Results on K49

[janzd]

Hi, I've trained some models on Kuzushiji-49 dataset. I tested some heavier models, DenseNet and Shake-Shake to see if and how much they outperform the shallow ResNet architecture, and I also tested different batch sizes and different augmentation techniques (Mixup, Cutout). The Shake-Shake model is actually probably the same as hysts used to report his results. I was able to reach a balanced accuracy of 98.29 on K49 with Shake-Shake, a large batch size, and cutout. Here is my repo with the results.

[hardmaru]

Thanks for running these experiments! Would be great additions to the table in the repo

On Sun, Feb 17, 2019 at 10:49 PM Jan Zdenek notifications@github.com wrote:

Hi, I've trained some models on Kuzushiji-49 dataset. I tested some heavier models, DenseNet and Shake-Shake to see if and how much they outperform the shallow ResNet architecture, and I also tested different batch sizes and different augmentation techniques (Mixup, Cutout). The Shake-Shake model is actually probably the same as hysts used to report his results. I was able to reach a balanced accuracy of 98.29 on K49 with Shake-Shake, a large batch size, and cutout. Here is my repo https://github.com/kurapan/pytorch_image_classification with the results.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rois-codh/kmnist/issues/8, or mute the thread https://github.com/notifications/unsubscribe-auth/AGBoHriTTRgHfzcHoFvQJzm4bBZo-Luyks5vOV37gaJpZM4a_kWi .

[janzd]

Should I send a PR with an updated table?

[mxbi]

Thanks a lot, and sorry about the delay! I went ahead and added your best result using both Shake-Shake and Densenet to the table in the repo.