google-research / nested-transformer Goto Github PK

Could you please let me know the hyperparameters you used to train Swin Transformer on CIFAR100?

I noticed some possible discrepancies of the architecture parameters here vs those in table A1 of the paper

For ImageNet models, is it correct that:

1. The table should say h=[3,3,4]?
2. The order of the scale_hidden_dims in the table is inverted. That is, hierarchies 1, 2 and 3 should say [4d, 4h] × 2, 1 [2d, 2h] × 2, 4 [d, h] × k, 16?

Just suggesting you try a similar architecture with sequences rather than images.

I would like to know the training details. Thanks a lot if you can help me.

I am training on a medium-scale dataset that consists of 100,000 images. The learning rate and weight decay as the same as your config but still not working. Any opinion?

Regards,
Khawar Islam

could you share your patch size and window size when you run swin on CIFAR

Hello, thanks for sharing your interesting work.

I was trying to reproduce the NesT-T ImageNet result in this link using TPUs.

Here are my TPU-v3 8 cores result (link) by using exactly the same hyperparameters in imagenet_nest_tiny.py

As you can see, it takes 63 hours for training while your result takes 21 hours.
How can I reduce training hours such as your result?
If this difference came from the data loading time, could you tell me the types of data storage that you used?
Right now, I'm using the google cloud storage bucket for data storage.

Furthermore, I can see the accuracy difference around 0.5% (81.0 v.s. 81.5).
Could you explain this difference?

Hi
I'm interested to work with the Nest model, however I'm facing difficulty with the implementation of GradCAT. Could you please share the implementation?