search

PRIS-CV / Fine-Grained-or-Not

Code release for Your “Flamingo” is My “Bird”: Fine-Grained, or Not (CVPR 2021 Oral)
MIT License
58 stars 9 forks source link

Selection of some parameters and loss #7

Closed hit-dyl closed 2 years ago

hit-dyl commented 2 years ago

hi，this works very well, I have two questions

  1. Why Hiden_Number is 600？
  2. Why choose NLL loss，do not use CE loss directly？
dongliangchang commented 2 years ago
  1. To make the three different tasks have the same number of features, we chose a number that is closer to 512 and divisible by 3.
  2. In torch, NLL(log(softmax(x))) ==CE(x)
hit-dyl commented 2 years ago

Got it, thank you