search

wenjiaXu / APN-ZSL

This is a PyTorch implementation of the paper "Attribute Prototype Network for Zero-Shot Learning".
67 stars 8 forks source link

Question about paper and code. #19

Open fire717 opened 1 year ago

fire717 commented 1 year ago

Hi, I am now working on ZSL and interested in your nice work, after reading the code, I have some questions need your help:

  1. In your paper, after the image encoder( means resnet101 backbone in code), here are ProtoMod and BaseMod, but in code ,there isn't any mod, but just two simple branches(conv-pool-mm-softmax), is this the Mod?
  2. In code the two branches seems is the same, while output branch is (conv-pool-mm-softmax), the other extract['layer4'] branch is also same structure. And in loss computing, they also do same thing.
  3. By the question of 2, I delete branch ['layer4'] related loss computing, got acc 68.7... (while the origin code I ran is 68.6 which is the same as your readme on AwA2)

Looking forward to your reply, thx!