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

Function reorg_gradient

dlib/cuda/cpu_dlib.cpp:2436–2476  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

2434 }
2435
2436 void reorg_gradient(
2437 bool add_to,
2438 tensor& grad,
2439 const int row_stride,
2440 const int col_stride,
2441 const tensor& gradient_input
2442 )
2443 {
2444 DLIB_CASSERT(!is_same_object(grad, gradient_input), "Grad and gradient_input must be distinct objects.");
2445 DLIB_CASSERT(grad.nr() % row_stride == 0, "The number of rows in grad must be divisible by row_stride.");
2446 DLIB_CASSERT(grad.nc() % col_stride == 0, "The number of columns in grad must be divisible by col_stride.");
2447 DLIB_CASSERT(grad.num_samples() == gradient_input.num_samples(), "The number of samples in grad and gradient_input must match.");
2448 DLIB_CASSERT(grad.k() == gradient_input.k() / row_stride / col_stride, "The number of channels in grad must be gradient_input.k() divided by row_stride and col_stride.");
2449 DLIB_CASSERT(grad.nr() == gradient_input.nr() * row_stride, "The number of rows in grad must be gradient_input.nr() multiplied by row_stride.");
2450 DLIB_CASSERT(grad.nc() == gradient_input.nc() * col_stride, "The number of columns in grad must be gradient_input.nc() multiplied by col_stride.");
2451
2452 const float* gi = gradient_input.host();
2453 float* g = grad.host();
2454
2455 parallel_for(0, gradient_input.num_samples(), [&](long n)
2456 {
2457 for (long k = 0; k < gradient_input.k(); ++k)
2458 {
2459 for (long r = 0; r < gradient_input.nr(); ++r)
2460 {
2461 for (long c = 0; c < gradient_input.nc(); ++c)
2462 {
2463 const auto in_idx = tensor_index(gradient_input, n, k, r, c);
2464 const auto out_idx = tensor_index(grad,
2465 n,
2466 k % grad.k(),
2467 r * row_stride + (k / grad.k()) / col_stride,
2468 c * col_stride + (k / grad.k()) % col_stride);
2469
2470 if (add_to) g[out_idx] += gi[in_idx];
2471 else g[out_idx] = gi[in_idx];
2472 }
2473 }
2474 }
2475 });
2476 }
2477
2478 // ------------------------------------------------------------------------------------
2479

Callers 2

test_reorgFunction · 0.50
backwardMethod · 0.50

Calls 8

parallel_forFunction · 0.85
tensor_indexFunction · 0.85
is_same_objectFunction · 0.50
nrMethod · 0.45
ncMethod · 0.45
num_samplesMethod · 0.45
kMethod · 0.45
hostMethod · 0.45

Tested by 1

test_reorgFunction · 0.40