there is some wrong with the train-net you provided

[chengyao0929]

Hi，I used MobileNetSSD_train.prototxt you provided to train the model,but when i used MobileNetSSD_deploy.prototxt to test the model,there is some wrong with MobileNetSSD_deploy.prototxt, "Incompatible error "，so i fixed the wrong and test the model，but could not find any detections，so can you tell me how to train the model？the MobileNetSSD_train.prototxt and MobileNetSSD_deploy.prototxt you provided is right？ thank you very much！

[chuanqi305]

Because the batch_norm, scale layer can be merged to conv layer at test stage, so I merged it from MobileNetSSD_train.caffemodel to MobileNetSSD_deploy.caffemodel. I just uploaded the merge_bn.py, you can generate your own caffemodel by this script. If you want to train other dataset than PASCAL VOC, you should change the conv_xx_mbox_conf layer output num like this: 84 --> (class count) * 4 126-->(class count) * 6 And you should change the layer name if your training is start with MobileNetSSD_train.caffemodel, or with no change of layer name but converge slowly if start with MobileNet_nofc.caffemodel. After several(about 100K) training iterator, your loss will be 0.6 ~ 1.5, check your data set if this is not reached.

[kaishijeng]

Does you have demo.py for testing MobileNetSSD_deploy.prototxt and your trained model? Any idea about fps of test speed and how it compares yolov2?

Thanks,

[chuanqi305]

The caffe implement of MobileNet is slow because there is no optimization for depth-wise convolution. I have tested it on my in-house Android CNN framework, it's much faster than yolov2, and also faster than tiny-yolo.

[Numeria-Jun]

Hello,I have also have problems with chenyao0929 which send error message "couldn't find any detections", and I follow your answer to use merge.py creating MobileNetSSD_deploy.caffemodel,but it still doesn't work. the layer output has also been revise . the MobileNetSSD_train.prototxt and MobileNetSSD_deploy.prototxt you provided is right？thank you.