search

AojunZhou / Incremental-Network-Quantization

Caffe Implementation for Incremental network quantization
Other
191 stars 74 forks source link

The weight is too large. #18

Open lamperouge11 opened 6 years ago

lamperouge11 commented 6 years ago

The weight is 2,4,8,16,32. is it normal? As result ,When retrain the model , the loss can't converged. How to make the weight within -1,1 ?Thanks.@Zhouaojun

AojunZhou commented 6 years ago

@lamperouge11 for ImageNet, most convNet models (ResNet, GoogleNet, AlexNet, VGG) weight range are within -1,1.