shicai
shicai
you dont need add a new layer, just change the name and num_output of the last conv layer. of course, you can also remove the last conv layer, then add...
check your data processing steps. the last conv is equivalent to fc layer. there is no difference.
you can try `./build/tools/caffe` first, with this transform params: ``` transform_param { scale: 0.017 mirror: false crop_size: 224 mean_value: [103.94,116.78,123.68] } ```
you can not use single `classify.py` script for all kinds of models, since different models are trained in different ways. for this mobilenet model, you should change your script, and...
for natural RGB images, you can use my default setttings to finetune your model.
1/0.017=58.8, it is used as the approximated std values for image processing.
it has been computed by facebook before. it is [ 58.395, 57.12 , 57.375] for RGB channels. pls see: https://github.com/facebook/fb.resnet.torch/blob/master/datasets/imagenet.lua#L69
certainly yes.
yes, you can train from scratch in this way. i think it makes very little difference.
~100 epochs. more is better.