Idea from kratzert/finetune_alexnet_with_tensorflow and fellow the main route to finetune Vgg-Face on Morph dataset for age estimation. When finish finetune. Can get MAE 5.5 for age estimation.
Main Finetune Steps:
-
Download the Caffe pretrained model, (if the model is old version, use Caffe util (upgrade_net_proto_binary upgrade_net_proto_text ) to upgrade the model file).
-
Clone the git caffe-tensorflow and use fellow command to generate tensorflow network file and numpy binary dump file. ./convert.py examples/mnist/lenet.prototxt --code-output-path=mynet.py ./convert.py examples/mnist/lenet.prototxt --caffemodel examples/mnist/lenet_iter_10000.caffemodel --data-output-path=mynet.npy
-
Write you own tensorflow finetune file, and load the weights from the numpy dumped file.
-
You should better validate it before finetune you own data too make sure everything is ok