mirror/darknet
1
0
Fork 0
mirror of https://github.com/pjreddie/darknet.git synced 2023-08-10 21:13:14 +03:00
Code Issues Projects Releases Wiki Activity

Fixed Depth-wise (grouped) convolution for CPU and GPU (CUDNN=0/1)

Browse Source
This commit is contained in:
AlexeyAB
2019-06-15 02:32:57 +03:00
parent a1abd07e23
commit f0582446f2
5 changed files with 46 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/convolutional_kernels.cu
View File

@ -578,7 +578,8 @@ void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
state.workspace); // output
}
gemm_ongpu(0, 0, m, n, k, 1., a, k, b, n, 1., c + i*m*n, n);
//gemm_ongpu(0, 0, m, n, k, 1., a, k, b, n, 1., c + i*m*n, n);
gemm_ongpu(0, 0, m, n, k, 1, a, k, b, n, 1, c, n);
}
}
@ -817,7 +818,8 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
l.stride, l.stride, // stride (h, w)
l.dilation, l.dilation, // dilation (h, w)
state.workspace); // output
gemm_ongpu(0, 1, m, n, k, 1, a + i*m*k, k, b, k, 1, c, n);
//gemm_ongpu(0, 1, m, n, k, 1, a + i*m*k, k, b, k, 1, c, n);
gemm_ongpu(0, 1, m, n, k, 1, a, k, b, k, 1, c, n);
if (state.delta) {
if (l.binary || l.xnor) swap_binary(&l);
@ -825,7 +827,9 @@ void backward_convolutional_layer_gpu(convolutional_layer l, network_state state
float * b = l.delta_gpu + (i*l.groups + j)*m*k;
float * c = state.workspace;
gemm_ongpu(1, 0, n, k, m, 1, a, n, b + i*k*m, k, 0, c, k);
//gemm_ongpu(1, 0, n, k, m, 1, a, n, b + i*k*m, k, 0, c, k);
gemm_ongpu(1, 0, n, k, m, 1, a, n, b, k, 0, c, k);
float *delta = state.delta + (i*l.groups + j)*l.c / l.groups*l.h*l.w;

34
src/convolutional_layer.c
View File

@ -547,7 +547,16 @@ convolutional_layer make_convolutional_layer(int batch, int steps, int h, int w,
if (l.xnor && l.use_bin_output) fprintf(stderr, "convXB");
else if (l.xnor) fprintf(stderr, "convX ");
else fprintf(stderr, "conv ");
fprintf(stderr, "%5d %2d x%2d /%2d(%d)%4d x%4d x%4d -> %4d x%4d x%4d %5.3f BF\n", n, size, size, stride, dilation, w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
if(groups > 1) fprintf(stderr, "%5d/%3d ", n, groups);
else fprintf(stderr, "%5d ", n);
if(dilation > 1) fprintf(stderr, "%2d x%2d/%2d(%1d)", size, size, stride, dilation);
else fprintf(stderr, "%2d x%2d/%2d ", size, size, stride);
fprintf(stderr, "%4d x%4d x%4d -> %4d x%4d x%4d %5.3f BF\n", w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
//fprintf(stderr, "%5d/%2d %2d x%2d /%2d(%d)%4d x%4d x%4d -> %4d x%4d x%4d %5.3f BF\n", n, groups, size, size, stride, dilation, w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
return l;
}
@ -897,9 +906,10 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
static int u = 0;
u++;
for(i = 0; i < l.batch; ++i){
for (j = 0; j < l.groups; ++j) {
for(i = 0; i < l.batch; ++i)
{
for (j = 0; j < l.groups; ++j)
{
float *a = l.weights +j*l.nweights / l.groups;
float *b = state.workspace;
float *c = l.output +(i*l.groups + j)*n*m;
@ -1025,7 +1035,11 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
}
else {
//printf(" l.index = %d - FP32 \n", l.index);
float *im = state.input + (i*l.groups + j)*l.c / l.groups*l.h*l.w;
float *im = state.input + (i*l.groups + j)*(l.c / l.groups)*l.h*l.w;
if (l.size == 1) {
b = im;
}
else {
//im2col_cpu(im, l.c / l.groups, l.h, l.w, l.size, l.stride, l.pad, b);
im2col_cpu_ext(im, // input
@ -1037,11 +1051,13 @@ void forward_convolutional_layer(convolutional_layer l, network_state state)
l.dilation, l.dilation, // dilation (h, w)
b); // output
}
gemm(0, 0, m, n, k, 1, a, k, b, n, 1, c, n);
// bit-count to float
}
c += n*m;
state.input += l.c*l.h*l.w;
//c += n*m;
//state.input += l.c*l.h*l.w;
}
}
@ -1079,7 +1095,7 @@ void backward_convolutional_layer(convolutional_layer l, network_state state)
float *b = state.workspace;
float *c = l.weight_updates + j*l.nweights / l.groups;
float *im = state.input + (i*l.groups + j)*l.c / l.groups*l.h*l.w;
float *im = state.input + (i*l.groups + j)* (l.c / l.groups)*l.h*l.w;
//im2col_cpu(im, l.c / l.groups, l.h, l.w, l.size, l.stride, l.pad, b);
im2col_cpu_ext(
@ -1112,7 +1128,7 @@ void backward_convolutional_layer(convolutional_layer l, network_state state)
l.pad, l.pad, // padding (h, w)
l.stride, l.stride, // stride (h, w)
l.dilation, l.dilation, // dilation (h, w)
state.delta + (i*l.groups + j)*l.c / l.groups*l.h*l.w); // output (delta)
state.delta + (i*l.groups + j)* (l.c / l.groups)*l.h*l.w); // output (delta)
}
}
}

2
src/maxpool_layer.c
View File

@ -78,7 +78,7 @@ maxpool_layer make_maxpool_layer(int batch, int h, int w, int c, int size, int s
#endif // GPU
l.bflops = (l.size*l.size*l.c * l.out_h*l.out_w) / 1000000000.;
fprintf(stderr, "max %d x %d / %d %4d x%4d x%4d -> %4d x%4d x%4d %5.3f BF\n", size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
fprintf(stderr, "max %d x %d/%2d %4d x%4d x%4d -> %4d x%4d x%4d %5.3f BF\n", size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c, l.bflops);
return l;
}

2
src/parser.c
View File

@ -834,7 +834,7 @@ network parse_network_cfg_custom(char *filename, int batch, int time_steps)
n = n->next;
int count = 0;
free_section(s);
fprintf(stderr, "layer filters size input output\n");
fprintf(stderr, " layer filters size/strd(dil) input output\n");
while(n){
params.index = count;
fprintf(stderr, "%4d ", count);