I wonder why the code under ./model/pkgs/clip are slightly different from openai's official repository? For example, this is the architecture of Bottlenet in ./model/pkgs/clip/model.py
So what is the consideration behind these alteration? Why not use the implemented model by simply importing clip package? And are there any references for these changes?
Also I'm confused about the 'pretrained' option. Since the model used is different from openai's implementation, why it is possible for the model to load the state dict value downloaded from openai's official URL? I tried to set it True and found the code worked, but I'm more confused than it didn't.
Thanks for any help.
I wonder why the code under ./model/pkgs/clip are slightly different from openai's official repository? For example, this is the architecture of Bottlenet in ./model/pkgs/clip/model.py
However, this is the architecture of Bottlenet in "https://github.com/openai/CLIP/blob/main/clip/model.py"
So what is the consideration behind these alteration? Why not use the implemented model by simply importing clip package? And are there any references for these changes? Also I'm confused about the 'pretrained' option. Since the model used is different from openai's implementation, why it is possible for the model to load the state dict value downloaded from openai's official URL? I tried to set it True and found the code worked, but I'm more confused than it didn't. Thanks for any help.