leo-zhou / tripletloss

tripletloss in caffe

TripletLoss Example

This aims to reproduce the loss function used in Google's FaceNet paper.

This is just an example which shows how to use the method to train a model.

Online sample selection

I provide an online triplet sample selection usage. Just a runnable strategy. YOU MAY CHANGE IT to fit your own strategy.

Setup

Rebuild your caffe directory and makesure your python could find the added layers.

Go to your caffe root path:

cp Makefile.configexample Makefile.config

Open Makefile.config uncomment the line :

WITH_PYTHON_LAYER := 1

Then return to caffe root create build directory:

mkdir build
cd build
cmake ..
make all & make pycaffe

Cluster Sample

I trained a 256 dim output model, the results for clustering is followed:

Usage

(You should review the code first~)

a.Pre-training your model with softmax loss. Initialize the parameters of the CNN(layer's lr_mult,decay_mult) which you find in your caffe folder. Don't forget the 'base_lr' in solver.prototxt.

b.Uncomment the lines of 'softmax' in train.prototxt and comment 'tripletloss'.

c.Change the configs in ./tripletloss/config.py, Makesure your image path is exists, (my path is exampled)

python train.py

d.Uncomment the lines of 'tripletloss' in train.prototxt and comment 'softmax'.

e.Initialize the parameters mentioned above.

python train.py

I provide a pretrained example model training from a data set of 997 indentities. You could change the top fc9 layer's name and finetune this model. But the best way is to make the model to fit your own dataset smoothly. Re-training the model on your own data set.

My approach is like the Baidu's paper. (also similiar with the vgg_face's method)

Firstly, pretraining the model with softmax, Here means to fix your featrue output to satisfy the classification model. Then do a kind of metric learning. Using triplet method to train the linear transform layer fc9_1, making the feature's affine projection fit the expected Euclidean distance.

notation: maybe your need a really well cropped face dataset to do this.

If the proposal is helpful for you, please star it. thanks~