From d9f1b0b16edeb59281355a855e18a8be343fc33c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 08 Aug 2014 19:04:15 +0000
Subject: [PATCH] probably how maxpool layers should be
---
src/network.c | 260 +++++++++++++++++++++++++---------------------------
1 files changed, 125 insertions(+), 135 deletions(-)
diff --git a/src/network.c b/src/network.c
index b75eddf..ed927a8 100644
--- a/src/network.c
+++ b/src/network.c
@@ -9,6 +9,7 @@
#include "maxpool_layer.h"
#include "normalization_layer.h"
#include "softmax_layer.h"
+#include "dropout_layer.h"
network make_network(int n, int batch)
{
@@ -25,98 +26,9 @@
return net;
}
-void print_convolutional_cfg(FILE *fp, convolutional_layer *l, int first)
-{
- int i;
- fprintf(fp, "[convolutional]\n");
- if(first) fprintf(fp, "batch=%d\n"
- "height=%d\n"
- "width=%d\n"
- "channels=%d\n",
- l->batch,l->h, l->w, l->c);
- fprintf(fp, "filters=%d\n"
- "size=%d\n"
- "stride=%d\n"
- "activation=%s\n",
- l->n, l->size, l->stride,
- get_activation_string(l->activation));
- fprintf(fp, "data=");
- for(i = 0; i < l->n; ++i) fprintf(fp, "%g,", l->biases[i]);
- for(i = 0; i < l->n*l->c*l->size*l->size; ++i) fprintf(fp, "%g,", l->filters[i]);
- fprintf(fp, "\n\n");
-}
-void print_connected_cfg(FILE *fp, connected_layer *l, int first)
-{
- int i;
- fprintf(fp, "[connected]\n");
- if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs);
- fprintf(fp, "output=%d\n"
- "activation=%s\n",
- l->outputs,
- get_activation_string(l->activation));
- fprintf(fp, "data=");
- for(i = 0; i < l->outputs; ++i) fprintf(fp, "%g,", l->biases[i]);
- for(i = 0; i < l->inputs*l->outputs; ++i) fprintf(fp, "%g,", l->weights[i]);
- fprintf(fp, "\n\n");
-}
-
-void print_maxpool_cfg(FILE *fp, maxpool_layer *l, int first)
-{
- fprintf(fp, "[maxpool]\n");
- if(first) fprintf(fp, "batch=%d\n"
- "height=%d\n"
- "width=%d\n"
- "channels=%d\n",
- l->batch,l->h, l->w, l->c);
- fprintf(fp, "stride=%d\n\n", l->stride);
-}
-
-void print_normalization_cfg(FILE *fp, normalization_layer *l, int first)
-{
- fprintf(fp, "[localresponsenormalization]\n");
- if(first) fprintf(fp, "batch=%d\n"
- "height=%d\n"
- "width=%d\n"
- "channels=%d\n",
- l->batch,l->h, l->w, l->c);
- fprintf(fp, "size=%d\n"
- "alpha=%g\n"
- "beta=%g\n"
- "kappa=%g\n\n", l->size, l->alpha, l->beta, l->kappa);
-}
-
-void print_softmax_cfg(FILE *fp, softmax_layer *l, int first)
-{
- fprintf(fp, "[softmax]\n");
- if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs);
- fprintf(fp, "\n");
-}
-
-void save_network(network net, char *filename)
-{
- FILE *fp = fopen(filename, "w");
- if(!fp) file_error(filename);
- int i;
- for(i = 0; i < net.n; ++i)
- {
- if(net.types[i] == CONVOLUTIONAL)
- print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i], i==0);
- else if(net.types[i] == CONNECTED)
- print_connected_cfg(fp, (connected_layer *)net.layers[i], i==0);
- else if(net.types[i] == MAXPOOL)
- print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i], i==0);
- else if(net.types[i] == NORMALIZATION)
- print_normalization_cfg(fp, (normalization_layer *)net.layers[i], i==0);
- else if(net.types[i] == SOFTMAX)
- print_softmax_cfg(fp, (softmax_layer *)net.layers[i], i==0);
- }
- fclose(fp);
-}
-
+#ifdef GPU
void forward_network(network net, float *input, int train)
{
- int i;
- #ifdef GPU
cl_setup();
size_t size = get_network_input_size(net);
if(!net.input_cl){
@@ -126,16 +38,12 @@
}
cl_write_array(net.input_cl, input, size);
cl_mem input_cl = net.input_cl;
- #endif
+ int i;
for(i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- #ifdef GPU
forward_convolutional_layer_gpu(layer, input_cl);
input_cl = layer.output_cl;
- #else
- forward_convolutional_layer(layer, input);
- #endif
input = layer.output;
}
else if(net.types[i] == CONNECTED){
@@ -161,13 +69,53 @@
}
}
-void update_network(network net, float step, float momentum, float decay)
+#else
+
+void forward_network(network net, float *input, int train)
{
int i;
for(i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- update_convolutional_layer(layer, step, momentum, decay);
+ forward_convolutional_layer(layer, input);
+ input = layer.output;
+ }
+ else if(net.types[i] == CONNECTED){
+ connected_layer layer = *(connected_layer *)net.layers[i];
+ forward_connected_layer(layer, input);
+ input = layer.output;
+ }
+ else if(net.types[i] == SOFTMAX){
+ softmax_layer layer = *(softmax_layer *)net.layers[i];
+ forward_softmax_layer(layer, input);
+ input = layer.output;
+ }
+ else if(net.types[i] == MAXPOOL){
+ maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ forward_maxpool_layer(layer, input);
+ input = layer.output;
+ }
+ else if(net.types[i] == NORMALIZATION){
+ normalization_layer layer = *(normalization_layer *)net.layers[i];
+ forward_normalization_layer(layer, input);
+ input = layer.output;
+ }
+ else if(net.types[i] == DROPOUT){
+ if(!train) continue;
+ dropout_layer layer = *(dropout_layer *)net.layers[i];
+ forward_dropout_layer(layer, input);
+ }
+ }
+}
+#endif
+
+void update_network(network net)
+{
+ int i;
+ for(i = 0; i < net.n; ++i){
+ if(net.types[i] == CONVOLUTIONAL){
+ convolutional_layer layer = *(convolutional_layer *)net.layers[i];
+ update_convolutional_layer(layer);
}
else if(net.types[i] == MAXPOOL){
//maxpool_layer layer = *(maxpool_layer *)net.layers[i];
@@ -180,7 +128,7 @@
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer(layer, step, momentum, decay);
+ update_connected_layer(layer);
}
}
}
@@ -196,6 +144,8 @@
} else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.output;
+ } else if(net.types[i] == DROPOUT){
+ return get_network_output_layer(net, i-1);
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.output;
@@ -221,6 +171,8 @@
} else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.delta;
+ } else if(net.types[i] == DROPOUT){
+ return get_network_delta_layer(net, i-1);
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.delta;
@@ -238,10 +190,13 @@
float sum = 0;
float *delta = get_network_delta(net);
float *out = get_network_output(net);
- int i, k = get_network_output_size(net);
- for(i = 0; i < k; ++i){
- //printf("%f, ", out[i]);
+ int i;
+ for(i = 0; i < get_network_output_size(net)*net.batch; ++i){
+ //if(i %get_network_output_size(net) == 0) printf("\n");
+ //printf("%5.2f %5.2f, ", out[i], truth[i]);
+ //if(i == get_network_output_size(net)) printf("\n");
delta[i] = truth[i] - out[i];
+ //printf("%.10f, ", out[i]);
sum += delta[i]*delta[i];
}
//printf("\n");
@@ -293,32 +248,50 @@
return error;
}
-float train_network_datum(network net, float *x, float *y, float step, float momentum, float decay)
+float train_network_datum(network net, float *x, float *y)
{
forward_network(net, x, 1);
//int class = get_predicted_class_network(net);
float error = backward_network(net, x, y);
- update_network(net, step, momentum, decay);
+ update_network(net);
//return (y[class]?1:0);
return error;
}
-float train_network_sgd(network net, data d, int n, float step, float momentum,float decay)
+float train_network_sgd(network net, data d, int n)
{
- int i;
- float error = 0;
- int correct = 0;
- int pos = 0;
+ int batch = net.batch;
+ float *X = calloc(batch*d.X.cols, sizeof(float));
+ float *y = calloc(batch*d.y.cols, sizeof(float));
+
+ int i,j;
+ float sum = 0;
for(i = 0; i < n; ++i){
- int index = rand()%d.X.rows;
- float err = train_network_datum(net, d.X.vals[index], d.y.vals[index], step, momentum, decay);
+ for(j = 0; j < batch; ++j){
+ int index = rand()%d.X.rows;
+ memcpy(X+j*d.X.cols, d.X.vals[index], d.X.cols*sizeof(float));
+ memcpy(y+j*d.y.cols, d.y.vals[index], d.y.cols*sizeof(float));
+ }
+ float err = train_network_datum(net, X, y);
+ sum += err;
+ //train_network_datum(net, X, y);
+ /*
float *y = d.y.vals[index];
int class = get_predicted_class_network(net);
correct += (y[class]?1:0);
- if(y[1]){
- error += err;
- ++pos;
+ */
+
+/*
+ for(j = 0; j < d.y.cols*batch; ++j){
+ printf("%6.3f ", y[j]);
}
+ printf("\n");
+ for(j = 0; j < d.y.cols*batch; ++j){
+ printf("%6.3f ", get_network_output(net)[j]);
+ }
+ printf("\n");
+ printf("\n");
+ */
//printf("%d %f %f\n", i,net.output[0], d.y.vals[index][0]);
@@ -327,33 +300,35 @@
//}
}
//printf("Accuracy: %f\n",(float) correct/n);
- return error/pos;
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
}
-float train_network_batch(network net, data d, int n, float step, float momentum,float decay)
+float train_network_batch(network net, data d, int n)
{
- int i;
- int correct = 0;
+ int i,j;
+ float sum = 0;
+ int batch = 2;
for(i = 0; i < n; ++i){
- int index = rand()%d.X.rows;
- float *x = d.X.vals[index];
- float *y = d.y.vals[index];
- forward_network(net, x, 1);
- int class = get_predicted_class_network(net);
- backward_network(net, x, y);
- correct += (y[class]?1:0);
+ for(j = 0; j < batch; ++j){
+ int index = rand()%d.X.rows;
+ float *x = d.X.vals[index];
+ float *y = d.y.vals[index];
+ forward_network(net, x, 1);
+ sum += backward_network(net, x, y);
+ }
+ update_network(net);
}
- update_network(net, step, momentum, decay);
- return (float)correct/n;
-
+ return (float)sum/(n*batch);
}
-void train_network(network net, data d, float step, float momentum, float decay)
+void train_network(network net, data d)
{
int i;
int correct = 0;
for(i = 0; i < d.X.rows; ++i){
- correct += train_network_datum(net, d.X.vals[i], d.y.vals[i], step, momentum, decay);
+ correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]);
if(i%100 == 0){
visualize_network(net);
cvWaitKey(10);
@@ -377,6 +352,9 @@
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.inputs;
+ } else if(net.types[i] == DROPOUT){
+ dropout_layer layer = *(dropout_layer *) net.layers[i];
+ return layer.inputs;
}
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
@@ -400,6 +378,9 @@
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
+ } else if(net.types[i] == DROPOUT){
+ dropout_layer layer = *(dropout_layer *) net.layers[i];
+ return layer.inputs;
}
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
@@ -448,7 +429,7 @@
int get_network_input_size(network net)
{
- return get_network_output_size_layer(net, 0);
+ return get_network_input_size_layer(net, 0);
}
image get_network_image_layer(network net, int i)
@@ -505,15 +486,24 @@
matrix network_predict_data(network net, data test)
{
- int i,j;
+ int i,j,b;
int k = get_network_output_size(net);
matrix pred = make_matrix(test.X.rows, k);
- for(i = 0; i < test.X.rows; ++i){
- float *out = network_predict(net, test.X.vals[i]);
- for(j = 0; j < k; ++j){
- pred.vals[i][j] = out[j];
+ float *X = calloc(net.batch*test.X.rows, sizeof(float));
+ for(i = 0; i < test.X.rows; i += net.batch){
+ for(b = 0; b < net.batch; ++b){
+ if(i+b == test.X.rows) break;
+ memcpy(X+b*test.X.cols, test.X.vals[i+b], test.X.cols*sizeof(float));
+ }
+ float *out = network_predict(net, X);
+ for(b = 0; b < net.batch; ++b){
+ if(i+b == test.X.rows) break;
+ for(j = 0; j < k; ++j){
+ pred.vals[i+b][j] = out[j+b*k];
+ }
}
}
+ free(X);
return pred;
}
--
Gitblit v1.10.0