self.ln_pre or self.ln_post?
https://github.com/sdc17/CrossGET/blob/dc3297534ecc2cbd7c0d3806df3e6426dc5c1a53/CLIP/clip/model.py#L296 https://github.com/sdc17/CrossGET/blob/dc3297534ecc2cbd7c0d3806df3e6426dc5c1a53/CLIP/clip/model.py#L297 https://github.com/sdc17/CrossGET/blob/dc3297534ecc2cbd7c0d3806df3e6426dc5c1a53/CLIP/clip/model.py#L298
Hello, why is self.ln_pre used here instead of self.ln_post? In def encode_text, self.ln_final is used, which represents the output layer_norm.
Hi,
This should be self.In_post like what we did for other models. It seems that the difference between self.ln_pre and self.ln_post was not so big, so it didn't result in an issue for CLIP model. Thanks.