loss/accuracy different between bvlc/caffe with NVIDIA/caffe

NVIDIA / caffe

Caffe: a fast open framework for deep learning.

http://caffe.berkeleyvision.org/

Other

672 stars 263 forks source link

loss/accuracy different between bvlc/caffe with NVIDIA/caffe #532

Closed zlheos closed 5 years ago

zlheos commented 6 years ago

Hello, I have some questions when using NVIDIA/caffe, and hope you can help me. finetune model： with bvlc/caffe , the accuracy is always 0.97 with NVIDIA/caffe , the accuracy drop from 0.98 to 0.58 and then restore to 0.7

prototxt and log is below

I don't know why train_prototxt.txt run_bvlc.log run_nvidia.log

@drnikolaev

rotorliu commented 6 years ago

I have the same question. Same model prototxt The loss of NVIDIA Caffe is more than the loss of BLVC Caffe when training. Why?

drnikolaev commented 5 years ago

@zlheos @rotorliu Please use NVCaffe-specific implementation of BatchNorm layer.