我已经训练了一个收集神经网络，现在我正在分析验证集上的结果。
我正在使用VGG-16启发模型来生成图像嵌入（类似于分类，但不是在最后一层上应用softmax，而是使用tf.nn.l2_normalize并输出其结果）。
这是我评估验证集上的结果的方法：
for q in range(total_batch):
s = q * batch_size
e = (q+1) *batch_size
input1,input2, input3 = training.next_batch(s,e)
m1 = sess.run([model1], feed_dict={x_anchor:input1})
当我将 2张或更多图片提供给网络时，我会得到以下结果（这看似正确）：
model 1[[ 0.54123181 0.80775619 0.67851502 -0.9016602 -0.32920051 -0.39620045
-0.57724983 0.39768949 -0.71639627 -0.92292905 -0.85529834 -0.80494177
-0.85227519 0.79539144 0.70412946 -0.84128755 0.24074456 -0.63105983
-0.47843483 0.51363963 -0.93949044 -0.64124674 0.64220959 -0.85175914
0.12223328 -0.90112907 -0.24197945 -0.59273964 -0.66464484 0.53093082
0.99998033 -0.44616687 -0.76501042 0.86774755 -0.5939005 -0.73198503
-0.86760223 -0.69168311 0.50087672 -0.84233558 0.27748042 0.72440165
-0.10699634 -0.69008458 0.59424591 0.28699273 0.9249208 0.39881933
-0.59371895 0.43215483 -0.55006212 -0.43135682 0.36412039 0.48564553
-0.83413011 -0.72999161 0.45997152 0.32711393 0.25188333 -0.25339019
-0.79056996 0.83432364 -0.90330607 -0.5077076 0.44556049 0.44817379
-0.99124968 -0.92694545 0.9209342 0.47476801 0.9576726 0.63066739
-0.90548855 -0.45749596 -0.6577583 -0.50995409 -0.23899333 -0.402354
0.69469041 0.88706255 -0.70948708 -0.86379802 -0.70613575 -0.99586815
0.18220347 -0.41845426 -0.56594819 -0.19674864 -0.3409965 0.99766964
0.76333064 -0.44313303 0.99999446 0.43300939 0.81194556 -0.41180304
0.52517533 0.40369266 -0.51925427 -0.39346993 0.80244076 -0.8112722
0.70244497 -0.74119097 0.9619118 0.68992859 0.46409786 0.87701035
-0.74074823 -0.23732661 -0.77201623 0.27954161 0.68868124 -0.99404013
-0.42262578 0.83058596 -0.70084429 -0.634561 0.95630938 0.34737089
0.67471945 -0.5309912 -0.99926859 -0.65541434 0.26393309 0.39613721
-0.43749622 -0.78592688]
[ 0.84087348 -0.58951676 0.7345866 -0.43244514 -0.94426006 -0.91816401
-0.81656754 0.91751999 -0.69769353 -0.38497004 -0.51813591 -0.59335381
-0.52309358 0.60609615 0.71007174 -0.54058796 0.97058851 0.77573431
-0.87812304 0.858006 -0.34257504 -0.76733476 0.76652908 -0.52393353
0.99250138 -0.43355095 -0.9702813 -0.80539411 -0.74715942 0.84741515
0.0062686 -0.89494973 -0.64401799 0.49700516 -0.80453843 -0.68132073
-0.4972589 -0.72220105 0.86551863 -0.53895342 -0.96073133 0.68937814
-0.99425936 -0.72372872 0.80428344 0.95793277 0.3801598 0.91702944
-0.80467254 0.90179944 -0.83512384 -0.90218145 0.93135184 0.87415588
-0.55156773 -0.68345618 0.88793361 -0.94498491 0.96775752 -0.96736413
0.61237168 0.5512749 0.42899671 0.86152947 -0.89525187 0.89394641
-0.13200052 -0.37519619 0.38971815 0.88011098 0.28785968 0.77605313
-0.42437062 -0.88921171 -0.75322908 -0.86020166 0.97102123 -0.91548419
0.71930879 0.46164933 -0.70471841 -0.5038383 -0.70807648 -0.09081167
0.98326087 -0.90823787 -0.82444084 -0.98045391 -0.94006455 0.06823064
0.64600801 -0.89645582 -0.00331163 -0.90138936 -0.58373314 -0.91127288
0.85099399 0.91489464 -0.85461974 -0.91933751 0.59673184 0.58466858
0.71173817 -0.67129421 -0.27336016 0.72387743 0.88578391 0.48047155
-0.67178267 0.97142994 -0.63560289 -0.96013361 0.72506416 -0.10901433
-0.9063043 0.55689049 -0.71331424 -0.77287281 0.29235667 0.93772775
0.7380743 0.84737724 -0.03824199 -0.75526953 -0.96454096 0.91819137
0.89922023 -0.61831951]]
但是当我将批量大小设置为1张图像时，嵌入会发生变化：
model 1[[ 1. -0.99999994 1. -1. -1. -1. -1.
1. -1. -1. -0.99999994 1. 1. -1.
0.99999994 1. 1. 1. -1. 1.
0.99999994 1. 1. 1. -0.99999994 -1. -1.
1.00000012 -1. 1. -1. -1. -1.
0.99999994 -1. -1. 1. -1. 1. -1.
-0.99999994 -0.99999994 1. -0.99999994 1. 1. -1.
1. -1. 0.99999994 1. -1. 0.99999994
1. 1. 0.99999994 1. -1. -1. -1.
1. -1. 1. 1. -1. 1. 1.
-1. 1. -1. -0.99999994 1. 1. -1.
-1. -1. 1. -1. 1. -0.99999994
-0.99999994 -0.99999994 -1. 1. 1. -1. 1.
-1. -0.99999994 -1. -0.99999994 -1. -0.99999994
-1. -1. -1. 1. 1. 1.
-0.99999994 -1. 1. 0.99999994 0.99999994 -1.
0.99999994 1. -1. 1. 0.99999994 1.
-0.99999994 -1. 1. -0.99999994 1.00000012 -1. -1.
-1. 1. 1. 1. 0.99999994 -1. -1.
1. 0.99999994 -1.00000012]]
--------------------------------------
为什么会这样？
稍后编辑：
当我批量改变图像数量时会发生这种情况：对于2,3和10图像没有任何变化。
模型1 - 2批图像
[[ 0.2743271 0.91178268 0.5872615 -0.98093092 -0.02889113 -0.03134303
-0.04661543 0.45045683 -0.72907752 -0.99990839 -0.98905557 -0.9919185
-0.91167277 0.30989948 0.04958323 -0.21191983 0.04454518 -0.44197813
-0.99413317 0.03212294 -0.1868417 -0.94774926 0.55885661 -0.9531188
0.29837319 -0.91574109 -0.02016532 -0.87102389 -0.74571055 0.03648382
0.88016605 -0.0329459 -0.82466573 0.49874297 -0.91681188 -0.75305611
-0.9985494 -0.71474195 0.0363951 -0.87884218 0.27167588 0.28420082
-0.22817914 -0.04758676 0.84340817 0.99999797 0.13699238 0.03059645
-0.90965801 0.41122833 -0.40550923 -0.0283178 0.02512571 0.3185125
-0.9999544 -0.89122361 0.03199656 0.02893824 0.4112269 -0.38217896
-0.85480034 0.97177809 -0.6597178 -0.04604623 0.03689161 0.78157973
-0.91014808 -0.90671444 0.93194991 0.60108137 0.90480918 0.6574111
-0.18326542 -0.02942787 -0.04679498 -0.04116671 -0.05752649 -0.02889678
0.04707722 0.14425533 -0.93019849 -0.9645502 -0.66847634 -0.06463299
0.01547062 -0.49338332 -0.41203031 -0.33844921 -0.35335046 0.7529974
0.39456338 -0.52115846 0.08408639 0.04551683 0.22299762 -0.02990589
0.96985286 0.44175237 -0.9540993 -0.47797546 0.92866713 -0.86449015
0.72105265 -0.26022291 0.56368029 0.91930032 0.02831106 0.14027278
-0.74462092 -0.01777472 -0.25965354 0.08585165 0.41533408 -0.73962849
-0.41705173 0.96994138 -0.04474245 -0.37005508 0.5702228 0.01939926
0.91649407 -0.08130386 -0.82676262 -0.6780622 0.02080963 0.02526239
-0.21303178 -0.85304713]
[ 0.96163648 -0.41067314 0.8093974 -0.19435681 -0.99958253 -0.99950862
-0.99891287 0.89279819 -0.68443102 0.01353077 -0.1475431 0.12687652
0.41091695 -0.95076936 0.99877 0.97728705 0.99900734 0.89702582
0.10816276 0.99948382 0.98239011 0.31901628 0.82926422 0.30259636
-0.9544493 -0.40176904 -0.99979657 0.49124065 -0.66626996 0.99933416
-0.47466597 -0.99945712 -0.56562042 0.86675 -0.39931941 -0.6579563
0.05384279 -0.69938827 0.99933743 -0.47711253 -0.96238875 -0.95876473
0.97361922 -0.99886709 0.53727323 -0.00204876 -0.99057204 0.99953175
-0.41535798 0.91153234 0.91409093 -0.99959892 0.99968427 0.94791865
0.00955276 0.45356417 0.99948794 -0.99958116 -0.91153306 -0.92408824
0.51895714 -0.23589683 0.75151342 0.99893922 -0.99931931 0.62380534
0.4142831 -0.42174494 0.36258689 -0.79918784 -0.42581734 0.75353211
0.98306352 -0.99956691 -0.99890453 -0.99915218 0.99834388 -0.99958235
0.99889123 -0.98954052 -0.36705688 -0.26389948 -0.74373341 0.99790913
0.99988031 -0.86981201 0.91117018 -0.94098461 -0.93549109 -0.65802348
-0.91886866 -0.85345995 -0.99645847 -0.99896359 -0.974819 -0.99955267
-0.2436911 0.89713699 0.29949048 -0.87837315 -0.37091413 0.50264978
0.69288033 0.96554852 -0.82599306 0.39355665 0.9995991 -0.9901129
0.6674875 0.99984199 0.96570182 -0.99630791 -0.90966898 0.67301542
-0.9088828 0.24333875 -0.99899858 -0.92900985 -0.82149017 0.99981183
0.40004814 0.99668926 0.56255096 -0.73500454 -0.99978352 0.99968088
0.97704524 -0.52183366]]
--------------------------------------
模型1 - 3批图像
[[ 0.0844328 -0.77254468 0.02411784 -0.00955349 -0.94160694 -0.94066721
-0.88625979 0.1149781 -0.02488928 -0.0051736 -0.01160398 -0.02201086
-0.03521669 0.06318482 0.90330708 0.22742757 0.88465506 0.69219536
-0.09456988 0.94272465 0.78211629 -0.54738432 0.03007991 -0.02421509
0.76738203 -0.00945781 -0.94134593 -0.07806421 -0.02830387 0.93667924
-0.77421671 -0.9394843 -0.01889145 0.02541017 -0.12410384 -0.02550597
-0.01388813 -0.05055251 0.93260431 -0.01248641 -0.81171089 0.32357386
-0.58824182 -0.89974731 0.12902646 0.7322554 -0.73108768 0.94067121
-0.11807644 0.07763585 -0.55410898 -0.94702822 0.91589779 0.060298
-0.01905362 -0.09003559 0.93983525 -0.01288702 0.73827815 -0.13281474
0.7777003 0.04422731 0.87029737 0.92503053 -0.90938175 0.11651902
0.59449238 -0.21909271 0.00718945 0.65837044 -0.48126209 0.0279669
0.58864224 -0.94854271 -0.90761161 -0.92175764 0.15994184 -0.94678712
0.90896451 -0.48565978 -0.02508452 -0.0084288 -0.02446353 0.92655313
0.94136518 -0.12520707 -0.5488351 -0.13468495 -0.10819445 -0.28640082
0.04190208 -0.07546353 -0.88280475 -0.90112931 -0.04243152 -0.94591689
0.09673197 0.11114503 -0.66887909 -0.11806588 0.01580797 0.66838497
0.0267245 -0.30286619 -0.9696154 0.03818087 0.94815379 -0.49568826
-0.06035382 0.89603657 -0.20980616 -0.14883731 0.36203983 0.26767451
-0.0712135 0.01647491 -0.91310835 -0.07197624 -0.11620539 0.94761276
0.11829913 0.0154587 0.25997716 -0.02925223 -0.88819623 0.94844997
0.89705831 -0.01577406]
[ 0.06992867 -0.31909466 0.02536829 -0.01694431 -0.33658543 -0.3355417
-0.43915471 0.08860384 -0.02602066 -0.00991058 -0.01751714 -0.0475461
-0.17316175 0.9174453 0.42861098 -0.88639027 0.43813875 0.29763162
-0.08426359 0.33345035 -0.4970766 -0.83511752 0.02875736 -0.21122056
0.60217178 -0.01281438 -0.31258523 -0.07853186 -0.03347619 0.34996024
0.09653269 -0.34235266 -0.02185564 0.03394587 -0.11823756 -0.02901799
-0.02814217 -0.06023972 0.35179645 -0.01470418 -0.44020525 0.92780703
-0.60654044 -0.43617141 0.12288523 0.61880648 0.57744491 0.33927616
-0.11079701 0.05827573 -0.8260603 -0.32088026 0.32848966 0.04612691
-0.04382927 -0.13040958 0.34031361 -0.00834278 0.61152399 -0.09501366
0.32608342 0.21010639 0.21165863 0.28613895 -0.31460357 0.09254558
-0.34285307 -0.94639134 0.00946684 0.71495456 0.84588212 0.02944838
-0.62024397 -0.31660575 -0.41654193 -0.38240901 0.0965206 -0.32049972
0.412752 0.77851194 -0.03856483 -0.01161524 -0.02505984 -0.11366697
0.30903593 -0.09858466 -0.82802993 -0.09703995 -0.07281385 0.05652994
0.11177502 -0.0600725 0.05701264 -0.33766046 -0.00946549 -0.32439664
0.09326869 0.08744934 -0.73073256 -0.09487063 0.03088553 0.32274199
0.02751617 -0.90462279 -0.17905805 0.05551227 0.31582361 0.7906518
-0.11476897 0.36962783 -0.91509163 -0.08848038 0.92485541 -0.07423788
-0.05653551 0.0251725 -0.40672073 -0.05931272 0.09049221 0.29042748
0.1437183 0.01071607 -0.03970633 -0.03280446 -0.3763141 0.30833092
0.41030872 -0.01896038]
[-0.99397242 0.54895651 0.9993872 -0.99981076 -0.00930807 -0.05056711
0.14726442 -0.98940867 -0.9993515 -0.99993747 -0.99977916 -0.99862647
0.9842636 -0.39281276 0.01813768 0.40322316 -0.15943591 0.65748084
0.99194562 0.00900941 0.37577793 0.05430508 0.99913377 0.97713834
-0.22025895 -0.9998731 0.12711462 0.99385047 -0.99903864 0.01264404
-0.62551564 -0.01279757 -0.99958265 0.99910063 0.98519951 -0.99925339
-0.99950749 -0.996903 -0.0805502 -0.99981397 0.38385507 -0.18567207
0.5348646 -0.01446672 -0.9839977 -0.28439516 0.36341196 0.00541824
0.98680389 -0.99527711 0.10289539 -0.01316902 -0.23070674 -0.99711412
-0.99885732 0.98736358 0.02993833 0.9998821 0.28457668 0.9865762
-0.53744942 -0.97667766 0.44472823 -0.24988614 0.27212018 -0.9888674
-0.72734493 -0.23736458 0.99992937 0.23534702 -0.22993524 0.99917501
0.51845711 -0.00524166 -0.05228702 -0.06423581 -0.98239625 -0.02956322
0.0584744 0.39755991 -0.99894112 -0.99989706 -0.99938661 -0.3585794
-0.13538241 0.98722047 0.11465792 0.98612529 0.99145955 -0.95644069
0.99284977 0.99533743 0.46626735 0.27194008 0.99905449 -0.00283846
-0.9909308 -0.98994917 -0.13649452 0.98846346 0.99939793 -0.67014861
0.99926412 0.29988277 -0.16668546 0.99772769 0.03549499 0.3593924
0.99155712 -0.24595471 0.34436712 0.98489523 -0.11648855 0.96064514
0.99585754 0.9995473 -0.0284851 0.99564117 -0.98909426 -0.13297313
-0.98252243 -0.99982309 0.96479797 -0.99903357 0.26361924 0.07331257
0.16411291 -0.99969572]]
模型1 - 10批图像
[[ 0.04818952 -0.05918095 0.01802165 ..., 0.05679128 0.11442375
-0.01418886]
[ 0.05519219 -0.1301478 0.01969874 ..., 0.07767208 0.18997246
-0.01425523]
[ 0.08626937 -0.25595629 0.02016677 ..., 0.63242364 0.43534788
-0.01190863]
...,
[-0.14989793 0.49290144 0.18958355 ..., 0.01846622 -0.25266901
-0.22219713]
[ 0.09586276 -0.28965887 0.02260818 ..., 0.69641638 0.4886305
-0.01328262]
[-0.01896885 0.5706526 0.03627656 ..., 0.0253126 -0.36303344
-0.05107375]]
model1 [0]批量处理10张图片：
model 1[ 5.86070597e-01 1.95135430e-01 2.46683449e-01 -1.15100272e-01
-8.81946564e-01 -8.91003788e-01 -9.79653239e-01 4.03030038e-01
-3.37121964e-01 -8.15378204e-02 -1.51473030e-01 -6.80846691e-01
-9.61061358e-01 9.99767303e-01 9.17174935e-01 -9.18523490e-01
4.61293280e-01 -2.48836592e-01 -6.79437101e-01 8.90483081e-01
-9.21952069e-01 -9.33306396e-01 3.34582895e-01 -9.64680552e-01
2.17209637e-01 -9.24504027e-02 -8.05327177e-01 -9.81654227e-01
-5.13453305e-01 8.98176014e-01 4.65111285e-01 -8.81239653e-01
-2.21239150e-01 1.82802275e-01 -9.89428520e-01 -3.78754020e-01
-1.94648057e-01 -9.60758328e-01 9.99629676e-01 -1.34037957e-01
5.10413982e-02 9.51475441e-01 -1.16250336e-01 -9.02551413e-01
9.75457489e-01 3.81219059e-01 6.40182137e-01 8.98657262e-01
-9.81229663e-01 3.48009765e-01 -8.25251520e-01 -8.78372252e-01
7.62746096e-01 3.04450452e-01 -3.40841383e-01 -9.91277814e-01
8.71233225e-01 -5.17561100e-04 4.97255594e-01 -1.69827744e-01
-1.96575001e-01 9.88305330e-01 -3.83750439e-01 -1.71618521e-01
1.40758827e-01 7.19896436e-01 -3.79732788e-01 -5.24435103e-01
7.77314454e-02 8.57556820e-01 5.83529353e-01 3.34895849e-01
-8.37150455e-01 -8.86918783e-01 -8.40985894e-01 -8.68736804e-01
-1.96579075e-03 -8.52129221e-01 8.46933126e-01 6.71221852e-01
-6.03814125e-01 -1.12122051e-01 -2.26175249e-01 -5.59274018e-01
5.39124966e-01 -3.51777464e-01 -8.13132882e-01 -1.37575284e-01
-2.99482316e-01 7.41403461e-01 8.68456364e-01 -4.78453755e-01
3.74983132e-01 1.13467500e-01 1.33464202e-01 -8.81694794e-01
7.77389646e-01 3.82896155e-01 -9.17372942e-01 -3.26944888e-01
3.77652109e-01 -1.51620105e-01 2.89691836e-01 -9.90572333e-01
3.18472356e-01 8.29179347e-01 8.44251931e-01 7.09410369e-01
-9.93203878e-01 -4.48296890e-02 -9.96632695e-01 8.17295071e-03
9.04701591e-01 -7.95541167e-01 -4.05060053e-01 1.68647662e-01
-9.02341783e-01 -8.14558685e-01 9.71956789e-01 9.01448905e-01
9.96791184e-01 -5.61384344e-03 -7.10371912e-01 -4.87548530e-01
4.36763577e-02 8.43983173e-01 -1.29987687e-01 -1.70727238e-01]
后来编辑：这是我的火车评估功能。唯一改变的是batch_size：
def train(x_anchor, x_positive, x_negative, idx_model):
global weights, bias, batch_size, keep_rate, suma, nr
sess = tf.Session()
saver = tf.train.import_meta_graph('/home/bogdan/triplet/model_85_la_suta/my-model_'+str(idx_model)+'.meta')
saver.restore(sess,tf.train.latest_checkpoint('./triplet/model_85_la_suta/'))
graph = tf.get_default_graph()
with tf.device('/gpu:1'):
with tf.variable_scope("siamese") as scope:
model1 = siamese_convnet(x_anchor, graph)
scope.reuse_variables()
model2 = siamese_convnet(x_positive, graph)
scope.reuse_variables()
model3 = siamese_convnet(x_negative, graph)
eps = 1e-10
d_pos = tf.sqrt(tf.reduce_sum(tf.square(model1 - model2), 1) + eps)
d_neg = tf.sqrt(tf.reduce_sum(tf.square(model1 - model3), 1) + eps)
training = lfw_generated_test.inputData()
training.shuffle_epoca();
nr_training_examples = training.get_nr_training()
print ("nr_training_examples "+str(nr_training_examples))
total_batch = int(nr_training_examples/batch_size)
print("tb "+str(total_batch))
avg_acc_test = 0;
for q in range(total_batch):
s = q * batch_size
e = (q+1) *batch_size
input1,input2, input3 = training.next_batch(s,e)
distance1, distance2, m1, m2, m3 = sess.run([d_pos, d_neg, model1, model2, model3], feed_dict={x_anchor:input1, x_positive:input2, x_negative:input3})
'''print("input 1"+str(input1))
print("--------------------------------------")'''
'''print("dist1 = "+str(distance1))
print("--------------------------------------")
print("dist2 = "+str(distance2))
print("--------------------------------------")
print("--------------------------------------")'''
''''print(np.shape(distance1))
print(np.shape(distance2)) '''
print("model 1"+str(m1[0]))
print("--------------------------------------")
''''print("--------------------------------------")
print("model 2"+str(m2))
print("--------------------------------------")
print("--------------------------------------")
print("model 3"+str(m3))
print("--------------------------------------")
print("--------------------------------------")'''
''''print(m2)
print(m3)'''
''''print(np.shape(model1))
print(np.shape(model2))
print(np.shape(model3))'''
#print(str(q)+" ----------------------------------------------------------------")
test_acc = compute_accuracy(distance1, distance2)
avg_acc_test +=test_acc*100
print('Accuract TEST set %0.2f' % (avg_acc_test/(total_batch)))
batch_size = 2
#70x70 images
x_anchor = tf.placeholder('float', [None, 4900])
x_positive = tf.placeholder('float', [None, 4900])
x_negative = tf.placeholder('float', [None, 4900])
labels = tf.placeholder(tf.float32, [None, 1]) #0 sau 1 (impostor sau genuine)
train(x_anchor, x_positive, x_negative, 99);