------------Before train-------------------- A. replace specified file noted in replace/readme.txt B. compile caffe: make clean make all make pycaffe C. add python interface because there are .py writed by ourselves. If you have any questions, please move to:https://blog.csdn.net/zllljf/article/details/81670143 D. do datasets there is a datasets sample in ~/SaCNN-master/ShanghaiTech/Part_B -----------Train----------------------------- A. cd ~/SaCNN-master sudo sh train_sacnn.sh B. the average_loss is equal to your train datasets, more explanation can be seen in Caffe website ---------------Retrain---------------------- A. after train, there are log.txt in ~/result and .caffemodel in ~/result_c You can draw loss-iter curve by using tools provided by Caffe which in ~/result B. copy the .caffemodel to ~/model/mine C. cd ~caffe ./build/tools/caffe train --solver ~/solverretrain.prototxt --gpu 0 --weights /~/model/mine/~.caffemodel -----------------Test---------------------- A. After retrain, there is a .caffemodel in /result B. Copy .caffemodel to ~/model/mine/ C. run python test.py