Faster batch normalization

Joseph Redmon

2015-12-08 c2738835f0a2435ab03f411af3d168aec389d2a6

Faster batch normalization

 src/blas.h

@@ -36,14 +36,12 @@
void variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance);
void normalize_gpu(float *x, float *mean, float *variance, int batch, int filters, int spatial);

void mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta);
void variance_delta_gpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta);
void normalize_delta_gpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta);

void fast_mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *spatial_mean_delta, float *mean_delta);
void fast_variance_delta_gpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *spatial_variance_delta, float *variance_delta);
void fast_mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta);
void fast_variance_delta_gpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta);

void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *spatial_variance, float *variance);
void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *spatial_mean, float *mean);
void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance);
void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *mean);
#endif
#endif

 src/blas_kernels.cu

@@ -48,28 +48,6 @@
    variance_delta[i] *= -.5 * pow(variance[i] + .00001f, (float)(-3./2.));
}

__global__ void spatial_variance_delta_kernel(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *spatial_variance_delta)
{
    int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
    if (i >= batch*filters) return;
    int f = i%filters;
    int b = i/filters;

    int k;
    spatial_variance_delta[i] = 0;
    for (k = 0; k < spatial; ++k) {
        int index = b*filters*spatial + f*spatial + k;
        spatial_variance_delta[i] += delta[index]*(x[index] - mean[f]);
    }
    spatial_variance_delta[i] *= -.5 * pow(variance[f] + .00001f, (float)(-3./2.));
}

extern "C" void variance_delta_gpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta)
{
    variance_delta_kernel<<<cuda_gridsize(filters), BLOCK>>>(x, delta, mean, variance, batch, filters, spatial, variance_delta);
    check_error(cudaPeekAtLastError());
}

__global__ void accumulate_kernel(float *x, int n, int groups, float *sum)
{
    int k;
@@ -81,38 +59,62 @@
    }
}

extern "C" void fast_variance_delta_gpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *spatial_variance_delta, float *variance_delta)
__global__ void fast_mean_delta_kernel(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
{
    spatial_variance_delta_kernel<<<cuda_gridsize(filters*batch), BLOCK>>>(x, delta, mean, variance, batch, filters, spatial, spatial_variance_delta);
    check_error(cudaPeekAtLastError());
    accumulate_kernel<<<cuda_gridsize(filters), BLOCK>>>(spatial_variance_delta, batch, filters, variance_delta);
    check_error(cudaPeekAtLastError());
}
    const int threads = BLOCK;
    __shared__ float local[threads];

__global__ void spatial_mean_delta_kernel(float *delta, float *variance, int batch, int filters, int spatial, float *spatial_mean_delta)
{
    int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
    if (i >= batch*filters) return;
    int f = i%filters;
    int b = i/filters;
    int id = threadIdx.x;
    local[id] = 0;

    int k;
    spatial_mean_delta[i] = 0;
    for (k = 0; k < spatial; ++k) {
        int index = b*filters*spatial + f*spatial + k;
        spatial_mean_delta[i] += delta[index];
    int filter = blockIdx.x;

    int i, j;
    for(j = 0; j < batch; ++j){
        for(i = 0; i < spatial; i += threads){
            int index = j*spatial*filters + filter*spatial + i + id;
            local[id] += (i+id < spatial) ? delta[index] : 0;
        }
    }
    spatial_mean_delta[i] *= (-1./sqrt(variance[f] + .00001f));

    if(id == 0){
        mean_delta[filter] = 0;
        for(i = 0; i < threads; ++i){
            mean_delta[filter] += local[i];
        }
        mean_delta[filter] *= (-1./sqrt(variance[filter] + .00001f));
    }
}

extern "C" void fast_mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *spatial_mean_delta, float *mean_delta)
__global__ void  fast_variance_delta_kernel(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta)
{
    spatial_mean_delta_kernel<<<cuda_gridsize(filters*batch), BLOCK>>>(delta, variance, batch, filters, spatial, spatial_mean_delta);
    check_error(cudaPeekAtLastError());
    accumulate_kernel<<<cuda_gridsize(filters), BLOCK>>>(spatial_mean_delta, batch, filters, mean_delta);
    check_error(cudaPeekAtLastError());
    const int threads = BLOCK;
    __shared__ float local[threads];

    int id = threadIdx.x;
    local[id] = 0;

    int filter = blockIdx.x;

    int i, j;
    for(j = 0; j < batch; ++j){
        for(i = 0; i < spatial; i += threads){
            int index = j*spatial*filters + filter*spatial + i + id;

            local[id] += (i+id < spatial) ? delta[index]*(x[index] - mean[filter]) : 0;
        }
    }

    if(id == 0){
        variance_delta[filter] = 0;
        for(i = 0; i < threads; ++i){
            variance_delta[filter] += local[i];
        }
        variance_delta[filter] *= -.5 * pow(variance[filter] + .00001f, (float)(-3./2.));
    }
}


__global__ void mean_delta_kernel(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
{
    int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
@@ -134,6 +136,18 @@
    check_error(cudaPeekAtLastError());
}

extern "C" void fast_mean_delta_gpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
{
    fast_mean_delta_kernel<<<filters, BLOCK>>>(delta, variance, batch, filters, spatial, mean_delta);
    check_error(cudaPeekAtLastError());
}

extern "C" void fast_variance_delta_gpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta)
{
    fast_variance_delta_kernel<<<filters, BLOCK>>>(x, delta, mean, variance, batch, filters, spatial, variance_delta);
    check_error(cudaPeekAtLastError());
}

__global__ void  mean_kernel(float *x, int batch, int filters, int spatial, float *mean)
{
    float scale = 1./(batch * spatial);
@@ -150,23 +164,6 @@
    mean[i] *= scale;
}

__global__ void spatial_variance_kernel(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
    float scale = 1./(spatial*batch-1);
    int k;
    int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
    if (i >= batch*filters) return;
    int f = i%filters;
    int b = i/filters;

    variance[i] = 0;
    for(k = 0; k < spatial; ++k){
        int index = b*filters*spatial + f*spatial + k;
        variance[i] += pow((x[index] - mean[f]), 2);
    }
    variance[i] *= scale;
}

__global__ void variance_kernel(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
    float scale = 1./(batch * spatial);
@@ -238,28 +235,80 @@
    check_error(cudaPeekAtLastError());
}

__global__ void  fast_mean_kernel(float *x, int batch, int filters, int spatial, float *mean)
{
    const int threads = BLOCK;
    __shared__ float local[threads];

    int id = threadIdx.x;
    local[id] = 0;

    int filter = blockIdx.x;

    int i, j;
    for(j = 0; j < batch; ++j){
        for(i = 0; i < spatial; i += threads){
            int index = j*spatial*filters + filter*spatial + i + id;
            local[id] += (i+id < spatial) ? x[index] : 0;
        }
    }

    if(id == 0){
        mean[filter] = 0;
        for(i = 0; i < threads; ++i){
            mean[filter] += local[i];
        }
        mean[filter] /= spatial * batch;
    }
}

__global__ void  fast_variance_kernel(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
    const int threads = BLOCK;
    __shared__ float local[threads];

    int id = threadIdx.x;
    local[id] = 0;

    int filter = blockIdx.x;

    int i, j;
    for(j = 0; j < batch; ++j){
        for(i = 0; i < spatial; i += threads){
            int index = j*spatial*filters + filter*spatial + i + id;

            local[id] += (i+id < spatial) ? pow((x[index] - mean[filter]), 2) : 0;
        }
    }

    if(id == 0){
        variance[filter] = 0;
        for(i = 0; i < threads; ++i){
            variance[filter] += local[i];
        }
        variance[filter] /= spatial * batch;
    }
}

extern "C" void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *mean)
{
    fast_mean_kernel<<<filters, BLOCK>>>(x, batch, filters, spatial, mean);
    check_error(cudaPeekAtLastError());
}

extern "C" void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
    fast_variance_kernel<<<filters, BLOCK>>>(x, mean, batch, filters, spatial, variance);
    check_error(cudaPeekAtLastError());
}


extern "C" void mean_gpu(float *x, int batch, int filters, int spatial, float *mean)
{
    mean_kernel<<<cuda_gridsize(filters), BLOCK>>>(x, batch, filters, spatial, mean);
    check_error(cudaPeekAtLastError());
}

extern "C" void fast_mean_gpu(float *x, int batch, int filters, int spatial, float *spatial_mean, float *mean)
{
    mean_kernel<<<cuda_gridsize(filters*batch), BLOCK>>>(x, 1, filters*batch, spatial, spatial_mean);
    check_error(cudaPeekAtLastError());
    mean_kernel<<<cuda_gridsize(filters), BLOCK>>>(spatial_mean, batch, filters, 1, mean);
    check_error(cudaPeekAtLastError());
}

extern "C" void fast_variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *spatial_variance, float *variance)
{
    spatial_variance_kernel<<<cuda_gridsize(batch*filters), BLOCK>>>(x, mean, batch, filters, spatial, spatial_variance);
    check_error(cudaPeekAtLastError());
    accumulate_kernel<<<cuda_gridsize(filters), BLOCK>>>(spatial_variance, batch, filters, variance);
    check_error(cudaPeekAtLastError());
}

extern "C" void variance_gpu(float *x, float *mean, int batch, int filters, int spatial, float *variance)
{
    variance_kernel<<<cuda_gridsize(filters), BLOCK>>>(x, mean, batch, filters, spatial, variance);

 src/classifier.c

@@ -98,6 +98,11 @@
            sprintf(buff, "%s/%s_%d.weights",backup_directory,base, epoch);
            save_weights(net, buff);
        }
        if(*net.seen%1000 == 0){
            char buff[256];
            sprintf(buff, "%s/%s.backup",backup_directory,base);
            save_weights(net, buff);
        }
    }
    char buff[256];
    sprintf(buff, "%s/%s.weights", backup_directory, base);

 src/coco.c

@@ -20,7 +20,8 @@
void train_coco(char *cfgfile, char *weightfile)
{
    //char *train_images = "/home/pjreddie/data/voc/test/train.txt";
    char *train_images = "/home/pjreddie/data/coco/train.txt";
    //char *train_images = "/home/pjreddie/data/coco/train.txt";
    char *train_images = "data/coco.trainval.txt";
    char *backup_directory = "/home/pjreddie/backup/";
    srand(time(0));
    data_seed = time(0);

 src/connected_layer.c

@@ -148,6 +148,12 @@
    float * c = l.output_gpu;
    gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);

/*
    cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
    float avg = mean_array(l.output, l.outputs*l.batch);
    printf("%f\n", avg);
    */
}

void backward_connected_layer_gpu(connected_layer l, network_state state)

 src/convolutional_kernels.cu

@@ -119,17 +119,14 @@

    if(l.batch_normalize){
        if(state.train){
            fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.spatial_mean_gpu, l.mean_gpu);   
            fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.spatial_variance_gpu, l.variance_gpu);
            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);

            scal_ongpu(l.n, .95, l.rolling_mean_gpu, 1);
            axpy_ongpu(l.n, .05, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
            scal_ongpu(l.n, .95, l.rolling_variance_gpu, 1);
            axpy_ongpu(l.n, .05, l.variance_gpu, 1, l.rolling_variance_gpu, 1);

            // cuda_pull_array(l.variance_gpu, l.mean, l.n);
            // printf("%f\n", l.mean[0]);

            copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
            normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_h*l.out_w);
            copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
@@ -161,8 +158,8 @@

        scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.n, l.out_h*l.out_w);

        fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.spatial_mean_delta_gpu, l.mean_delta_gpu);
        fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.spatial_variance_delta_gpu, l.variance_delta_gpu);
        fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.mean_delta_gpu);
        fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.variance_delta_gpu);
        normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.n, l.out_w*l.out_h, l.delta_gpu);
    }


 src/convolutional_layer.c

@@ -86,9 +86,8 @@
        }

        l.mean = calloc(n, sizeof(float));
        l.spatial_mean = calloc(n*l.batch, sizeof(float));

        l.variance = calloc(n, sizeof(float));

        l.rolling_mean = calloc(n, sizeof(float));
        l.rolling_variance = calloc(n, sizeof(float));
    }
@@ -114,12 +113,6 @@
        l.rolling_mean_gpu = cuda_make_array(l.mean, n);
        l.rolling_variance_gpu = cuda_make_array(l.variance, n);

        l.spatial_mean_gpu = cuda_make_array(l.spatial_mean, n*l.batch);
        l.spatial_variance_gpu = cuda_make_array(l.spatial_mean, n*l.batch);

        l.spatial_mean_delta_gpu = cuda_make_array(l.spatial_mean, n*l.batch);
        l.spatial_variance_delta_gpu = cuda_make_array(l.spatial_mean, n*l.batch);

        l.mean_delta_gpu = cuda_make_array(l.mean, n);
        l.variance_delta_gpu = cuda_make_array(l.variance, n);