CompCars模型测试

resize image to 256*256

for file in data/CompCars/test/*.jpg; do convert $file -resize 256x256\! $file; done

生成lmdb文件和均值文件
最好用gedit打开编译，否则可能出项file not found,command not found之类的错误
./create_imagenet.sh
./make_imagenet_mean.sh

性能测试命令
./build/tools/caffe.bin test -model=examples/CompCars/train_val_googlenet.prototxt
-weights=examples/CompCars/googlenet_finetune_web_car_iter_10000.caffemodel -gpu=0

单张图片测试命令
相对路径
./build/examples/cpp_classification/classification.bin examples/CompCars/deploy_cars.prototxt examples/CompCars/googlenet_finetune_web_car_iter_10000.caffemodel data/CompCars/imagenet_mean_comp.binaryproto data/CompCars/synset_words.txt examples/images/4_509.jpg
绝对路径
/home/xxx/caffe-master/build/examples/cpp_classification/classification.bin /home/xxx/caffe-master/examples/CompCars/deploy_cars.prototxt /home/spgz/caffe-master/examples/CompCars/kitlien/googlenet_finetune_web_car_kit1222_iter_10000.caffemodel /home/xxx/caffe-master/data/CompCars/imagenet_mean_comp.binaryproto /home/xxx/caffe-master/data/CompCars/synset_words.txt /home/xxx/caffe-master/examples/images/Mazda.jpg

1.deploy文件需要根据train_val_googlenet.prototxt修改
2.imagenet_mean_comp.binaryproto data使用的是imagenet自带的
3.synset_words.txt为映射文件

微调网络命令
./build/tools/caffe train -solver examples/CompCars/solver_googlenet.prototxt -weights examples/CompCars/kitlien/google
net_finetune_web_car_iter_10000.caffemodel -gpu 0

/home/xxx/caffe-master/build/tools/caffe train -solver /home/xxx/caffe-master/examples/CompCars/solver_googlenet.prototxt -weights /home/xxx/caffe-master/examples/CompCars/kitlien/googlenet_finetune_web_car_iter_10000.caffemodel -gpu 0

错误原因

1.图像未归一化到256*256，会在make_imagenet_mean时出现如下错误
F0105 10:48:02.025462 4031 compute_image_mean.cpp:76] Check failed: size_in_datum == data_size (217800 vs. 187500) Incorrect data field size 217800