chuanqi305

Still can't download on Jan.5

@AustinVan I used show_tensor.py to analyze the tensorflow implementation, and replicated the implementation in caffe. In my opinion, the depthwise conv in prediction layers refers to layer19.

1、batch数调一下试试，有时较大的batch会更有效 2、学习速率一开始不能太大，甚至固定基础网络的权重只训练最后一层 3、建议将voc的person类的权重提取出来作为你的基础网络，这样一开始的loss就不会很大，在这个基础上finetune会好一些

@fraukej @mychina75 According to the [paper](https://arxiv.org/abs/1801.04381), MobileNetv2 is better than v1 in both FPS and mAP. But it's very difficult to train and finetune in my test. Maybe some details...

@itemhsu Make sure your model is [ssd_mobilenetv2_coco](http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz)

@madhavajay v1 is [here](https://github.com/chuanqi305/MobileNet-SSD)

@Mrils The test.prototxt is ready and the bug of demo_caffe_voc.py has been fixed, thank you for your attention.

@wtiandong Thank you very much, the code has been changed.

@CarryJzzZ this error is caused by depthwise convolution. Please use my version of depthwise convolution https://github.com/chuanqi305/MobileNetv2-SSDLite/tree/master/src

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...