From a392bbd0c957a00e3782c96e7ced84a29ff9dd88 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Tue, 15 Mar 2016 05:33:02 +0000
Subject: [PATCH] Play along w/ alphago
---
src/convolutional_kernels.cu | 61 +++++++++++++++++++++++++-----
1 files changed, 51 insertions(+), 10 deletions(-)
diff --git a/src/convolutional_kernels.cu b/src/convolutional_kernels.cu
index 4fdc1a1..85b92df 100644
--- a/src/convolutional_kernels.cu
+++ b/src/convolutional_kernels.cu
@@ -115,17 +115,57 @@
}
}
+__global__ void dot_kernel(float *output, float scale, int batch, int n, int size, float *delta)
+{
+ int index = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
+ int f1 = index / n;
+ int f2 = index % n;
+ if (f2 <= f1) return;
+
+ float sum = 0;
+ float norm1 = 0;
+ float norm2 = 0;
+ int b, i;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < size; ++i){
+ int i1 = b * size * n + f1 * size + i;
+ int i2 = b * size * n + f2 * size + i;
+ sum += output[i1] * output[i2];
+ norm1 += output[i1] * output[i1];
+ norm2 += output[i2] * output[i2];
+ }
+ }
+ norm1 = sqrt(norm1);
+ norm2 = sqrt(norm2);
+ float norm = norm1 * norm2;
+ sum = sum / norm;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < size; ++i){
+ int i1 = b * size * n + f1 * size + i;
+ int i2 = b * size * n + f2 * size + i;
+ delta[i1] += - scale * sum * output[i2] / norm;
+ delta[i2] += - scale * sum * output[i1] / norm;
+ }
+ }
+}
+
+void dot_error_gpu(layer l)
+{
+ dot_kernel<<<cuda_gridsize(l.n*l.n), BLOCK>>>(l.output_gpu, l.dot, l.batch, l.n, l.out_w * l.out_h, l.delta_gpu);
+ check_error(cudaPeekAtLastError());
+}
+
void backward_bias_gpu(float *bias_updates, float *delta, int batch, int n, int size)
{
backward_bias_kernel<<<n, BLOCK>>>(bias_updates, delta, batch, n, size);
check_error(cudaPeekAtLastError());
}
-void swap_binary(convolutional_layer l)
+void swap_binary(convolutional_layer *l)
{
- float *swap = l.filters_gpu;
- l.filters_gpu = l.binary_filters_gpu;
- l.binary_filters_gpu = swap;
+ float *swap = l->filters_gpu;
+ l->filters_gpu = l->binary_filters_gpu;
+ l->binary_filters_gpu = swap;
}
void forward_convolutional_layer_gpu(convolutional_layer l, network_state state)
@@ -139,7 +179,7 @@
fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
if(l.binary){
binarize_filters_gpu(l.filters_gpu, l.n, l.c*l.size*l.size, l.binary_filters_gpu);
- swap_binary(l);
+ swap_binary(&l);
}
for(i = 0; i < l.batch; ++i){
@@ -150,8 +190,8 @@
gemm_ongpu(0,0,m,n,k,1.,a,k,b,n,1.,c+i*m*n,n);
}
- if(l.batch_normalize){
- if(state.train){
+ if (l.batch_normalize) {
+ if (state.train) {
fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.mean_gpu);
fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.variance_gpu);
@@ -172,7 +212,8 @@
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.n, n);
activate_array_ongpu(l.output_gpu, m*n*l.batch, l.activation);
- if(l.binary) swap_binary(l);
+ if(l.dot > 0) dot_error_gpu(l);
+ if(l.binary) swap_binary(&l);
}
void backward_convolutional_layer_gpu(convolutional_layer l, network_state state)
@@ -206,7 +247,7 @@
gemm_ongpu(0,1,m,n,k,1,a + i*m*k,k,b,k,1,c,n);
if(state.delta){
- if(l.binary) swap_binary(l);
+ if(l.binary) swap_binary(&l);
float * a = l.filters_gpu;
float * b = l.delta_gpu;
float * c = l.col_image_gpu;
@@ -214,7 +255,7 @@
gemm_ongpu(1,0,n,k,m,1,a,n,b + i*k*m,k,0,c,k);
col2im_ongpu(l.col_image_gpu, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta + i*l.c*l.h*l.w);
- if(l.binary) swap_binary(l);
+ if(l.binary) swap_binary(&l);
}
}
}
--
Gitblit v1.10.0