From a19604bf5fed9ca1c708e96bfe4879144b33b90f Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 08 Sep 2016 05:52:11 +0000
Subject: [PATCH] ok back
---
src/convolutional_layer.c | 793 ++++++++++++++++++++++++++++++++++++++------------------
1 files changed, 536 insertions(+), 257 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 6a103f6..ad2d8a5 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,298 +1,577 @@
#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>
-image get_convolutional_image(convolutional_layer layer)
+#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,w,c;
- h = layer.out_h;
- w = layer.out_w;
- c = layer.n;
- return float_to_image(h,w,c,layer.output);
+ float *swap = l->filters;
+ l->filters = l->binary_filters;
+ l->binary_filters = swap;
+
+ #ifdef GPU
+ swap = l->filters_gpu;
+ l->filters_gpu = l->binary_filters_gpu;
+ l->binary_filters_gpu = swap;
+ #endif
}
-image get_convolutional_delta(convolutional_layer layer)
+void binarize_filters(float *filters, int n, int size, float *binary)
{
- int h,w,c;
- h = layer.out_h;
- w = layer.out_w;
- c = layer.n;
- return float_to_image(h,w,c,layer.delta);
+ int i, f;
+ for(f = 0; f < n; ++f){
+ float mean = 0;
+ for(i = 0; i < size; ++i){
+ mean += fabs(filters[f*size + i]);
+ }
+ mean = mean / size;
+ for(i = 0; i < size; ++i){
+ binary[f*size + i] = (filters[f*size + i] > 0) ? mean : -mean;
+ }
+ }
}
-convolutional_layer *make_convolutional_layer(int h, int w, int c, int n, int size, int stride, ACTIVATION activation)
+void binarize_cpu(float *input, int n, float *binary)
{
int i;
- int out_h,out_w;
- 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->h = h;
- layer->w = w;
- layer->c = c;
- layer->n = n;
- layer->stride = stride;
- layer->size = size;
-
- 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);
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform());
for(i = 0; i < n; ++i){
- //layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = 0;
+ binary[i] = (input[i] > 0) ? 1 : -1;
}
- out_h = (h-size)/stride + 1;
- out_w = (w-size)/stride + 1;
+}
- layer->col_image = calloc(out_h*out_w*size*size*c, sizeof(float));
- layer->output = calloc(out_h * out_w * n, sizeof(float));
- layer->delta = calloc(out_h * out_w * n, sizeof(float));
- layer->activation = activation;
- layer->out_h = out_h;
- layer->out_w = out_w;
+void binarize_input(float *input, int n, int size, float *binary)
+{
+ 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;
+ }
+ }
+}
+
+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.filterDesc,
+ l.convDesc,
+ l.dstTensorDesc,
+ l.fw_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dfilterDesc,
+ l.bf_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
+ l.filterDesc,
+ 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)
+{
+ cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->dfilterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->filterDesc, CUDNN_DATA_FLOAT, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION);
+ cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->filterDesc,
+ l->convDesc,
+ l->dstTensorDesc,
+ CUDNN_CONVOLUTION_FWD_PREFER_FASTEST,
+ 0,
+ &l->fw_algo);
+ cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
+ l->filterDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dsrcTensorDesc,
+ CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST,
+ 0,
+ &l->bd_algo);
+ cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dfilterDesc,
+ CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST,
+ 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 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.filters = calloc(c*n*size*size, sizeof(float));
+ l.filter_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.filters[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*out_h * out_w * n, sizeof(float));
+ l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float));
+
+ if(binary){
+ l.binary_filters = calloc(c*n*size*size, sizeof(float));
+ l.cfilters = calloc(c*n*size*size, sizeof(char));
+ l.scales = calloc(n, sizeof(float));
+ }
+ if(xnor){
+ l.binary_filters = 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.rolling_mean = calloc(n, sizeof(float));
+ l.rolling_variance = calloc(n, sizeof(float));
+ }
+
+#ifdef GPU
+ if(gpu_index >= 0){
+ l.filters_gpu = cuda_make_array(l.filters, c*n*size*size);
+ l.filter_updates_gpu = cuda_make_array(l.filter_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.scales_gpu = cuda_make_array(l.scales, n);
+ l.scale_updates_gpu = cuda_make_array(l.scale_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_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
+ }
+ if(xnor){
+ l.binary_filters_gpu = cuda_make_array(l.filters, 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.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.filterDesc);
+ cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.dfilterDesc);
+ cudnnCreateConvolutionDescriptor(&l.convDesc);
+ cudnn_convolutional_setup(&l);
+#endif
+ }
+#endif
+ l.workspace_size = get_workspace_size(l);
+ l.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);
- srand(0);
- return layer;
+ return l;
}
-void forward_convolutional_layer(const convolutional_layer layer, float *in)
+void denormalize_convolutional_layer(convolutional_layer l)
{
- int i;
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = ((layer.h-layer.size)/layer.stride + 1)*
- ((layer.w-layer.size)/layer.stride + 1);
-
- memset(layer.output, 0, m*n*sizeof(float));
-
- float *a = layer.filters;
- float *b = layer.col_image;
- float *c = layer.output;
-
- im2col_cpu(in, layer.c, layer.h, layer.w, layer.size, layer.stride, b);
- gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
-
- for(i = 0; i < m*n; ++i){
- layer.output[i] = activate(layer.output[i], layer.activation);
- }
- //for(i = 0; i < m*n; ++i) if(i%(m*n/10+1)==0) printf("%f, ", layer.output[i]); printf("\n");
-
-}
-
-void gradient_delta_convolutional_layer(convolutional_layer layer)
-{
- int i;
- for(i = 0; i < layer.out_h*layer.out_w*layer.n; ++i){
- layer.delta[i] *= gradient(layer.output[i], layer.activation);
- }
-}
-
-void learn_bias_convolutional_layer(convolutional_layer layer)
-{
- int i,j;
- int size = layer.out_h*layer.out_w;
- for(i = 0; i < layer.n; ++i){
- float sum = 0;
- for(j = 0; j < size; ++j){
- sum += layer.delta[j+i*size];
+ 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.filters[i*l.c*l.size*l.size + j] *= scale;
}
- layer.bias_updates[i] += sum/size;
+ l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
}
}
-void learn_convolutional_layer(convolutional_layer layer)
-{
- gradient_delta_convolutional_layer(layer);
- learn_bias_convolutional_layer(layer);
- int m = layer.n;
- int n = layer.size*layer.size*layer.c;
- int k = ((layer.h-layer.size)/layer.stride + 1)*
- ((layer.w-layer.size)/layer.stride + 1);
-
- float *a = layer.delta;
- float *b = layer.col_image;
- float *c = layer.filter_updates;
-
- gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
-}
-
-void backward_convolutional_layer(convolutional_layer layer, float *delta)
-{
- int m = layer.size*layer.size*layer.c;
- int k = layer.n;
- int n = ((layer.h-layer.size)/layer.stride + 1)*
- ((layer.w-layer.size)/layer.stride + 1);
-
- float *a = layer.filters;
- float *b = layer.delta;
- float *c = layer.col_image;
-
-
- memset(c, 0, m*n*sizeof(float));
- gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
-
- memset(delta, 0, layer.h*layer.w*layer.c*sizeof(float));
- col2im_cpu(c, layer.c, layer.h, layer.w, layer.size, layer.stride, delta);
-}
-
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
-{
- int i;
- int size = layer.size*layer.size*layer.c*layer.n;
- for(i = 0; i < layer.n; ++i){
- layer.biases[i] += step*layer.bias_updates[i];
- layer.bias_updates[i] *= momentum;
- }
- for(i = 0; i < size; ++i){
- layer.filters[i] += step*(layer.filter_updates[i] - decay*layer.filters[i]);
- layer.filter_updates[i] *= momentum;
- }
-}
-/*
-
-void backward_convolutional_layer2(convolutional_layer layer, float *input, float *delta)
-{
- image in_delta = float_to_image(layer.h, layer.w, layer.c, delta);
- image out_delta = get_convolutional_delta(layer);
- int i,j;
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
-
- zero_image(in_delta);
- upsample_image(out_delta, layer.stride, layer.upsampled);
- for(j = 0; j < in_delta.c; ++j){
- for(i = 0; i < layer.n; ++i){
- two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, in_delta, j, layer.edge);
- }
- }
-
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
-}
-
-
-void learn_convolutional_layer(convolutional_layer layer, float *input)
-{
- int i;
- image in_image = float_to_image(layer.h, layer.w, layer.c, input);
- image out_delta = get_convolutional_delta(layer);
- gradient_delta_convolutional_layer(layer);
- for(i = 0; i < layer.n; ++i){
- kernel_update(in_image, layer.kernel_updates[i], layer.stride, i, out_delta, layer.edge);
- layer.bias_updates[i] += avg_image_layer(out_delta, i);
- }
-}
-
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
-{
- int i,j;
- for(i = 0; i < layer.n; ++i){
- layer.bias_momentum[i] = step*(layer.bias_updates[i])
- + momentum*layer.bias_momentum[i];
- layer.biases[i] += layer.bias_momentum[i];
- layer.bias_updates[i] = 0;
- int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c;
- for(j = 0; j < pixels; ++j){
- layer.kernel_momentum[i].data[j] = step*(layer.kernel_updates[i].data[j] - decay*layer.kernels[i].data[j])
- + momentum*layer.kernel_momentum[i].data[j];
- layer.kernels[i].data[j] += layer.kernel_momentum[i].data[j];
- }
- zero_image(layer.kernel_updates[i]);
- }
-}
-*/
-
void test_convolutional_layer()
{
- convolutional_layer l = *make_convolutional_layer(4,4,1,1,3,1,LINEAR);
- float input[] = {1,2,3,4,
- 5,6,7,8,
- 9,10,11,12,
- 13,14,15,16};
- float filter[] = {.5, 0, .3,
- 0 , 1, 0,
- .2 , 0, 1};
- float delta[] = {1, 2,
- 3, 4};
- float in_delta[] = {.5,1,.3,.6,
- 5,6,7,8,
- 9,10,11,12,
- 13,14,15,16};
- l.filters = filter;
- forward_convolutional_layer(l, input);
- l.delta = delta;
- learn_convolutional_layer(l);
- image filter_updates = float_to_image(3,3,1,l.filter_updates);
- print_image(filter_updates);
- printf("Delta:\n");
- backward_convolutional_layer(l, in_delta);
- pm(4,4,in_delta);
+ convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 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);
}
-image get_convolutional_filter(convolutional_layer layer, int i)
+void resize_convolutional_layer(convolutional_layer *l, int w, int h)
{
- 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);
+ 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*out_h * out_w * l->n*sizeof(float));
+ l->delta = realloc(l->delta,
+ l->batch*out_h * out_w * l->n*sizeof(float));
+
+#ifdef GPU
+ cuda_free(l->delta_gpu);
+ cuda_free(l->output_gpu);
+
+ l->delta_gpu = cuda_make_array(l->delta, l->batch*out_h*out_w*l->n);
+ l->output_gpu = cuda_make_array(l->output, l->batch*out_h*out_w*l->n);
+#ifdef CUDNN
+ cudnn_convolutional_setup(l);
+#endif
+#endif
+ l->workspace_size = get_workspace_size(*l);
}
-void visualize_convolutional_layer(convolutional_layer layer, char *window)
+void add_bias(float *output, float *biases, int batch, int n, int size)
{
- int color = 1;
- int border = 1;
- int h,w,c;
- int size = layer.size;
- h = size;
- w = (size + border) * layer.n - border;
- c = layer.c;
- if(c != 3 || !color){
- h = (h+border)*c - border;
- c = 1;
- }
-
- image filters = make_image(h,w,c);
- int i,j;
- for(i = 0; i < layer.n; ++i){
- int w_offset = i*(size+border);
- image k = get_convolutional_filter(layer, i);
- //printf("%f ** ", layer.biases[i]);
- //print_image(k);
- image copy = copy_image(k);
- normalize_image(copy);
- for(j = 0; j < k.c; ++j){
- //set_pixel(copy,0,0,j,layer.biases[i]);
- }
- if(c == 3 && color){
- embed_image(copy, filters, 0, w_offset);
- }
- else{
- for(j = 0; j < k.c; ++j){
- int h_offset = j*(size+border);
- image layer = get_image_layer(k, j);
- embed_image(layer, filters, h_offset, w_offset);
- free_image(layer);
+ 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] += biases[i];
}
}
- free_image(copy);
}
- image delta = get_convolutional_delta(layer);
+}
+
+void scale_bias(float *output, float *scales, int batch, int n, int 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_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.binary){
+ binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
+ binarize_filters2(l.filters, l.n, l.c*l.size*l.size, l.cfilters, l.scales);
+ swap_binary(&l);
+ }
+ */
+
+ /*
+ if(l.binary){
+ int m = l.n;
+ int k = l.size*l.size*l.c;
+ int n = out_h*out_w;
+
+ char *a = l.cfilters;
+ 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_bin(m,n,k,1,a,k,b,n,c,n);
+ c += n*m;
+ state.input += l.c*l.h*l.w;
+ }
+ scale_bias(l.output, l.scales, l.batch, l.n, out_h*out_w);
+ add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
+ activate_array(l.output, m*n*l.batch, l.activation);
+ return;
+ }
+ */
+
+ if(l.xnor){
+ binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
+ 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;
+
+ if (l.xnor && l.c%32 == 0 && AI2) {
+ forward_xnor_layer(l, state);
+ printf("xnor\n");
+ } else {
+
+ float *a = l.filters;
+ 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 = l.n;
+ int n = l.size*l.size*l.c;
+ int k = convolutional_out_height(l)*
+ convolutional_out_width(l);
+
+ 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 < l.batch; ++i){
+ float *a = l.delta + i*m*k;
+ float *b = state.workspace;
+ float *c = l.filter_updates;
+
+ 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(state.delta){
+ a = l.filters;
+ 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(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 l, int batch, float learning_rate, float momentum, float decay)
+{
+ 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);
+
+ axpy_cpu(size, -decay*batch, l.filters, 1, l.filter_updates, 1);
+ axpy_cpu(size, learning_rate/batch, l.filter_updates, 1, l.filters, 1);
+ scal_cpu(size, momentum, l.filter_updates, 1);
+}
+
+
+image get_convolutional_filter(convolutional_layer l, int i)
+{
+ int h = l.size;
+ int w = l.size;
+ int c = l.c;
+ return float_to_image(w,h,c,l.filters+i*h*w*c);
+}
+
+void rgbgr_filters(convolutional_layer l)
+{
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_filter(l, i);
+ if (im.c == 3) {
+ rgbgr_image(im);
+ }
+ }
+}
+
+void rescale_filters(convolutional_layer l, float scale, float trans)
+{
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_filter(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 *get_filters(convolutional_layer l)
+{
+ image *filters = calloc(l.n, sizeof(image));
+ int i;
+ for(i = 0; i < l.n; ++i){
+ filters[i] = copy_image(get_convolutional_filter(l, i));
+ //normalize_image(filters[i]);
+ }
+ return filters;
+}
+
+image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_filters)
+{
+ image *single_filters = get_filters(l);
+ show_images(single_filters, l.n, window);
+
+ image delta = get_convolutional_image(l);
image dc = collapse_image_layers(delta, 1);
char buff[256];
- sprintf(buff, "%s: Delta", window);
- show_image(dc, buff);
+ sprintf(buff, "%s: Output", window);
+ //show_image(dc, buff);
+ //save_image(dc, buff);
free_image(dc);
- show_image(filters, window);
- free_image(filters);
+ return single_filters;
}
--
Gitblit v1.10.0