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,...

@he-chen Thank you very much, I have merged your code to my project.

@yja1 Use my version of depthwise convolution https://github.com/chuanqi305/MobileNetv2-SSDLite/tree/master/src

@HAN-Seul I‘m working on it and trying some of this: 1. Replace the ReLU layers by ReLU6 2. Use the same padding as tensorflow 3. Change the post processing layers....

@HAN-Seul Now all the weights checked, but maybe we need to finetune the model to recover the precision because the changes of ReLU6 and padding.

No, I can't do that, because: 1. Tensorflow MobileNet-SSD use ReLU6 as activation function, but there is no ReLU6 in caffe. You can replace it with ReLU and the MobileNet...

Maybe you were using CPU to train? On my machine(GTX1080 * 2) the training speed is about 1000 iterations per hour. Make sure you have compiled ssd with cudnn support.

The depth-wise convolution is implemented by 'group' parameter, that's the bottleneck. I have implemented it myself, and the performance is better than tinyyolo. I will open the source later.

@vsooda thanks for your solution of coredump，I converted the whole dataset to avoid this. The final loss of coco is about 4.0 - 5.0, I fixed the base MobileNet weights...

Yes，just like you said, I trained it on coco and finetuned it on VOC0712. I fixed the base MobileNet weights for the first 10000 iterations. Not so much tricky I...