lyatdawn / LightCNN-mxnet

mxnet implementation of LightCNN in the paper "A Light CNN for Deep Face Representation with Noisy Labels" (https://arxiv.org/abs/1511.02683)

Light CNN for Deep Face Recognition

• mxnet implementation of LightCNN in the paper "A Light CNN for Deep Face Representation with Noisy Labels" [Paper]
• Borrowed code and ideas from yxu0611's Tensorflow-implementation-of-LCNN: https://github.com/yxu0611/Tensorflow-implementation-of-LCNN.

Install Required Packages

First ensure that you have installed the following required packages:

• mxnet0.12.0 (instructions). Maybe other version is ok.
• Opencv (instructions). Here is opencv-2.4.9.

See requirements.txt for details.

Datasets

• In the implementation of the LightCNN, MsCelebV1-Faces-Aligned subset as the training data and validating data. Utilize the 10K, 70K download to generate the subset of the MsCelebV1-Faces-Aligned. Then, use the mxnet tools im2rec.py to get the Record file, the image size is 144x144. When training LightCNN, all the training data are converted to gray-scale images and randomly croped to 128x128.

Training a Model

• Run the following script to train and validate the model.

python train.py

You could change some arguments in the train.py, like num_epochs, gpus. If want to get the good results and save training time, there are some points should note:

• For weights initialization, use "xavier" initialization for "weight" of every layer, the "bias" will initializate to constant value "0.1". See common/fit.py initializer for details.
• For learning rate, use "AdamOptimizer" as optimization method, and initial learning rate is 0.00001. So in mxnet, you can set the initial learning rate is 0.0001, lr_step_epochs is '0, 100', in this way, the initial learning rate is 0.00001 in epoch 0.

Extract features of LFW

• Run the following script to extract every image's feature, moreover save the features and image path to MATLAB data.

cd extract_features
python extract_features.py

The script will load the trained LightCNN model to generate fc layer output, w.t. the feature of a image. You could change the arguments in test.sh depend on your machine config. You could download the aligned LFW datasets and lfw_patch_part.txt from here download.

Face Verification

• Utilize MATLAB to run the lfw_eval.m to verificate a pair of images. You must generate the features of the aligned LFW datasets in advance. See lfw_verification-matlab/lfw_eval.m for details.

Downloading trained model

• Pretrained model: download. Training about 50 epoches use 10k people, the loss has converged. But the accuracy on the aligned LFW datasets is not high, about 85%.

• Some improved solutions:

1. Change the optimizer method, like: when loss doesn't decrease, manually decrease it to 10 time smaller.
2. Improved by manaully aligning the images which are mis-algined in LFW datasets.
3. Utilize a new feature of a image, for example use mfm_fc1_output as a image's feature, it will increase the accuracy in LFW.
4. Improved by using metric learning method for similarity caculation, not just cos value.