The dimensions do not broadcast about vit-adapter HOT 3 CLOSED

Hi, great work! Trying to test it out. Maybe a bug:

Settings config: mask2former_beit_adapter_base_512_40k_cocostuff10k_ss.py checkpoint: mask2former_beit_adapter_base_512_40k_cocostuff10k.pth.tar

Error In this line in beit.py:

            attn = attn + relative_position_bias.unsqueeze(0)

The dimensions do not broadcast. If one takes an input image 1x3x128x128, then the dimensions are:

# attn.shape
# ([1, 12, 65, 65])

# relative_position_bias.unsqueeze(0).shape
# ([1, 12, 1025, 1025])

Yes, BEiT has a limitation on image resolution. When it is set to img_size=512, the input image must be 1x3x512x512.
It does not support a dynamic resolution (due to its implementation of relative position encoding), one possible way is to pad a 128x128 image to 512x512. Or if all images are 128x128, set img_size=128.

from vit-adapter.

Thank you for your fast reply. I fixed it in the code like this (rewriting in TF2). Works like charm:

attn = attn + tf.expand_dims(relative_position_bias[:,:N,:N], 0)

from vit-adapter.

Can you explain your solution in more detail please @rmihaylov . I'm using pytorch.
Thanks

from vit-adapter.