About float and half

[idata]

Now I am running a self caffe model and it's training in float point.When I run it on NCS with C++,the outputs are different with the results outputted on windows through VS.

Following is part of the output of NCS:Category 482 is: 10544.0Category 483 is: -8336.0Category 484 is: -2166.0Category 485 is: 1444.0Category 486 is: 2420.0Category 487 is: 552.0Category 488 is: 3742.0Category 489 is: -1145.0Category 490 is: -4592.0Category 491 is: -3924.0Category 492 is: -14144.0Category 493 is: 8608.0Category 494 is: -2808.0Category 495 is: -264.8Category 496 is: -2390.0Category 497 is: 3598.0Category 498 is: 5816.0Category 499 is: 5216.0Category 500 is: -4548.0Category 501 is: 12752.0Category 502 is: -4324.0Category 503 is: 7444.0Category 504 is: -2660.0Category 505 is: 5800.0Category 506 is: 4172.0Category 507 is: 8680.0Category 508 is: -3318.0Category 509 is: 951.5Category 510 is: 630.0Category 511 is: 4516.0Index of top result is: 113Probability of top result is: 15728.0

And following is the output of VS on windows:

0.693973 0.9956 -1.04383 0.302624 0.336853 1.54326 -2.62327 -2.1755 0.34646 1.33913 -1.11766 -1.1199 -1.95261 -0.837082 0.36523 0.724509 -0.865793 -1.12518 -0.146516 1.52155 -1.79436 -0.313601 -0.725422 1.05461 0.515218 2.54319 -0.38223 1.94542 -0.186538 -0.71533 0.163427 0.284998 -3.08611 1.00865 1.15777 -0.325071 -1.65323 1.51967 1.84158 -0.526452 1.01265 -0.175593 1.33656 0.0283866 0.338035 -0.675896 2.05167 -0.452004 -1.3773 -1.69321 0.575961 -1.79048 -0.436242 1.40409 0.586798 1.83048 1.57769 -1.21422 -0.43123 1.20651 -1.47478 -0.0310141 0.98217 0.674636 0.120956 0.506346 1.81047 -1.96575 0.612426 0.586477 -1.95771 0.995132 0.98117 1.24573 0.27711 -2.59334 -0.41862 0.182753 1.31437 -0.539988 1.07199 -0.17873 0.376313 1.5717 0.210307 0.596403 -1.29091 0.41806 -1.36455 0.232549 -1.37666 2.16447 1.79488 -1.01879 0.814555 1.38125 2.97567 0.461581 0.200207 -0.472004 -1.99846 -0.0923696 -0.888577 2.52875 -0.535223 0.509798 0.197736 0.102396 -0.549576 0.350901 2.04382 -0.897249 1.7332

The VS outputs is right,however the NCS's output is wrong. Is this the problem of training in float point??