search

yu-changqian / TorchSeg

Fast, modular reference implementation and easy training of Semantic Segmentation algorithms in PyTorch.
MIT License
1.41k stars 253 forks source link

The number of paramters #30

Open Jason93K opened 5 years ago

Jason93K commented 5 years ago

First of all, thank you for your work. However, I tried to get the total number of parameters of the BiSeNet(Xception39 based) with 19 classes(Cityscapes) by using pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad), the result is 1.54 M. It is not the same as 5.8 M in paper. How did you measure the number of parameters?

BebDong commented 4 years ago

@Jason93K @ycszen They seem to mistake the model size on disk as the number of model parameters.