MobilenetV2_tf_to_caffe

convert tensorflow mobilenet-v2 checkpoint to caffemodel

Requirements:

1. pycaffe

2. tensorflow >= 1.5

Clone https://github.com/tensorflow/models, and put 'models' folder in this repository

To convert tf mobilenet_v2 to caffe model:

Prepare checkpoints of mobilenet_v2:

sh download.sh mobilenet_v2_1.0_224

See: https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet

Restore tensorflow graph from checkpoint and revise protxt (for mobilenet_v2_1.0_224) :

python generate_prototxt.py --factor 1.0 --image_scale 224

revised prototxt will be generated in prototxt_mobilenet_v2/ directory.

Convert variables to caffe parameters:

python converter_v2.py --factor 1.0 --image_scale 224

caffemodel will be generated in caffemodel_fromckpt/ directory.

Note:

1. Fix padding problem by using pad=2 and slicing layers.

2. Set epsilon to 1e-3 in batch_norm layer which is equal to tensorflow.

3. ReLU6 not solved, but result is OK.