# --------------------------------------------------------
# TRIPLET LOSS
# Copyright (c) 2015 Pinguo Tech.
# Written by David Lu
# --------------------------------------------------------
"""The data layer used during training a VGG_FACE network by triplet loss.
The layer combines the input image into triplet.Priority select the semi-hard samples
"""
import caffe
import numpy as np
from numpy import *
import yaml
from multiprocessing import Process, Queue
from caffe._caffe import RawBlobVec
from sklearn import preprocessing
import math
import config
class TripletSelectLayer(caffe.Layer):
def setup(self, bottom, top):
"""Setup the TripletSelectLayer."""
self.triplet = config.BATCH_SIZE/3
top[0].reshape(self.triplet,shape(bottom[0].data)[1])
top[1].reshape(self.triplet,shape(bottom[0].data)[1])
top[2].reshape(self.triplet,shape(bottom[0].data)[1])
def forward(self, bottom, top):
"""Get blobs and copy them into this layer's top blob vector."""
top_archor = []
top_positive = []
top_negative = []
labels = []
self.tripletlist = []
self.no_residual_list=[]
aps = {}
ans = {}
archor_feature = bottom[0].data[0]
for i in range(self.triplet):
positive_feature = bottom[0].data[i+self.triplet]
a_p = archor_feature - positive_feature
ap = np.dot(a_p,a_p)
aps[i+self.triplet] = ap
aps = sorted(aps.items(), key = lambda d: d[1], reverse = True)
for i in range(self.triplet):
negative_feature = bottom[0].data[i+self.triplet*2]
a_n = archor_feature - negative_feature
an = np.dot(a_n,a_n)
ans[i+self.triplet*2] = an
ans = sorted(ans.items(), key = lambda d: d[1], reverse = True)
for i in range(self.triplet):
top_archor.append(bottom[0].data[i])
top_positive.append(bottom[0].data[aps[i][0]])
top_negative.append(bottom[0].data[ans[i][0]])
if aps[i][1] >= ans[i][1]:
self.no_residual_list.append(i)
self.tripletlist.append([i,aps[i][0],ans[i][0]])
top[0].data[...] = np.array(top_archor).astype(float32)
top[1].data[...] = np.array(top_positive).astype(float32)
top[2].data[...] = np.array(top_negative).astype(float32)
def backward(self, top, propagate_down, bottom):
for i in range(len(self.tripletlist)):
if not i in self.no_residual_list:
bottom[0].diff[self.tripletlist[i][0]] += top[0].diff[i]
bottom[0].diff[self.tripletlist[i][1]] += top[1].diff[i]
bottom[0].diff[self.tripletlist[i][2]] += top[2].diff[i]
else:
bottom[0].diff[self.tripletlist[i][0]] += np.zeros(shape(top[0].diff[i]))
bottom[0].diff[self.tripletlist[i][1]] += np.zeros(shape(top[1].diff[i]))
bottom[0].diff[self.tripletlist[i][2]] += np.zeros(shape(top[2].diff[i]))
#print 'backward-no_re:',bottom[0].diff[0][0]
#print 'tripletlist:',self.no_residual_list
def reshape(self, bottom, top):
"""Reshaping happens during the call to forward."""
pass
