search

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)
MIT License
3.9k stars 930 forks source link

Why the accuracy of training increases dramatically at 100th epoch in the Figure 3(a) of the paper? #164

Open jeff62802217 opened 5 years ago

jeff62802217 commented 5 years ago

I carried out the experiment again but found the training accuracy converged to 27% and remained unchanged.

g2-bernotas commented 5 years ago

I don't have an answer, but it could be that the NN found a more optimal solution at epoch 100. I have worked a bit with the Mask R-CNN before and I found that not all the time the best solution is found as the training, to my understanding, is a bit random. Or maybe the authors changed a parameter value at that epoch to result in such a dramatical change.

EDIT:

To further support my second opinion (change of parameter value), we can take a look at this paper: http://openaccess.thecvf.com/content_ICCV_2017_workshops/papers/w44/Hara_Learning_Spatio-Temporal_Features_ICCV_2017_paper.pdf in particular, Figure 3 (a). Here we see that value gets better at round numbers such as 100 and 150, indicating that the parameter value changes at certain intervals.