search

GeorgeChenZJ / deepcount

Deep Density-aware Count Regressor: a state-of-the-art method for crowd counting
27 stars 8 forks source link

loss divergence #2

Open zhuoxizeng1212 opened 5 years ago

zhuoxizeng1212 commented 5 years ago

Hi thank you for your great work and sharing it with the community. I am trying to re-train your model with a resnet backbone, training from scratch. However, it seems like model would not converge at all.

Have you experienced anything similar to this with VGG backbone?

GeorgeChenZJ commented 4 years ago

Hi. For a network with batch norm layers such as resnet, I believe a larger batch size would help (experience suggests a good start is from 64). Other training parameters matter too.