Caffe implementation of SSD detection on MobileNetv2, converted from tensorflow.
Tensorflow and Caffe version SSD is properly installed on your computer.
- Firstly you should download the original model from tensorflow.
- Use gen_model.py to generate the train.prototxt and deploy.prototxt (or use the default prototxt).
python gen_model.py -s deploy -c 91 >deploy.prototxt
- Use dump_tensorflow_weights.py to dump the weights of conv layer and batchnorm layer.
- Use load_caffe_weights.py to load the dumped weights to deploy.caffemodel.
- Use the code in src to accelerate your training if you have a cudnn7, or add "engine: CAFFE" to your depthwise convolution layer to solve the memory issue.
- The original tensorflow model is trained on MSCOCO dataset, maybe you need deploy.caffemodel for VOC dataset, use coco2voc.py to get deploy_voc.caffemodel.
- Generate the trainval_lmdb and test_lmdb from your dataset.
- Write a labelmap.prototxt
- Use gen_model.py to generate some prototxt files, replace the "CLASS_NUM" with class number of your own dataset.
python gen_model.py -s train -c CLASS_NUM >train.prototxt
python gen_model.py -s test -c CLASS_NUM >test.prototxt
python gen_model.py -s deploy -c CLASS_NUM >deploy.prototxt
- Copy coco/solver_train.prototxt and coco/train.sh to your project and start training.
There are some differences between caffe and tensorflow implementation:
- The padding method 'SAME' in tensorflow sometimes use the [0, 0, 1, 1] paddings, means that top=0, left=0, bottom=1, right=1 padding. In caffe, there is no parameters can be used to do that kind of padding.
- MobileNet on Tensorflow use ReLU6 layer y = min(max(x, 0), 6), but caffe has no ReLU6 layer. Replace ReLU6 with ReLU cause a bit accuracy drop in ssd-mobilenetv2, but very large drop in ssdlite-mobilenetv2. There is a ReLU6 layer implementation in my fork of ssd.