From 989ab8c38a02fa7ea9c25108151736c62e81c972 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 24 Apr 2015 17:27:50 +0000
Subject: [PATCH] IOU loss function
---
src/network.c | 640 ++++++++++++++++++++++++++-------------------------------
1 files changed, 290 insertions(+), 350 deletions(-)
diff --git a/src/network.c b/src/network.c
index 69942e8..3247a31 100644
--- a/src/network.c
+++ b/src/network.c
@@ -4,247 +4,121 @@
#include "image.h"
#include "data.h"
#include "utils.h"
+#include "params.h"
#include "crop_layer.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
+#include "deconvolutional_layer.h"
+#include "detection_layer.h"
#include "maxpool_layer.h"
#include "cost_layer.h"
#include "normalization_layer.h"
-#include "freeweight_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
-network make_network(int n, int batch)
+char *get_layer_string(LAYER_TYPE a)
+{
+ switch(a){
+ case CONVOLUTIONAL:
+ return "convolutional";
+ case DECONVOLUTIONAL:
+ return "deconvolutional";
+ case CONNECTED:
+ return "connected";
+ case MAXPOOL:
+ return "maxpool";
+ case SOFTMAX:
+ return "softmax";
+ case DETECTION:
+ return "detection";
+ case NORMALIZATION:
+ return "normalization";
+ case DROPOUT:
+ return "dropout";
+ case CROP:
+ return "crop";
+ case COST:
+ return "cost";
+ default:
+ break;
+ }
+ return "none";
+}
+
+network make_network(int n)
{
network net;
net.n = n;
- net.batch = batch;
net.layers = calloc(net.n, sizeof(void *));
net.types = calloc(net.n, sizeof(LAYER_TYPE));
net.outputs = 0;
net.output = 0;
+ net.seen = 0;
+ net.batch = 0;
+ net.inputs = 0;
+ net.h = net.w = net.c = 0;
#ifdef GPU
- net.input_cl = calloc(1, sizeof(cl_mem));
- net.truth_cl = calloc(1, sizeof(cl_mem));
+ net.input_gpu = calloc(1, sizeof(float *));
+ net.truth_gpu = calloc(1, sizeof(float *));
#endif
return net;
}
-#ifdef GPU
-
-void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
-{
- //printf("start\n");
- int i;
- for(i = 0; i < net.n; ++i){
- //clock_t time = clock();
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- forward_convolutional_layer_gpu(layer, input);
- input = layer.output_cl;
- }
- else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- forward_cost_layer_gpu(layer, input, truth);
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- forward_connected_layer_gpu(layer, input);
- input = layer.output_cl;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- forward_maxpool_layer_gpu(layer, input);
- input = layer.output_cl;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- forward_softmax_layer_gpu(layer, input);
- input = layer.output_cl;
- }
- //printf("%d %f\n", i, sec(clock()-time));
- /*
- else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_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;
- }
- */
- }
-}
-
-void backward_network_gpu(network net, cl_mem input)
-{
- int i;
- cl_mem prev_input;
- cl_mem prev_delta;
- for(i = net.n-1; i >= 0; --i){
- if(i == 0){
- prev_input = input;
- prev_delta = 0;
- }else{
- prev_input = get_network_output_cl_layer(net, i-1);
- prev_delta = get_network_delta_cl_layer(net, i-1);
- }
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- backward_convolutional_layer_gpu(layer, prev_delta);
- }
- else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- backward_cost_layer_gpu(layer, prev_input, prev_delta);
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- backward_connected_layer_gpu(layer, prev_input, prev_delta);
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- backward_maxpool_layer_gpu(layer, prev_delta);
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- backward_softmax_layer_gpu(layer, prev_delta);
- }
- }
-}
-
-void update_network_gpu(network net)
+void forward_network(network net, network_state state)
{
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_gpu(layer);
+ forward_convolutional_layer(*(convolutional_layer *)net.layers[i], state);
+ }
+ else if(net.types[i] == DECONVOLUTIONAL){
+ forward_deconvolutional_layer(*(deconvolutional_layer *)net.layers[i], state);
+ }
+ else if(net.types[i] == DETECTION){
+ forward_detection_layer(*(detection_layer *)net.layers[i], state);
}
else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer_gpu(layer);
- }
- }
-}
-
-cl_mem get_network_output_cl_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output_cl;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- return layer.output_cl;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output_cl;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.output_cl;
- }
- return 0;
-}
-
-cl_mem get_network_delta_cl_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.delta_cl;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- return layer.delta_cl;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.delta_cl;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.delta_cl;
- }
- return 0;
-}
-
-#endif
-
-void forward_network(network net, float *input, float *truth, int train)
-{
- int i;
- for(i = 0; i < net.n; ++i){
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- 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;
+ forward_connected_layer(*(connected_layer *)net.layers[i], state);
}
else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- forward_crop_layer(layer, input);
- input = layer.output;
+ forward_crop_layer(*(crop_layer *)net.layers[i], state);
}
else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- forward_cost_layer(layer, input, truth);
+ forward_cost_layer(*(cost_layer *)net.layers[i], state);
}
else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- forward_softmax_layer(layer, input);
- input = layer.output;
+ forward_softmax_layer(*(softmax_layer *)net.layers[i], state);
}
else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- forward_maxpool_layer(layer, input);
- input = layer.output;
+ forward_maxpool_layer(*(maxpool_layer *)net.layers[i], state);
}
else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- forward_normalization_layer(layer, input);
- input = layer.output;
+ forward_normalization_layer(*(normalization_layer *)net.layers[i], state);
}
else if(net.types[i] == DROPOUT){
- if(!train) continue;
- dropout_layer layer = *(dropout_layer *)net.layers[i];
- forward_dropout_layer(layer, input);
+ forward_dropout_layer(*(dropout_layer *)net.layers[i], state);
}
- else if(net.types[i] == FREEWEIGHT){
- if(!train) continue;
- freeweight_layer layer = *(freeweight_layer *)net.layers[i];
- forward_freeweight_layer(layer, input);
- }
+ state.input = get_network_output_layer(net, i);
}
}
void update_network(network net)
{
int i;
+ int update_batch = net.batch*net.subdivisions;
for(i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- update_convolutional_layer(layer);
+ update_convolutional_layer(layer, update_batch, net.learning_rate, net.momentum, net.decay);
}
- else if(net.types[i] == MAXPOOL){
- //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- }
- else if(net.types[i] == SOFTMAX){
- //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- }
- else if(net.types[i] == NORMALIZATION){
- //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ update_deconvolutional_layer(layer, net.learning_rate, net.momentum, net.decay);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- update_connected_layer(layer);
+ update_connected_layer(layer, update_batch, net.learning_rate, net.momentum, net.decay);
}
}
}
@@ -252,27 +126,27 @@
float *get_network_output_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.output;
+ return ((convolutional_layer *)net.layers[i]) -> output;
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ return ((deconvolutional_layer *)net.layers[i]) -> output;
} else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- return layer.output;
+ return ((maxpool_layer *)net.layers[i]) -> output;
+ } else if(net.types[i] == DETECTION){
+ return ((detection_layer *)net.layers[i]) -> output;
} else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- return layer.output;
+ return ((softmax_layer *)net.layers[i]) -> output;
} else if(net.types[i] == DROPOUT){
return get_network_output_layer(net, i-1);
- } else if(net.types[i] == FREEWEIGHT){
- 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;
+ return ((connected_layer *)net.layers[i]) -> output;
+ } else if(net.types[i] == CROP){
+ return ((crop_layer *)net.layers[i]) -> output;
} else if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- return layer.output;
+ return ((normalization_layer *)net.layers[i]) -> output;
}
return 0;
}
+
float *get_network_output(network net)
{
int i;
@@ -285,15 +159,20 @@
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
return layer.delta;
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ return layer.delta;
} else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.delta;
} else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.delta;
+ } else if(net.types[i] == DETECTION){
+ detection_layer layer = *(detection_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] == FREEWEIGHT){
+ if(i == 0) return 0;
return get_network_delta_layer(net, i-1);
} else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
@@ -307,6 +186,9 @@
if(net.types[net.n-1] == COST){
return ((cost_layer *)net.layers[net.n-1])->output[0];
}
+ if(net.types[net.n-1] == DETECTION){
+ return ((detection_layer *)net.layers[net.n-1])->cost[0];
+ }
return 0;
}
@@ -315,24 +197,6 @@
return get_network_delta_layer(net, net.n-1);
}
-float calculate_error_network(network net, float *truth)
-{
- float sum = 0;
- float *delta = get_network_delta(net);
- float *out = get_network_output(net);
- 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");
- return sum;
-}
-
int get_predicted_class_network(network net)
{
float *out = get_network_output(net);
@@ -340,121 +204,70 @@
return max_index(out, k);
}
-void backward_network(network net, float *input)
+void backward_network(network net, network_state state)
{
int i;
- float *prev_input;
- float *prev_delta;
+ float *original_input = state.input;
for(i = net.n-1; i >= 0; --i){
if(i == 0){
- prev_input = input;
- prev_delta = 0;
+ state.input = original_input;
+ state.delta = 0;
}else{
- prev_input = get_network_output_layer(net, i-1);
- prev_delta = get_network_delta_layer(net, i-1);
+ state.input = get_network_output_layer(net, i-1);
+ state.delta = get_network_delta_layer(net, i-1);
}
+
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- backward_convolutional_layer(layer, prev_delta);
+ backward_convolutional_layer(layer, state);
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ backward_deconvolutional_layer(layer, state);
}
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- if(i != 0) backward_maxpool_layer(layer, prev_delta);
+ if(i != 0) backward_maxpool_layer(layer, state);
+ }
+ else if(net.types[i] == DROPOUT){
+ dropout_layer layer = *(dropout_layer *)net.layers[i];
+ backward_dropout_layer(layer, state);
+ }
+ else if(net.types[i] == DETECTION){
+ detection_layer layer = *(detection_layer *)net.layers[i];
+ backward_detection_layer(layer, state);
}
else if(net.types[i] == NORMALIZATION){
normalization_layer layer = *(normalization_layer *)net.layers[i];
- if(i != 0) backward_normalization_layer(layer, prev_input, prev_delta);
+ if(i != 0) backward_normalization_layer(layer, state);
}
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
- if(i != 0) backward_softmax_layer(layer, prev_delta);
+ if(i != 0) backward_softmax_layer(layer, state);
}
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
- backward_connected_layer(layer, prev_input, prev_delta);
+ backward_connected_layer(layer, state);
}
else if(net.types[i] == COST){
cost_layer layer = *(cost_layer *)net.layers[i];
- backward_cost_layer(layer, prev_input, prev_delta);
+ backward_cost_layer(layer, state);
}
}
}
-
-#ifdef GPU
-float train_network_datum_gpu(network net, float *x, float *y)
-{
- int x_size = get_network_input_size(net)*net.batch;
- int y_size = get_network_output_size(net)*net.batch;
- clock_t time = clock();
- if(!*net.input_cl){
- *net.input_cl = cl_make_array(x, x_size);
- *net.truth_cl = cl_make_array(y, y_size);
- }else{
- cl_write_array(*net.input_cl, x, x_size);
- cl_write_array(*net.truth_cl, y, y_size);
- }
- //printf("trans %f\n", sec(clock()-time));
- time = clock();
- forward_network_gpu(net, *net.input_cl, *net.truth_cl, 1);
- //printf("forw %f\n", sec(clock()-time));
- time = clock();
- backward_network_gpu(net, *net.input_cl);
- //printf("back %f\n", sec(clock()-time));
- time = clock();
- float error = get_network_cost(net);
- update_network_gpu(net);
- //printf("updt %f\n", sec(clock()-time));
- time = clock();
- return error;
-}
-
-float train_network_sgd_gpu(network net, data d, int n)
-{
- int batch = net.batch;
- float *X = calloc(batch*d.X.cols, sizeof(float));
- float *y = calloc(batch*d.y.cols, sizeof(float));
-
- int i;
- float sum = 0;
- for(i = 0; i < n; ++i){
- get_random_batch(d, batch, X, y);
- float err = train_network_datum_gpu(net, X, y);
- sum += err;
- }
- free(X);
- free(y);
- return (float)sum/(n*batch);
-}
-
-float train_network_data_gpu(network net, data d, int n)
-{
- int batch = net.batch;
- float *X = calloc(batch*d.X.cols, sizeof(float));
- float *y = calloc(batch*d.y.cols, sizeof(float));
-
- int i;
- float sum = 0;
- for(i = 0; i < n; ++i){
- get_next_batch(d, batch, i*batch, X, y);
- float err = train_network_datum_gpu(net, X, y);
- sum += err;
- }
- free(X);
- free(y);
- return (float)sum/(n*batch);
-}
-#endif
-
-
float train_network_datum(network net, float *x, float *y)
{
- forward_network(net, x, y, 1);
- //int class = get_predicted_class_network(net);
- backward_network(net, x);
+ #ifdef GPU
+ if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
+ #endif
+ network_state state;
+ state.input = x;
+ state.truth = y;
+ state.train = 1;
+ forward_network(net, state);
+ backward_network(net, state);
float error = get_network_cost(net);
- update_network(net);
- //return (y[class]?1:0);
+ if((net.seen/net.batch)%net.subdivisions == 0) update_network(net);
return error;
}
@@ -467,6 +280,7 @@
int i;
float sum = 0;
for(i = 0; i < n; ++i){
+ net.seen += batch;
get_random_batch(d, batch, X, y);
float err = train_network_datum(net, X, y);
sum += err;
@@ -475,18 +289,41 @@
free(y);
return (float)sum/(n*batch);
}
+
+float train_network(network net, data d)
+{
+ int batch = net.batch;
+ int n = d.X.rows / batch;
+ float *X = calloc(batch*d.X.cols, sizeof(float));
+ float *y = calloc(batch*d.y.cols, sizeof(float));
+
+ int i;
+ float sum = 0;
+ for(i = 0; i < n; ++i){
+ get_next_batch(d, batch, i*batch, X, y);
+ net.seen += batch;
+ float err = train_network_datum(net, X, y);
+ sum += err;
+ }
+ free(X);
+ free(y);
+ return (float)sum/(n*batch);
+}
+
float train_network_batch(network net, data d, int n)
{
int i,j;
+ network_state state;
+ state.train = 1;
float sum = 0;
int batch = 2;
for(i = 0; i < n; ++i){
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, y, 1);
- backward_network(net, x);
+ state.input = d.X.vals[index];
+ state.truth = d.y.vals[index];
+ forward_network(net, state);
+ backward_network(net, state);
sum += get_network_cost(net);
}
update_network(net);
@@ -494,29 +331,58 @@
return (float)sum/(n*batch);
}
-
-void train_network(network net, data d)
+void set_batch_network(network *net, int b)
{
+ net->batch = b;
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]);
- if(i%100 == 0){
- visualize_network(net);
- cvWaitKey(10);
+ for(i = 0; i < net->n; ++i){
+ if(net->types[i] == CONVOLUTIONAL){
+ convolutional_layer *layer = (convolutional_layer *)net->layers[i];
+ layer->batch = b;
+ }else if(net->types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer *layer = (deconvolutional_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == MAXPOOL){
+ maxpool_layer *layer = (maxpool_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == CONNECTED){
+ connected_layer *layer = (connected_layer *)net->layers[i];
+ layer->batch = b;
+ } else if(net->types[i] == DROPOUT){
+ dropout_layer *layer = (dropout_layer *) net->layers[i];
+ layer->batch = b;
+ } else if(net->types[i] == DETECTION){
+ detection_layer *layer = (detection_layer *) net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == SOFTMAX){
+ softmax_layer *layer = (softmax_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == COST){
+ cost_layer *layer = (cost_layer *)net->layers[i];
+ layer->batch = b;
+ }
+ else if(net->types[i] == CROP){
+ crop_layer *layer = (crop_layer *)net->layers[i];
+ layer->batch = b;
}
}
- visualize_network(net);
- cvWaitKey(100);
- fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}
+
int get_network_input_size_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
return layer.h*layer.w*layer.c;
}
+ if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ return layer.h*layer.w*layer.c;
+ }
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return layer.h*layer.w*layer.c;
@@ -527,15 +393,18 @@
} else if(net.types[i] == DROPOUT){
dropout_layer layer = *(dropout_layer *) net.layers[i];
return layer.inputs;
- }
- else if(net.types[i] == FREEWEIGHT){
- freeweight_layer layer = *(freeweight_layer *) net.layers[i];
+ } else if(net.types[i] == DETECTION){
+ detection_layer layer = *(detection_layer *) net.layers[i];
return layer.inputs;
+ } else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *) net.layers[i];
+ return layer.c*layer.h*layer.w;
}
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.inputs;
}
+ fprintf(stderr, "Can't find input size\n");
return 0;
}
@@ -546,11 +415,24 @@
image output = get_convolutional_image(layer);
return output.h*output.w*output.c;
}
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ image output = get_deconvolutional_image(layer);
+ return output.h*output.w*output.c;
+ }
+ else if(net.types[i] == DETECTION){
+ detection_layer layer = *(detection_layer *)net.layers[i];
+ return get_detection_layer_output_size(layer);
+ }
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
image output = get_maxpool_image(layer);
return output.h*output.w*output.c;
}
+ else if(net.types[i] == CROP){
+ crop_layer layer = *(crop_layer *) net.layers[i];
+ return layer.c*layer.crop_height*layer.crop_width;
+ }
else if(net.types[i] == CONNECTED){
connected_layer layer = *(connected_layer *)net.layers[i];
return layer.outputs;
@@ -559,14 +441,11 @@
dropout_layer layer = *(dropout_layer *) net.layers[i];
return layer.inputs;
}
- else if(net.types[i] == FREEWEIGHT){
- freeweight_layer layer = *(freeweight_layer *) net.layers[i];
- return layer.inputs;
- }
else if(net.types[i] == SOFTMAX){
softmax_layer layer = *(softmax_layer *)net.layers[i];
return layer.inputs;
}
+ fprintf(stderr, "Can't find output size\n");
return 0;
}
@@ -576,21 +455,31 @@
for (i = 0; i < net.n; ++i){
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer *layer = (convolutional_layer *)net.layers[i];
- resize_convolutional_layer(layer, h, w, c);
+ resize_convolutional_layer(layer, h, w);
image output = get_convolutional_image(*layer);
h = output.h;
w = output.w;
c = output.c;
+ } else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer *layer = (deconvolutional_layer *)net.layers[i];
+ resize_deconvolutional_layer(layer, h, w);
+ image output = get_deconvolutional_image(*layer);
+ h = output.h;
+ w = output.w;
+ c = output.c;
}else if(net.types[i] == MAXPOOL){
maxpool_layer *layer = (maxpool_layer *)net.layers[i];
- resize_maxpool_layer(layer, h, w, c);
+ resize_maxpool_layer(layer, h, w);
image output = get_maxpool_image(*layer);
h = output.h;
w = output.w;
c = output.c;
+ }else if(net.types[i] == DROPOUT){
+ dropout_layer *layer = (dropout_layer *)net.layers[i];
+ resize_dropout_layer(layer, h*w*c);
}else if(net.types[i] == NORMALIZATION){
normalization_layer *layer = (normalization_layer *)net.layers[i];
- resize_normalization_layer(layer, h, w, c);
+ resize_normalization_layer(layer, h, w);
image output = get_normalization_image(*layer);
h = output.h;
w = output.w;
@@ -614,12 +503,28 @@
return get_network_input_size_layer(net, 0);
}
+detection_layer *get_network_detection_layer(network net)
+{
+ int i;
+ for(i = 0; i < net.n; ++i){
+ if(net.types[i] == DETECTION){
+ detection_layer *layer = (detection_layer *)net.layers[i];
+ return layer;
+ }
+ }
+ return 0;
+}
+
image get_network_image_layer(network net, int i)
{
if(net.types[i] == CONVOLUTIONAL){
convolutional_layer layer = *(convolutional_layer *)net.layers[i];
return get_convolutional_image(layer);
}
+ else if(net.types[i] == DECONVOLUTIONAL){
+ deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
+ return get_deconvolutional_image(layer);
+ }
else if(net.types[i] == MAXPOOL){
maxpool_layer layer = *(maxpool_layer *)net.layers[i];
return get_maxpool_image(layer);
@@ -628,6 +533,9 @@
normalization_layer layer = *(normalization_layer *)net.layers[i];
return get_normalization_image(layer);
}
+ else if(net.types[i] == DROPOUT){
+ return get_network_image_layer(net, i-1);
+ }
else if(net.types[i] == CROP){
crop_layer layer = *(crop_layer *)net.layers[i];
return get_crop_image(layer);
@@ -664,30 +572,26 @@
}
}
-void top_predictions(network net, int n, int *index)
+void top_predictions(network net, int k, int *index)
{
- int i,j;
- int k = get_network_output_size(net);
+ int size = get_network_output_size(net);
float *out = get_network_output(net);
- float thresh = FLT_MAX;
- for(i = 0; i < n; ++i){
- float max = -FLT_MAX;
- int max_i = -1;
- for(j = 0; j < k; ++j){
- float val = out[j];
- if(val > max && val < thresh){
- max = val;
- max_i = j;
- }
- }
- index[i] = max_i;
- thresh = max;
- }
+ top_k(out, size, k, index);
}
+
float *network_predict(network net, float *input)
{
- forward_network(net, input, 0, 0);
+#ifdef GPU
+ if(gpu_index >= 0) return network_predict_gpu(net, input);
+#endif
+
+ network_state state;
+ state.input = input;
+ state.truth = 0;
+ state.train = 0;
+ state.delta = 0;
+ forward_network(net, state);
float *out = get_network_output(net);
return out;
}
@@ -722,7 +626,7 @@
int i,j,b;
int k = get_network_output_size(net);
matrix pred = make_matrix(test.X.rows, k);
- float *X = calloc(net.batch*test.X.rows, sizeof(float));
+ float *X = calloc(net.batch*test.X.cols, 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;
@@ -784,18 +688,54 @@
}
}
+void compare_networks(network n1, network n2, data test)
+{
+ matrix g1 = network_predict_data(n1, test);
+ matrix g2 = network_predict_data(n2, test);
+ int i;
+ int a,b,c,d;
+ a = b = c = d = 0;
+ for(i = 0; i < g1.rows; ++i){
+ int truth = max_index(test.y.vals[i], test.y.cols);
+ int p1 = max_index(g1.vals[i], g1.cols);
+ int p2 = max_index(g2.vals[i], g2.cols);
+ if(p1 == truth){
+ if(p2 == truth) ++d;
+ else ++c;
+ }else{
+ if(p2 == truth) ++b;
+ else ++a;
+ }
+ }
+ printf("%5d %5d\n%5d %5d\n", a, b, c, d);
+ float num = pow((abs(b - c) - 1.), 2.);
+ float den = b + c;
+ printf("%f\n", num/den);
+}
+
float network_accuracy(network net, data d)
{
matrix guess = network_predict_data(net, d);
- float acc = matrix_accuracy(d.y, guess);
+ float acc = matrix_topk_accuracy(d.y, guess,1);
free_matrix(guess);
return acc;
}
+float *network_accuracies(network net, data d)
+{
+ static float acc[2];
+ matrix guess = network_predict_data(net, d);
+ acc[0] = matrix_topk_accuracy(d.y, guess,1);
+ acc[1] = matrix_topk_accuracy(d.y, guess,5);
+ free_matrix(guess);
+ return acc;
+}
+
+
float network_accuracy_multi(network net, data d, int n)
{
matrix guess = network_predict_data_multi(net, d, n);
- float acc = matrix_accuracy(d.y, guess);
+ float acc = matrix_topk_accuracy(d.y, guess,1);
free_matrix(guess);
return acc;
}
--
Gitblit v1.10.0