batch 32, LRG 0.0004, LRD 0.0008\n- Cropped images, selecting only those with ratio 1:50\n- Transformations: Flip Horizontal, Random Crop\n- Early stop at best FID\n- Select best random images\n\nGood luck for all participants!", "600306": "Wow, simple and effective. Congratulations. Great job.\n\n(Do you mean to say, select only with ratio less than `1 : 1.5` or less?)", "600310": "Thanks Chris. The idea was selecting \"squared\" images to better fit in 64x64, so W / H < 1.50 (or H / W if H > W). I tried with different ratios (1:25, 1:50, 1:75) and 1:50 seems to perform better.", "600361": "That's smart. I think selecting quality training images and using data augmentation (random crops etc) improves LB a lot. I also see that you don't use dogs if the resolution of the bounding box is less than 64x64.", "600396": "Yes, you are right. I forgot to mention that.", "600523": "Simple, Nice and clean code. Good luck with stage II.", "600601": "alvaroma Good work! But as pointed out by @bestfitting, kaggle will use another datasets for private LB, so the MiFID of it might be different if the model was only trained to learn from selective squared dog images?", "600866": "For those that are interesting, here are samples of Alvaro's dogs. They look great!\n", "601386": "This one is very interesting. I wasn't able to get good results with my DCGAN implementation. Good to know that such a high result could be achieved with relatively simple architectures and not only BigGANs." }, "source": "meta" }