myTools
I keep this respository for some useful tools.
Description
caffe_ccl
- coupled-clusters loss layer source code implemented by nicklhy
- check out paper for detail about coupled-clusters loss
caffe_data_augment
- data augmentation layer source code
- color_cast [30]
- aspect_ratio [1.4]
- rotation [10]
- scale_jittering [1.5]
caffe_lbn
- batch normalization source code
- check out paper for detail about batch normalization
caffe_normalization
- L1 and L2 normalization layer source code
- mainly use in front of the triplet loss layer
caffe_siamese
- siamese training strategy
- siamese data layer and contrasitive loss layer
genList
- generate proper training list for caffe image_data_layer
- for triplet loss based training
- one closest negative sample (hard negative) and one farthest positive sample to anchor
- check out paper for details about triplet loss
- for lifted-structured based training
- n closest hard negative samples and (n-1) farthest positive samples
- bath size is 2n (include anchor)
- check out paper for details about lifted-structured feature embedding
circular_train.sh
- circularly train caffe model shell example
- mainly for training the triplet loss based network
- include hard negative and positive samples mining
convert_lmdb_to_numpy.py
- convert the format of image feature extracted from caffemodel
- convert to .npy file for further processing