search

fucusy / Cross-View-Gait-Based-Human-Identification-with-Deep-CNNs

Code for 2016 TPAMI(IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE) A Comprehensive Study on Cross-View Gait Based Human Identification with Deep CNNs
MIT License
22 stars 13 forks source link

two more issues #2

Closed ifangcheng closed 7 years ago

ifangcheng commented 7 years ago

many thx for the previous reply！

two more issues：

  1. have you tested the model on the CASIA_B dataset？ can you achieve the 90% around accuracy？

  2. have you observed the final performance difference between using nn.LogSoftMax and nn.SoftMax （since you used nn.LogSoftMax here）

fucusy commented 7 years ago
  1. I have not tried, I guess you can achieve around 90%.
  2. If you use nn.ClassNLLCriterion as loss function you need to add nn.LogSoftMax as your last layer. you can see more from nn.ClassNLLCriterion document here., if you use nn.Softmax, I do not find suitable Criterion class as loss function, you may write the Criterion youself. Besides, combining Log and SoftMax is for speed purpose.