shicai
shicai
models are trained with my customized caffe, and convert to standard caffe. accuracy results are tested under standard caffe, please make sure that you use the original images from ImageNet...
It's so weird. there should be no problem with the pretrained weights. but there might be some problems with different image processing steps or different implementations. when it is converted...
``` layer { name: "data" type: "Input" top: "data" input_param { shape: { dim: 1 dim: 3 dim: 224 dim: 224 } } } ``` this should be work with...
there is no big difference between deploy prototxt and train prototxt. you'd better check your settings again.
if training from sctrach, please follow the original papers. if finetuning, it depends, you can try to use lr=0.01 or 0.001 to start your training.
ReLU is used instead of ReLU6. It is ok to use ReLU to train the model.
fixed, i changed it to utf8 without bom.
I have not met this error before, maybe there is something wrong with your caffe or cudnn? you may try https://github.com/yonghenglh6/DepthwiseConvolution for speedup.
yes, i did use data shuffling for each epoch. btw, is this the result from the model you trained from scratch?
due to poor implementations of caffe group convolutions.