From dda993f3dd3c753dfd580d485b39c1001830fee4 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Thu, 22 Feb 2018 19:54:40 +0000
Subject: [PATCH] Use half_float16 instead of float32 if defined both CUDNN and CUDNN_HALF. Use Tensor Cores.
---
src/convolutional_layer.c | 911 +++++++++++++++++++++++++++++++++++---------------------
1 files changed, 562 insertions(+), 349 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index bae06d3..6865808 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,419 +1,632 @@
#include "convolutional_layer.h"
#include "utils.h"
-#include "mini_blas.h"
+#include "batchnorm_layer.h"
+#include "im2col.h"
+#include "col2im.h"
+#include "blas.h"
+#include "gemm.h"
#include <stdio.h>
#include <time.h>
-int convolutional_out_height(convolutional_layer layer)
+#ifdef CUDNN
+#pragma comment(lib, "cudnn.lib")
+#endif
+
+#ifdef AI2
+#include "xnor_layer.h"
+#endif
+
+#ifndef AI2
+#define AI2 0
+void forward_xnor_layer(layer l, network_state state);
+#endif
+
+void swap_binary(convolutional_layer *l)
{
- int h = layer.h;
- if (!layer.pad) h -= layer.size;
- else h -= 1;
- return h/layer.stride + 1;
+ float *swap = l->weights;
+ l->weights = l->binary_weights;
+ l->binary_weights = swap;
+
+ #ifdef GPU
+ swap = l->weights_gpu;
+ l->weights_gpu = l->binary_weights_gpu;
+ l->binary_weights_gpu = swap;
+ #endif
}
-int convolutional_out_width(convolutional_layer layer)
+void binarize_weights(float *weights, int n, int size, float *binary)
{
- int w = layer.w;
- if (!layer.pad) w -= layer.size;
- else w -= 1;
- return w/layer.stride + 1;
+ int i, f;
+ for(f = 0; f < n; ++f){
+ float mean = 0;
+ for(i = 0; i < size; ++i){
+ mean += fabs(weights[f*size + i]);
+ }
+ mean = mean / size;
+ for(i = 0; i < size; ++i){
+ binary[f*size + i] = (weights[f*size + i] > 0) ? mean : -mean;
+ }
+ }
}
-image get_convolutional_image(convolutional_layer layer)
-{
- int h,w,c;
- h = convolutional_out_height(layer);
- w = convolutional_out_width(layer);
- c = layer.n;
- return float_to_image(h,w,c,layer.output);
-}
-
-image get_convolutional_delta(convolutional_layer layer)
-{
- int h,w,c;
- h = convolutional_out_height(layer);
- w = convolutional_out_width(layer);
- c = layer.n;
- return float_to_image(h,w,c,layer.delta);
-}
-
-convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, float learning_rate, float momentum, float decay)
+void binarize_cpu(float *input, int n, float *binary)
{
int i;
- size = 2*(size/2)+1; //HA! And you thought you'd use an even sized filter...
- convolutional_layer *layer = calloc(1, sizeof(convolutional_layer));
-
- layer->learning_rate = learning_rate;
- layer->momentum = momentum;
- layer->decay = decay;
-
- layer->h = h;
- layer->w = w;
- layer->c = c;
- layer->n = n;
- layer->batch = batch;
- layer->stride = stride;
- layer->size = size;
- layer->pad = pad;
-
- layer->filters = calloc(c*n*size*size, sizeof(float));
- layer->filter_updates = calloc(c*n*size*size, sizeof(float));
- layer->filter_momentum = calloc(c*n*size*size, sizeof(float));
-
- layer->biases = calloc(n, sizeof(float));
- layer->bias_updates = calloc(n, sizeof(float));
- layer->bias_momentum = calloc(n, sizeof(float));
- float scale = 1./(size*size*c);
- scale = .01;
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*2*(rand_uniform()-.5);
for(i = 0; i < n; ++i){
- //layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = .5;
+ binary[i] = (input[i] > 0) ? 1 : -1;
}
- int out_h = convolutional_out_height(*layer);
- int out_w = convolutional_out_width(*layer);
-
- layer->col_image = calloc(out_h*out_w*size*size*c, sizeof(float));
- layer->output = calloc(layer->batch*out_h * out_w * n, sizeof(float));
- layer->delta = calloc(layer->batch*out_h * out_w * n, sizeof(float));
- #ifdef GPU
- layer->filters_cl = cl_make_array(layer->filters, c*n*size*size);
- layer->filter_updates_cl = cl_make_array(layer->filter_updates, c*n*size*size);
- layer->filter_momentum_cl = cl_make_array(layer->filter_momentum, c*n*size*size);
-
- layer->biases_cl = cl_make_array(layer->biases, n);
- layer->bias_updates_cl = cl_make_array(layer->bias_updates, n);
- layer->bias_momentum_cl = cl_make_array(layer->bias_momentum, n);
-
- layer->col_image_cl = cl_make_array(layer->col_image, out_h*out_w*size*size*c);
- layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n);
- layer->output_cl = cl_make_array(layer->output, layer->batch*out_h*out_w*n);
- #endif
- layer->activation = activation;
-
- fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n);
-
- return layer;
}
-void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c)
+void binarize_input(float *input, int n, int size, float *binary)
{
- layer->h = h;
- layer->w = w;
- layer->c = c;
- int out_h = convolutional_out_height(*layer);
- int out_w = convolutional_out_width(*layer);
-
- layer->col_image = realloc(layer->col_image,
- out_h*out_w*layer->size*layer->size*layer->c*sizeof(float));
- layer->output = realloc(layer->output,
- layer->batch*out_h * out_w * layer->n*sizeof(float));
- layer->delta = realloc(layer->delta,
- layer->batch*out_h * out_w * layer->n*sizeof(float));
+ int i, s;
+ for(s = 0; s < size; ++s){
+ float mean = 0;
+ for(i = 0; i < n; ++i){
+ mean += fabs(input[i*size + s]);
+ }
+ mean = mean / n;
+ for(i = 0; i < n; ++i){
+ binary[i*size + s] = (input[i*size + s] > 0) ? mean : -mean;
+ }
+ }
}
-void bias_output(const convolutional_layer layer)
+int convolutional_out_height(convolutional_layer l)
+{
+ return (l.h + 2*l.pad - l.size) / l.stride + 1;
+}
+
+int convolutional_out_width(convolutional_layer l)
+{
+ return (l.w + 2*l.pad - l.size) / l.stride + 1;
+}
+
+image get_convolutional_image(convolutional_layer l)
+{
+ int h,w,c;
+ h = convolutional_out_height(l);
+ w = convolutional_out_width(l);
+ c = l.n;
+ return float_to_image(w,h,c,l.output);
+}
+
+image get_convolutional_delta(convolutional_layer l)
+{
+ int h,w,c;
+ h = convolutional_out_height(l);
+ w = convolutional_out_width(l);
+ c = l.n;
+ return float_to_image(w,h,c,l.delta);
+}
+
+size_t get_workspace_size(layer l){
+#ifdef CUDNN
+ if(gpu_index >= 0){
+ size_t most = 0;
+ size_t s = 0;
+ cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.weightDesc,
+ l.convDesc,
+ l.dstTensorDesc,
+ l.fw_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dweightDesc,
+ l.bf_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
+ l.weightDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dsrcTensorDesc,
+ l.bd_algo,
+ &s);
+ if (s > most) most = s;
+ return most;
+ }
+ #endif
+ return (size_t)l.out_h*l.out_w*l.size*l.size*l.c*sizeof(float);
+}
+
+#ifdef GPU
+#ifdef CUDNN
+void cudnn_convolutional_setup(layer *l, int cudnn_preference)
+{
+
+#ifdef CUDNN_HALF
+ // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0):
+ // Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
+ const cudnnDataType_t data_type = CUDNN_DATA_HALF;
+#else
+ cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
+#endif
+ // Tensor Core uses CUDNN_TENSOR_OP_MATH instead of CUDNN_DEFAULT_MATH
+ cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH);
+
+ // INT8_CONFIG, INT8_EXT_CONFIG, INT8x4_CONFIG and INT8x4_EXT_CONFIG are only supported
+ // on architectures with DP4A support (compute capability 6.1 and later).
+ //cudnnDataType_t data_type = CUDNN_DATA_INT8;
+
+ cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->dweightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->weightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+#if(CUDNN_MAJOR >= 6)
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, data_type); // cudnn >= 6.0
+#else
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION); // cudnn 5.1
+#endif
+ int forward_algo = CUDNN_CONVOLUTION_FWD_PREFER_FASTEST;
+ int backward_algo = CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST;
+ int backward_filter = CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST;
+ if (cudnn_preference == cudnn_smallest)
+ {
+ forward_algo = CUDNN_CONVOLUTION_FWD_NO_WORKSPACE;
+ backward_algo = CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE;
+ backward_filter = CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE;
+ }
+
+ cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->weightDesc,
+ l->convDesc,
+ l->dstTensorDesc,
+ forward_algo,
+ 0,
+ &l->fw_algo);
+ cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
+ l->weightDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dsrcTensorDesc,
+ backward_algo,
+ 0,
+ &l->bd_algo);
+ cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dweightDesc,
+ backward_filter,
+ 0,
+ &l->bf_algo);
+}
+#endif
+#endif
+
+convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int padding, ACTIVATION activation, int batch_normalize, int binary, int xnor, int adam)
+{
+ int i;
+ convolutional_layer l = {0};
+ l.type = CONVOLUTIONAL;
+
+ l.h = h;
+ l.w = w;
+ l.c = c;
+ l.n = n;
+ l.binary = binary;
+ l.xnor = xnor;
+ l.batch = batch;
+ l.stride = stride;
+ l.size = size;
+ l.pad = padding;
+ l.batch_normalize = batch_normalize;
+
+ l.weights = calloc(c*n*size*size, sizeof(float));
+ l.weight_updates = calloc(c*n*size*size, sizeof(float));
+
+ l.biases = calloc(n, sizeof(float));
+ l.bias_updates = calloc(n, sizeof(float));
+
+ // float scale = 1./sqrt(size*size*c);
+ float scale = sqrt(2./(size*size*c));
+ for(i = 0; i < c*n*size*size; ++i) l.weights[i] = scale*rand_uniform(-1, 1);
+ int out_h = convolutional_out_height(l);
+ int out_w = convolutional_out_width(l);
+ l.out_h = out_h;
+ l.out_w = out_w;
+ l.out_c = n;
+ l.outputs = l.out_h * l.out_w * l.out_c;
+ l.inputs = l.w * l.h * l.c;
+
+ l.output = calloc(l.batch*l.outputs, sizeof(float));
+ l.delta = calloc(l.batch*l.outputs, sizeof(float));
+
+ l.forward = forward_convolutional_layer;
+ l.backward = backward_convolutional_layer;
+ l.update = update_convolutional_layer;
+ if(binary){
+ l.binary_weights = calloc(c*n*size*size, sizeof(float));
+ l.cweights = calloc(c*n*size*size, sizeof(char));
+ l.scales = calloc(n, sizeof(float));
+ }
+ if(xnor){
+ l.binary_weights = calloc(c*n*size*size, sizeof(float));
+ l.binary_input = calloc(l.inputs*l.batch, sizeof(float));
+ }
+
+ if(batch_normalize){
+ l.scales = calloc(n, sizeof(float));
+ l.scale_updates = calloc(n, sizeof(float));
+ for(i = 0; i < n; ++i){
+ l.scales[i] = 1;
+ }
+
+ l.mean = calloc(n, sizeof(float));
+ l.variance = calloc(n, sizeof(float));
+
+ l.mean_delta = calloc(n, sizeof(float));
+ l.variance_delta = calloc(n, sizeof(float));
+
+ l.rolling_mean = calloc(n, sizeof(float));
+ l.rolling_variance = calloc(n, sizeof(float));
+ l.x = calloc(l.batch*l.outputs, sizeof(float));
+ l.x_norm = calloc(l.batch*l.outputs, sizeof(float));
+ }
+ if(adam){
+ l.adam = 1;
+ l.m = calloc(c*n*size*size, sizeof(float));
+ l.v = calloc(c*n*size*size, sizeof(float));
+ }
+
+#ifdef GPU
+ l.forward_gpu = forward_convolutional_layer_gpu;
+ l.backward_gpu = backward_convolutional_layer_gpu;
+ l.update_gpu = update_convolutional_layer_gpu;
+
+ if(gpu_index >= 0){
+ if (adam) {
+ l.m_gpu = cuda_make_array(l.m, c*n*size*size);
+ l.v_gpu = cuda_make_array(l.v, c*n*size*size);
+ }
+
+ l.weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+#ifdef CUDNN_HALF
+ l.weights_gpu16 = cuda_make_array(l.weights, c*n*size*size/2);
+#endif
+ l.weight_updates_gpu = cuda_make_array(l.weight_updates, c*n*size*size);
+
+ l.biases_gpu = cuda_make_array(l.biases, n);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
+
+ l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n);
+ l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+
+ if(binary){
+ l.binary_weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+ }
+ if(xnor){
+ l.binary_weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+ l.binary_input_gpu = cuda_make_array(0, l.inputs*l.batch);
+ }
+
+ if(batch_normalize){
+ l.mean_gpu = cuda_make_array(l.mean, n);
+ l.variance_gpu = cuda_make_array(l.variance, n);
+
+ l.rolling_mean_gpu = cuda_make_array(l.mean, n);
+ l.rolling_variance_gpu = cuda_make_array(l.variance, n);
+
+ l.mean_delta_gpu = cuda_make_array(l.mean, n);
+ l.variance_delta_gpu = cuda_make_array(l.variance, n);
+
+ l.scales_gpu = cuda_make_array(l.scales, n);
+ l.scale_updates_gpu = cuda_make_array(l.scale_updates, n);
+
+ l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ }
+#ifdef CUDNN
+ cudnnCreateTensorDescriptor(&l.srcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.weightDesc);
+ cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.dweightDesc);
+ cudnnCreateConvolutionDescriptor(&l.convDesc);
+ cudnn_convolutional_setup(&l, cudnn_fastest);
+#endif
+ }
+#endif
+ l.workspace_size = get_workspace_size(l);
+ l.activation = activation;
+
+ fprintf(stderr, "conv %5d %2d x%2d /%2d %4d x%4d x%4d -> %4d x%4d x%4d\n", n, size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c);
+
+ return l;
+}
+
+void denormalize_convolutional_layer(convolutional_layer l)
+{
+ int i, j;
+ for(i = 0; i < l.n; ++i){
+ float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .00001);
+ for(j = 0; j < l.c*l.size*l.size; ++j){
+ l.weights[i*l.c*l.size*l.size + j] *= scale;
+ }
+ l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
+ }
+}
+
+void test_convolutional_layer()
+{
+ convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 0, 0, 0);
+ l.batch_normalize = 1;
+ float data[] = {1,1,1,1,1,
+ 1,1,1,1,1,
+ 1,1,1,1,1,
+ 1,1,1,1,1,
+ 1,1,1,1,1,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 3,3,3,3,3,
+ 3,3,3,3,3,
+ 3,3,3,3,3,
+ 3,3,3,3,3,
+ 3,3,3,3,3};
+ network_state state = {0};
+ state.input = data;
+ forward_convolutional_layer(l, state);
+}
+
+void resize_convolutional_layer(convolutional_layer *l, int w, int h)
+{
+ int old_w = l->w;
+ int old_h = l->h;
+ l->w = w;
+ l->h = h;
+ int out_w = convolutional_out_width(*l);
+ int out_h = convolutional_out_height(*l);
+
+ l->out_w = out_w;
+ l->out_h = out_h;
+
+ l->outputs = l->out_h * l->out_w * l->out_c;
+ l->inputs = l->w * l->h * l->c;
+
+ l->output = realloc(l->output, l->batch*l->outputs*sizeof(float));
+ l->delta = realloc(l->delta, l->batch*l->outputs*sizeof(float));
+ if(l->batch_normalize){
+ l->x = realloc(l->x, l->batch*l->outputs*sizeof(float));
+ l->x_norm = realloc(l->x_norm, l->batch*l->outputs*sizeof(float));
+ }
+
+#ifdef GPU
+ if (old_w < w || old_h < h) {
+ cuda_free(l->delta_gpu);
+ cuda_free(l->output_gpu);
+
+ l->delta_gpu = cuda_make_array(l->delta, l->batch*l->outputs);
+ l->output_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+
+ if (l->batch_normalize) {
+ cuda_free(l->x_gpu);
+ cuda_free(l->x_norm_gpu);
+
+ l->x_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ l->x_norm_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ }
+ }
+#ifdef CUDNN
+ cudnn_convolutional_setup(l, cudnn_fastest);
+#endif
+#endif
+ l->workspace_size = get_workspace_size(*l);
+
+#ifdef CUDNN
+ // check for excessive memory consumption
+ size_t free_byte;
+ size_t total_byte;
+ check_error(cudaMemGetInfo(&free_byte, &total_byte));
+ if (l->workspace_size > free_byte || l->workspace_size >= total_byte / 2) {
+ printf(" used slow CUDNN algo without Workspace! \n");
+ cudnn_convolutional_setup(l, cudnn_smallest);
+ l->workspace_size = get_workspace_size(*l);
+ }
+#endif
+}
+
+void add_bias(float *output, float *biases, int batch, int n, int size)
{
int i,j,b;
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- for(b = 0; b < layer.batch; ++b){
- for(i = 0; i < layer.n; ++i){
- for(j = 0; j < out_h*out_w; ++j){
- layer.output[(b*layer.n + i)*out_h*out_w + j] = layer.biases[i];
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ for(j = 0; j < size; ++j){
+ output[(b*n + i)*size + j] += biases[i];
}
}
}
}
-void forward_convolutional_layer(const convolutional_layer layer, float *in)
+void scale_bias(float *output, float *scales, int batch, int n, int size)
{
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- int i;
-
- bias_output(layer);
-
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = out_h*out_w;
-
- float *a = layer.filters;
- float *b = layer.col_image;
- float *c = layer.output;
-
-
- for(i = 0; i < layer.batch; ++i){
- im2col_cpu(in, layer.c, layer.h, layer.w,
- layer.size, layer.stride, layer.pad, b);
- gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
- c += n*m;
- in += layer.c*layer.h*layer.w;
- }
- activate_array(layer.output, m*n*layer.batch, layer.activation);
-}
-
-void learn_bias_convolutional_layer(convolutional_layer layer)
-{
- int i,b;
- int size = convolutional_out_height(layer)
- *convolutional_out_width(layer);
- for(b = 0; b < layer.batch; ++b){
- for(i = 0; i < layer.n; ++i){
- layer.bias_updates[i] += sum_array(layer.delta+size*(i+b*layer.n), size);
+ int i,j,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ for(j = 0; j < size; ++j){
+ output[(b*n + i)*size + j] *= scales[i];
+ }
}
}
}
-void backward_convolutional_layer(convolutional_layer layer, float *in, float *delta)
+void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
+{
+ int i,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ bias_updates[i] += sum_array(delta+size*(i+b*n), size);
+ }
+ }
+}
+
+void forward_convolutional_layer(convolutional_layer l, network_state state)
+{
+ int out_h = convolutional_out_height(l);
+ int out_w = convolutional_out_width(l);
+ int i;
+
+ fill_cpu(l.outputs*l.batch, 0, l.output, 1);
+
+ if(l.xnor){
+ binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
+ swap_binary(&l);
+ binarize_cpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input);
+ state.input = l.binary_input;
+ }
+
+ int m = l.n;
+ int k = l.size*l.size*l.c;
+ int n = out_h*out_w;
+
+
+ float *a = l.weights;
+ float *b = state.workspace;
+ float *c = l.output;
+
+ for(i = 0; i < l.batch; ++i){
+ im2col_cpu(state.input, l.c, l.h, l.w,
+ l.size, l.stride, l.pad, b);
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ c += n*m;
+ state.input += l.c*l.h*l.w;
+ }
+
+ if(l.batch_normalize){
+ forward_batchnorm_layer(l, state);
+ }
+ add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
+
+ activate_array(l.output, m*n*l.batch, l.activation);
+ if(l.binary || l.xnor) swap_binary(&l);
+}
+
+void backward_convolutional_layer(convolutional_layer l, network_state state)
{
int i;
- int m = layer.n;
- int n = layer.size*layer.size*layer.c;
- int k = convolutional_out_height(layer)*
- convolutional_out_width(layer);
- gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
- learn_bias_convolutional_layer(layer);
+ int m = l.n;
+ int n = l.size*l.size*l.c;
+ int k = convolutional_out_height(l)*
+ convolutional_out_width(l);
- if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+ gradient_array(l.output, m*k*l.batch, l.activation, l.delta);
+ backward_bias(l.bias_updates, l.delta, l.batch, l.n, k);
- for(i = 0; i < layer.batch; ++i){
- float *a = layer.delta + i*m*k;
- float *b = layer.col_image;
- float *c = layer.filter_updates;
+ if(l.batch_normalize){
+ backward_batchnorm_layer(l, state);
+ }
- float *im = in+i*layer.c*layer.h*layer.w;
+ for(i = 0; i < l.batch; ++i){
+ float *a = l.delta + i*m*k;
+ float *b = state.workspace;
+ float *c = l.weight_updates;
- im2col_cpu(im, layer.c, layer.h, layer.w,
- layer.size, layer.stride, layer.pad, b);
+ float *im = state.input+i*l.c*l.h*l.w;
+
+ im2col_cpu(im, l.c, l.h, l.w,
+ l.size, l.stride, l.pad, b);
gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
- if(delta){
- a = layer.filters;
- b = layer.delta + i*m*k;
- c = layer.col_image;
+ if(state.delta){
+ a = l.weights;
+ b = l.delta + i*m*k;
+ c = state.workspace;
gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
- col2im_cpu(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta+i*layer.c*layer.h*layer.w);
+ col2im_cpu(state.workspace, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w);
}
}
}
-void update_convolutional_layer(convolutional_layer layer)
+void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay)
{
- int size = layer.size*layer.size*layer.c*layer.n;
- axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
- scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1);
+ int size = l.size*l.size*l.c*l.n;
+ axpy_cpu(l.n, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
+ scal_cpu(l.n, momentum, l.bias_updates, 1);
- scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1);
- axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
- scal_cpu(size, layer.momentum, layer.filter_updates, 1);
+ if(l.scales){
+ axpy_cpu(l.n, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
+ scal_cpu(l.n, momentum, l.scale_updates, 1);
+ }
+
+ axpy_cpu(size, -decay*batch, l.weights, 1, l.weight_updates, 1);
+ axpy_cpu(size, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
+ scal_cpu(size, momentum, l.weight_updates, 1);
}
-image get_convolutional_filter(convolutional_layer layer, int i)
+image get_convolutional_weight(convolutional_layer l, int i)
{
- int h = layer.size;
- int w = layer.size;
- int c = layer.c;
- return float_to_image(h,w,c,layer.filters+i*h*w*c);
+ int h = l.size;
+ int w = l.size;
+ int c = l.c;
+ return float_to_image(w,h,c,l.weights+i*h*w*c);
}
-image *weighted_sum_filters(convolutional_layer layer, image *prev_filters)
+void rgbgr_weights(convolutional_layer l)
{
- image *filters = calloc(layer.n, sizeof(image));
- int i,j,k,c;
- if(!prev_filters){
- for(i = 0; i < layer.n; ++i){
- filters[i] = copy_image(get_convolutional_filter(layer, i));
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_weight(l, i);
+ if (im.c == 3) {
+ rgbgr_image(im);
}
}
- else{
- image base = prev_filters[0];
- for(i = 0; i < layer.n; ++i){
- image filter = get_convolutional_filter(layer, i);
- filters[i] = make_image(base.h, base.w, base.c);
- for(j = 0; j < layer.size; ++j){
- for(k = 0; k < layer.size; ++k){
- for(c = 0; c < layer.c; ++c){
- float weight = get_pixel(filter, j, k, c);
- image prev_filter = copy_image(prev_filters[c]);
- scale_image(prev_filter, weight);
- add_into_image(prev_filter, filters[i], 0,0);
- free_image(prev_filter);
- }
- }
- }
- }
- }
- return filters;
}
-image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters)
+void rescale_weights(convolutional_layer l, float scale, float trans)
{
- image *single_filters = weighted_sum_filters(layer, 0);
- show_images(single_filters, layer.n, window);
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_weight(l, i);
+ if (im.c == 3) {
+ scale_image(im, scale);
+ float sum = sum_array(im.data, im.w*im.h*im.c);
+ l.biases[i] += sum*trans;
+ }
+ }
+}
- image delta = get_convolutional_image(layer);
+image *get_weights(convolutional_layer l)
+{
+ image *weights = calloc(l.n, sizeof(image));
+ int i;
+ for(i = 0; i < l.n; ++i){
+ weights[i] = copy_image(get_convolutional_weight(l, i));
+ //normalize_image(weights[i]);
+ }
+ return weights;
+}
+
+image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_weights)
+{
+ image *single_weights = get_weights(l);
+ show_images(single_weights, l.n, window);
+
+ image delta = get_convolutional_image(l);
image dc = collapse_image_layers(delta, 1);
char buff[256];
sprintf(buff, "%s: Output", window);
//show_image(dc, buff);
//save_image(dc, buff);
free_image(dc);
- return single_filters;
+ return single_weights;
}
-#ifdef GPU
-
-cl_kernel get_convolutional_learn_bias_kernel()
-{
- static int init = 0;
- static cl_kernel kernel;
- if(!init){
- kernel = get_kernel("src/convolutional_layer.cl", "learn_bias", 0);
- init = 1;
- }
- return kernel;
-}
-
-void learn_bias_convolutional_layer_ongpu(convolutional_layer layer)
-{
- int size = convolutional_out_height(layer) * convolutional_out_width(layer);
-
- cl_setup();
- cl_kernel kernel = get_convolutional_learn_bias_kernel();
- cl_command_queue queue = cl.queue;
-
- cl_uint i = 0;
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.batch), (void*) &layer.batch);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n);
- cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.delta_cl), (void*) &layer.delta_cl);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.bias_updates_cl), (void*) &layer.bias_updates_cl);
- check_error(cl);
-
- const size_t global_size[] = {layer.n};
-
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0);
- check_error(cl);
-}
-
-cl_kernel get_convolutional_bias_kernel()
-{
- static int init = 0;
- static cl_kernel kernel;
- if(!init){
- kernel = get_kernel("src/convolutional_layer.cl", "bias", 0);
- init = 1;
- }
- return kernel;
-}
-
-void bias_output_gpu(const convolutional_layer layer)
-{
- int out_h = convolutional_out_height(layer);
- int out_w = convolutional_out_width(layer);
- int size = out_h*out_w;
-
- cl_setup();
- cl_kernel kernel = get_convolutional_bias_kernel();
- cl_command_queue queue = cl.queue;
-
- cl_uint i = 0;
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.n), (void*) &layer.n);
- cl.error = clSetKernelArg(kernel, i++, sizeof(size), (void*) &size);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.biases_cl), (void*) &layer.biases_cl);
- cl.error = clSetKernelArg(kernel, i++, sizeof(layer.output_cl), (void*) &layer.output_cl);
- check_error(cl);
-
- const size_t global_size[] = {layer.n*size, layer.batch};
-
- cl.error = clEnqueueNDRangeKernel(queue, kernel, 2, 0, global_size, 0, 0, 0, 0);
- check_error(cl);
-}
-
-//#define TIMEIT
-
-void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
-{
- int i;
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = convolutional_out_height(layer)*
- convolutional_out_width(layer);
-
- bias_output_gpu(layer);
-
- for(i = 0; i < layer.batch; ++i){
- im2col_ongpu(in, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
- cl_mem a = layer.filters_cl;
- cl_mem b = layer.col_image_cl;
- cl_mem c = layer.output_cl;
- gemm_ongpu_offset(0,0,m,n,k,1.,a,0,k,b,0,n,1.,c,i*m*n,n);
- }
- activate_array_ongpu(layer.output_cl, m*n*layer.batch, layer.activation);
-}
-
-void backward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in, cl_mem delta_cl)
-{
- int i;
- int m = layer.n;
- int n = layer.size*layer.size*layer.c;
- int k = convolutional_out_height(layer)*
- convolutional_out_width(layer);
- gradient_array_ongpu(layer.output_cl, m*k*layer.batch, layer.activation, layer.delta_cl);
- learn_bias_convolutional_layer_ongpu(layer);
-
- if(delta_cl) scal_ongpu(layer.batch*layer.h*layer.w*layer.c, 0, delta_cl, 1);
-
- for(i = 0; i < layer.batch; ++i){
- cl_mem a = layer.delta_cl;
- cl_mem b = layer.col_image_cl;
- cl_mem c = layer.filter_updates_cl;
-
- im2col_ongpu(in, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, layer.col_image_cl);
- gemm_ongpu_offset(0,1,m,n,k,1,a,i*m*k,k,b,0,k,1,c,0,n);
-
- if(delta_cl){
-
- cl_mem a = layer.filters_cl;
- cl_mem b = layer.delta_cl;
- cl_mem c = layer.col_image_cl;
-
- gemm_ongpu_offset(1,0,n,k,m,1,a,0,n,b,i*k*m,k,0,c,0,k);
-
- col2im_ongpu(layer.col_image_cl, i*layer.c*layer.h*layer.w, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta_cl);
- }
- }
-}
-
-void pull_convolutional_layer(convolutional_layer layer)
-{
- cl_read_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
- cl_read_array(layer.biases_cl, layer.biases, layer.n);
-}
-
-void push_convolutional_layer(convolutional_layer layer)
-{
- cl_write_array(layer.filters_cl, layer.filters, layer.c*layer.n*layer.size*layer.size);
- cl_write_array(layer.biases_cl, layer.biases, layer.n);
-}
-
-void update_convolutional_layer_gpu(convolutional_layer layer)
-{
- int size = layer.size*layer.size*layer.c*layer.n;
- axpy_ongpu(layer.n, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
- scal_ongpu(layer.n,layer.momentum, layer.bias_updates_cl, 1);
-
- scal_ongpu(size, 1.-layer.learning_rate*layer.decay, layer.filters_cl, 1);
- axpy_ongpu(size, layer.learning_rate, layer.filter_updates_cl, 1, layer.filters_cl, 1);
- scal_ongpu(size, layer.momentum, layer.filter_updates_cl, 1);
- pull_convolutional_layer(layer);
-}
-
-
-#endif
-
--
Gitblit v1.10.0