DenseNet , ResNet , WRN models

[shuaibaslam2019]

Hi,

Thank you so much for giving us new data augmentation technique. I am bit confused, when I explore your models folder.
My main question is, are you training your ImageNet models with Pre-trained weights? It seems, you don't use pre-trained but you are building your own models and training without using "imagenet" weights.. Please, correct me If I am wrong here..

Secondly, the context of the first question is in the second question about baseline accuracies. Are you getting the baseline accuracies without "Imagenet" weights?

Any guideline to use FMix in keras?

Thank you!

[ethanwharris]

Hi,

Thanks for the issue :)

1. You are correct that we don't use ImageNet weights, all of the models we used were trained from scratch (random initialisation)
2. The baseline results were also trained from scratch, but used standard augmentations found in datasets/datasets.py
3. We don't currently have a guide for FMix on keras but will take a look at adding one

Hope that helps!

[shuaibaslam2019]

Thank you man, it helped a lot. I was in depression while figuring out my questions!!!

Thank you again🙂

[shuaibaslam2019]

Hi @ethanwharris ,

Could you please tell me what input image sizes you used for fashion-mnist? I can't find in your paper. Default is 28 by 28 as you know but what size you kept in your experiments? It would be a great help.. thank you!

Regards,
Aslam

[shuaibaslam2019]

Hi @ethanwharris ,

If I am not wrong, you are using 28 fashion-mnist images size in your experiments.

[ethanwharris]

Hi @shuaibaslam2019, sorry for the delayed reply. Yes, we use 28 x 28 for Fashion-MNIST. It's worth mentioning that I think the Fashion-MNIST version we used is older than what you get with torchvision. You can find a Dataset that will load the older version here: https://github.com/ecs-vlc/FMix/blob/master/datasets/fashion.py

[innat]

@shuaibaslam2019 I've played with fmix, cutmix, mixup with keras sequence generator, notebook here, you may find it as a guide starter in keras.

[Lupin1998]

Hi, we have reimplemented FMix in OpenMixup for image classification benchmarks on various datasets (ImageNet, CIFAR-10/100, Tiny-ImageNet, etc.). Various backbone architectures (ResNet, ResNeXt, Wide-ResNet, ViT, Swin Transformer, etc.) are supported.