search

Event-AHU / OpenPAR

[OpenPAR] An open-source framework for Pedestrian Attribute Recognition, based on PyTorch
MIT License
79 stars 11 forks source link

about the visual_encoder #4

Closed Dinosaurcubs closed 10 months ago

Dinosaurcubs commented 10 months ago

Ncie work, congratulations! I got a question recently, it seems you removed the last ln_post layer before the projection layer in the visual_encoder， What is the reason for doing this?

1125178969 commented 10 months ago

We will feed the visual features of CLIP to the multimodal encoder for feature fusion, so we think that the LN here is not necessary, but experiments show that with or without this LN has no effect on the results.

Dinosaurcubs commented 10 months ago

got it, thanks!