The "alpha" and "beta" in the paper are the opposite of the "alpha" and "beta" in the code of Tip-Adapter

[euminds]

In Code,
"alpha_list = [i * (6.0 - 1.0) / 20 + 1 for i in range(20)] " "beta_list = [i * (7 - 0.1) / 200 + 0.1 for i in range(200)]"
In paper,

[euminds]

The optimal hyper-parameters of Tip-Adapter, the results I achieved are Residual Ratio1.0, Sharpness Ratio 5.0, acc 62.02%. Also what is the optimal hyperparameter for Tip-Adapter-F with acc 65.51% (16-shot)? My current result is 65.45%.

[ZrrSkywalker]

@euminds Thanks for pointing out.
We have fix this and release a new code base in a repo.

Concerning 65.45% for Tip-Adapter, the released code would achieve 65.51% on my original device, but has variance on others. Thus, it's common to get +-0.1% accuracy jittor.

[euminds]

@ZrrSkywalker
Thank you for your amazing paper,

I am trying to evaluate CLIP with RN50x16 on ImageNet,
output = model.encode_image(test_image)
but get error:
File "", line 1, in <cell line: 1>
output = model.encode_image(test_image)
File "/home/user/anaconda3/envs/yolov5_4/lib/python3.8/site-packages/clip/model.py", line 337, in encode_image
return self.visual(image.type(self.dtype))
File "/home/user/anaconda3/envs/yolov5_4/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/anaconda3/envs/yolov5_4/lib/python3.8/site-packages/clip/model.py", line 148, in forward
x = self.attnpool(x)
File "/home/user/anaconda3/envs/yolov5_4/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/anaconda3/envs/yolov5_4/lib/python3.8/site-packages/clip/model.py", line 69, in forward
x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC
RuntimeError: The size of tensor a (50) must match the size of tensor b (145) at non-singleton dimension 0

Thanks