MCPcopy Create free account
hub / github.com/davisking/dlib / rms_normalize

Function rms_normalize

dlib/cuda/cpu_dlib.cpp:1454–1515  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

1452// -----------------------------------------------------------------------------------
1453
1454 void rms_normalize(
1455 const double eps,
1456 resizable_tensor& dest,
1457 resizable_tensor& scale,
1458 const tensor& src,
1459 const tensor& gamma
1460 )
1461 {
1462 DLIB_CASSERT(
1463 gamma.k() == src.k() &&
1464 gamma.nr() == 1 &&
1465 gamma.nc() == 1 &&
1466 eps > 0,
1467 "\nsrc.k(): " << src.k() <<
1468 "\ngamma.k(): " << gamma.k() <<
1469 "\ngamma.nr(): " << gamma.nr() <<
1470 "\ngamma.nc(): " << gamma.nc() <<
1471 "\neps: " << eps
1472 );
1473
1474 const long ns = src.num_samples();
1475 const long ks = src.k();
1476 const long num = src.nr() * src.nc();
1477
1478 dest.copy_size(src);
1479 scale.set_size(ns);
1480
1481 // Compute RMS values
1482 scale = 0;
1483 const float* p_src = src.host();
1484 float* p_scale = scale.host();
1485 for (long n = 0; n < ns; ++n)
1486 {
1487 for (long k = 0; k < ks; ++k)
1488 {
1489 for (long i = 0; i < num; ++i)
1490 {
1491 p_scale[n] += (*p_src) * (*p_src);
1492 ++p_src;
1493 }
1494 }
1495 p_scale[n] = 1.0f / std::sqrt(p_scale[n] / (ks * num) + static_cast<float>(eps));
1496 }
1497 scale.host();
1498
1499 // Apply RMS normalization
1500 p_src = src.host();
1501 float* p_dest = dest.host();
1502 const float* p_gamma = gamma.host();
1503 for (long n = 0; n < ns; ++n)
1504 {
1505 for (long k = 0; k < ks; ++k)
1506 {
1507 for (long i = 0; i < num; ++i)
1508 {
1509 *p_dest = (*p_src) * p_scale[n] * p_gamma[k];
1510 ++p_src;
1511 ++p_dest;

Callers 2

test_rms_normalizeFunction · 0.50
forwardMethod · 0.50

Calls 8

copy_sizeMethod · 0.80
sqrtFunction · 0.50
kMethod · 0.45
nrMethod · 0.45
ncMethod · 0.45
num_samplesMethod · 0.45
set_sizeMethod · 0.45
hostMethod · 0.45

Tested by 1

test_rms_normalizeFunction · 0.40