VGG + CUB params

[Ahmed Taha]

Thanks for sharing your code. Can you please share or at least confirm the params used to train VGG on CUB.
According to the readme example

• base-lr 0.01 [Not mentioned in the paper]
• batch 32 [Not mentioned in the paper]
• epoch = 105 [Not mentioned in the paper, deduced from utils_args]
• attdrop 3 4 53 [slight mentioned in the paper]
• threshold 0.80 [clearly mentioned in the paper]
• keep_prob 0.25 [clearly mentioned in the paper]

are these the right params?
I noticed you use different params in some of the closed issues
#2
epoch = 200
base-lr 0.001

[Junsuk Choe]

Hi ahmdtaha,

I used these parameters:

• base-lr: 0.01
• batch: 128
• attdrop: 3 4 53
• threshold: 0.80
• keep_prob: 0.25

See also #2.

[Ahmed Taha]

Thanks for your reply and best regards with your future plans
Batch_size is different from utils_args, this is a minor change.
Can you confirm the number of epochs as well?

[Junsuk Choe]

Thank you!

You don't need modify the number of epochs.

In utils_args.py, I set the appropriate the number of epochs for CUB.

if args.cub:
    args.laststride = 1
    args.stepscale = 5.0
    args.epoch = 105

So you will actually run the 525 epochs. Note that 200 epochs also work well.

[Won Bae]

Can you also share the params you used for resnet50 on CUB as well as Imagenet?

[Junsuk Choe]

@won-bae

For ResNet, I use threshold 0.90 and keep_prob 0.25. Batch size is 128 for both datasets.

[Ahmed Taha]

For ResNet50-SE what is the --attdrop? i.e., where are adl components inserted?
For Cam_vgg, attdrop is [3,4,53]. what is the list for resnet50-se

[Won Bae]

@ahmdtaha It's 31 41 5 for resent50-se