From f0abcfa02b2094396f955c743f7f11fcdb2e3d13 Mon Sep 17 00:00:00 2001
From: IlyaOvodov <b@ovdv.ru>
Date: Mon, 04 Jun 2018 15:57:15 +0000
Subject: [PATCH] Merge branch 'master' of https://github.com/AlexeyAB/darknet into Fix_get_color_depth
---
src/network.c | 856 +++++++++++++++++++++++++++++++-------------------------
1 files changed, 475 insertions(+), 381 deletions(-)
diff --git a/src/network.c b/src/network.c
index 3247a31..050d334 100644
--- a/src/network.c
+++ b/src/network.c
@@ -1,45 +1,171 @@
#include <stdio.h>
#include <time.h>
+#include <assert.h>
#include "network.h"
#include "image.h"
#include "data.h"
#include "utils.h"
-#include "params.h"
+#include "blas.h"
#include "crop_layer.h"
#include "connected_layer.h"
+#include "gru_layer.h"
+#include "rnn_layer.h"
+#include "crnn_layer.h"
+#include "local_layer.h"
#include "convolutional_layer.h"
-#include "deconvolutional_layer.h"
+#include "activation_layer.h"
#include "detection_layer.h"
-#include "maxpool_layer.h"
-#include "cost_layer.h"
+#include "region_layer.h"
#include "normalization_layer.h"
+#include "batchnorm_layer.h"
+#include "maxpool_layer.h"
+#include "reorg_layer.h"
+#include "avgpool_layer.h"
+#include "cost_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
+#include "route_layer.h"
+#include "shortcut_layer.h"
+#include "yolo_layer.h"
+#include "upsample_layer.h"
+#include "parser.h"
+
+network *load_network_custom(char *cfg, char *weights, int clear, int batch)
+{
+ printf(" Try to load cfg: %s, weights: %s, clear = %d \n", cfg, weights, clear);
+ network *net = calloc(1, sizeof(network));
+ *net = parse_network_cfg_custom(cfg, batch);
+ if (weights && weights[0] != 0) {
+ load_weights(net, weights);
+ }
+ if (clear) (*net->seen) = 0;
+ return net;
+}
+
+network *load_network(char *cfg, char *weights, int clear)
+{
+ return load_network_custom(cfg, weights, clear, 0);
+}
+
+int get_current_batch(network net)
+{
+ int batch_num = (*net.seen)/(net.batch*net.subdivisions);
+ return batch_num;
+}
+
+void reset_momentum(network net)
+{
+ if (net.momentum == 0) return;
+ net.learning_rate = 0;
+ net.momentum = 0;
+ net.decay = 0;
+ #ifdef GPU
+ //if(net.gpu_index >= 0) update_network_gpu(net);
+ #endif
+}
+
+void reset_network_state(network *net, int b)
+{
+ int i;
+ for (i = 0; i < net->n; ++i) {
+#ifdef GPU
+ layer l = net->layers[i];
+ if (l.state_gpu) {
+ fill_ongpu(l.outputs, 0, l.state_gpu + l.outputs*b, 1);
+ }
+ if (l.h_gpu) {
+ fill_ongpu(l.outputs, 0, l.h_gpu + l.outputs*b, 1);
+ }
+#endif
+ }
+}
+
+void reset_rnn(network *net)
+{
+ reset_network_state(net, 0);
+}
+
+float get_current_rate(network net)
+{
+ int batch_num = get_current_batch(net);
+ int i;
+ float rate;
+ if (batch_num < net.burn_in) return net.learning_rate * pow((float)batch_num / net.burn_in, net.power);
+ switch (net.policy) {
+ case CONSTANT:
+ return net.learning_rate;
+ case STEP:
+ return net.learning_rate * pow(net.scale, batch_num/net.step);
+ case STEPS:
+ rate = net.learning_rate;
+ for(i = 0; i < net.num_steps; ++i){
+ if(net.steps[i] > batch_num) return rate;
+ rate *= net.scales[i];
+ //if(net.steps[i] > batch_num - 1 && net.scales[i] > 1) reset_momentum(net);
+ }
+ return rate;
+ case EXP:
+ return net.learning_rate * pow(net.gamma, batch_num);
+ case POLY:
+ return net.learning_rate * pow(1 - (float)batch_num / net.max_batches, net.power);
+ //if (batch_num < net.burn_in) return net.learning_rate * pow((float)batch_num / net.burn_in, net.power);
+ //return net.learning_rate * pow(1 - (float)batch_num / net.max_batches, net.power);
+ case RANDOM:
+ return net.learning_rate * pow(rand_uniform(0,1), net.power);
+ case SIG:
+ return net.learning_rate * (1./(1.+exp(net.gamma*(batch_num - net.step))));
+ default:
+ fprintf(stderr, "Policy is weird!\n");
+ return net.learning_rate;
+ }
+}
char *get_layer_string(LAYER_TYPE a)
{
switch(a){
case CONVOLUTIONAL:
return "convolutional";
+ case ACTIVE:
+ return "activation";
+ case LOCAL:
+ return "local";
case DECONVOLUTIONAL:
return "deconvolutional";
case CONNECTED:
return "connected";
+ case RNN:
+ return "rnn";
+ case GRU:
+ return "gru";
+ case CRNN:
+ return "crnn";
case MAXPOOL:
return "maxpool";
+ case REORG:
+ return "reorg";
+ case AVGPOOL:
+ return "avgpool";
case SOFTMAX:
return "softmax";
case DETECTION:
return "detection";
- case NORMALIZATION:
- return "normalization";
+ case REGION:
+ return "region";
case DROPOUT:
return "dropout";
case CROP:
return "crop";
case COST:
return "cost";
+ case ROUTE:
+ return "route";
+ case SHORTCUT:
+ return "shortcut";
+ case NORMALIZATION:
+ return "normalization";
+ case BATCHNORM:
+ return "batchnorm";
default:
break;
}
@@ -48,58 +174,34 @@
network make_network(int n)
{
- network net;
+ network net = {0};
net.n = n;
- 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.layers = calloc(net.n, sizeof(layer));
+ net.seen = calloc(1, sizeof(int));
+#ifdef GPU
net.input_gpu = calloc(1, sizeof(float *));
net.truth_gpu = calloc(1, sizeof(float *));
- #endif
+
+ net.input16_gpu = calloc(1, sizeof(float *));
+ net.output16_gpu = calloc(1, sizeof(float *));
+ net.max_input16_size = calloc(1, sizeof(size_t));
+ net.max_output16_size = calloc(1, sizeof(size_t));
+#endif
return net;
}
void forward_network(network net, network_state state)
{
+ state.workspace = net.workspace;
int i;
for(i = 0; i < net.n; ++i){
- if(net.types[i] == CONVOLUTIONAL){
- forward_convolutional_layer(*(convolutional_layer *)net.layers[i], state);
+ state.index = i;
+ layer l = net.layers[i];
+ if(l.delta){
+ scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
}
- 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){
- forward_connected_layer(*(connected_layer *)net.layers[i], state);
- }
- else if(net.types[i] == CROP){
- forward_crop_layer(*(crop_layer *)net.layers[i], state);
- }
- else if(net.types[i] == COST){
- forward_cost_layer(*(cost_layer *)net.layers[i], state);
- }
- else if(net.types[i] == SOFTMAX){
- forward_softmax_layer(*(softmax_layer *)net.layers[i], state);
- }
- else if(net.types[i] == MAXPOOL){
- forward_maxpool_layer(*(maxpool_layer *)net.layers[i], state);
- }
- else if(net.types[i] == NORMALIZATION){
- forward_normalization_layer(*(normalization_layer *)net.layers[i], state);
- }
- else if(net.types[i] == DROPOUT){
- forward_dropout_layer(*(dropout_layer *)net.layers[i], state);
- }
- state.input = get_network_output_layer(net, i);
+ l.forward(l, state);
+ state.input = l.output;
}
}
@@ -107,94 +209,37 @@
{
int i;
int update_batch = net.batch*net.subdivisions;
+ float rate = get_current_rate(net);
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_batch, net.learning_rate, net.momentum, net.decay);
- }
- 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_batch, net.learning_rate, net.momentum, net.decay);
+ layer l = net.layers[i];
+ if(l.update){
+ l.update(l, update_batch, rate, net.momentum, net.decay);
}
}
}
-float *get_network_output_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- 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){
- 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){
- 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] == CONNECTED){
- 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){
- return ((normalization_layer *)net.layers[i]) -> output;
- }
- return 0;
-}
-
float *get_network_output(network net)
{
+#ifdef GPU
+ if (gpu_index >= 0) return get_network_output_gpu(net);
+#endif
int i;
- for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
- return get_network_output_layer(net, i);
-}
-
-float *get_network_delta_layer(network net, int i)
-{
- 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){
- 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];
- return layer.delta;
- }
- return 0;
+ for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
+ return net.layers[i].output;
}
float get_network_cost(network net)
{
- if(net.types[net.n-1] == COST){
- return ((cost_layer *)net.layers[net.n-1])->output[0];
+ int i;
+ float sum = 0;
+ int count = 0;
+ for(i = 0; i < net.n; ++i){
+ if(net.layers[i].cost){
+ sum += net.layers[i].cost[0];
+ ++count;
+ }
}
- if(net.types[net.n-1] == DETECTION){
- return ((detection_layer *)net.layers[net.n-1])->cost[0];
- }
- return 0;
-}
-
-float *get_network_delta(network net)
-{
- return get_network_delta_layer(net, net.n-1);
+ return sum/count;
}
int get_predicted_class_network(network net)
@@ -208,66 +253,41 @@
{
int i;
float *original_input = state.input;
+ float *original_delta = state.delta;
+ state.workspace = net.workspace;
for(i = net.n-1; i >= 0; --i){
+ state.index = i;
if(i == 0){
state.input = original_input;
- state.delta = 0;
+ state.delta = original_delta;
}else{
- state.input = get_network_output_layer(net, i-1);
- state.delta = get_network_delta_layer(net, i-1);
+ layer prev = net.layers[i-1];
+ state.input = prev.output;
+ state.delta = prev.delta;
}
-
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- 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, 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, state);
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- 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, state);
- }
- else if(net.types[i] == COST){
- cost_layer layer = *(cost_layer *)net.layers[i];
- backward_cost_layer(layer, state);
- }
+ layer l = net.layers[i];
+ if (l.stopbackward) break;
+ l.backward(l, state);
}
}
float train_network_datum(network net, float *x, float *y)
{
- #ifdef GPU
+#ifdef GPU
if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
- #endif
+#endif
network_state state;
+ *net.seen += net.batch;
+ state.index = 0;
+ state.net = net;
state.input = x;
+ state.delta = 0;
state.truth = y;
state.train = 1;
forward_network(net, state);
backward_network(net, state);
float error = get_network_cost(net);
- if((net.seen/net.batch)%net.subdivisions == 0) update_network(net);
+ if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);
return error;
}
@@ -280,7 +300,6 @@
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;
@@ -292,6 +311,7 @@
float train_network(network net, data d)
{
+ assert(d.X.rows % net.batch == 0);
int batch = net.batch;
int n = d.X.rows / batch;
float *X = calloc(batch*d.X.cols, sizeof(float));
@@ -301,7 +321,6 @@
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;
}
@@ -310,11 +329,15 @@
return (float)sum/(n*batch);
}
+
float train_network_batch(network net, data d, int n)
{
int i,j;
network_state state;
+ state.index = 0;
+ state.net = net;
state.train = 1;
+ state.delta = 0;
float sum = 0;
int batch = 2;
for(i = 0; i < n; ++i){
@@ -336,211 +359,138 @@
net->batch = b;
int i;
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;
+ net->layers[i].batch = b;
+#ifdef CUDNN
+ if(net->layers[i].type == CONVOLUTIONAL){
+ cudnn_convolutional_setup(net->layers + i, cudnn_fastest);
+ /*
+ layer *l = net->layers + i;
+ cudnn_convolutional_setup(l, cudnn_fastest);
+ // 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);
+ }
+ */
}
- 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;
- }
+#endif
}
}
-
-int get_network_input_size_layer(network net, int i)
+int resize_network(network *net, int w, int h)
{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- return layer.h*layer.w*layer.c;
+#ifdef GPU
+ cuda_set_device(net->gpu_index);
+ if(gpu_index >= 0){
+ cuda_free(net->workspace);
+ if (net->input_gpu) {
+ cuda_free(*net->input_gpu);
+ *net->input_gpu = 0;
+ cuda_free(*net->truth_gpu);
+ *net->truth_gpu = 0;
+ }
}
- 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;
- }
- 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] == 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;
-}
-
-int get_network_output_size_layer(network net, int i)
-{
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- 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;
- }
- 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];
- return layer.inputs;
- }
- fprintf(stderr, "Can't find output size\n");
- return 0;
-}
-
-int resize_network(network net, int h, int w, int c)
-{
+#endif
int i;
- 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);
- 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);
- 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);
- image output = get_normalization_image(*layer);
- h = output.h;
- w = output.w;
- c = output.c;
+ //if(w == net->w && h == net->h) return 0;
+ net->w = w;
+ net->h = h;
+ int inputs = 0;
+ size_t workspace_size = 0;
+ //fprintf(stderr, "Resizing to %d x %d...\n", w, h);
+ //fflush(stderr);
+ for (i = 0; i < net->n; ++i){
+ layer l = net->layers[i];
+ //printf(" %d: layer = %d,", i, l.type);
+ if(l.type == CONVOLUTIONAL){
+ resize_convolutional_layer(&l, w, h);
+ }else if(l.type == CROP){
+ resize_crop_layer(&l, w, h);
+ }else if(l.type == MAXPOOL){
+ resize_maxpool_layer(&l, w, h);
+ }else if(l.type == REGION){
+ resize_region_layer(&l, w, h);
+ }else if (l.type == YOLO) {
+ resize_yolo_layer(&l, w, h);
+ }else if(l.type == ROUTE){
+ resize_route_layer(&l, net);
+ }else if (l.type == SHORTCUT) {
+ resize_shortcut_layer(&l, w, h);
+ }else if (l.type == UPSAMPLE) {
+ resize_upsample_layer(&l, w, h);
+ }else if(l.type == REORG){
+ resize_reorg_layer(&l, w, h);
+ }else if(l.type == AVGPOOL){
+ resize_avgpool_layer(&l, w, h);
+ }else if(l.type == NORMALIZATION){
+ resize_normalization_layer(&l, w, h);
+ }else if(l.type == COST){
+ resize_cost_layer(&l, inputs);
}else{
+ fprintf(stderr, "Resizing type %d \n", (int)l.type);
error("Cannot resize this type of layer");
}
+ if(l.workspace_size > workspace_size) workspace_size = l.workspace_size;
+ inputs = l.outputs;
+ net->layers[i] = l;
+ w = l.out_w;
+ h = l.out_h;
+ if(l.type == AVGPOOL) break;
}
+#ifdef GPU
+ if(gpu_index >= 0){
+ printf(" try to allocate workspace = %zu * sizeof(float), ", (workspace_size - 1) / sizeof(float) + 1);
+ net->workspace = cuda_make_array(0, (workspace_size-1)/sizeof(float)+1);
+ printf(" CUDA allocate done! \n");
+ }else {
+ free(net->workspace);
+ net->workspace = calloc(1, workspace_size);
+ }
+#else
+ free(net->workspace);
+ net->workspace = calloc(1, workspace_size);
+#endif
+ //fprintf(stderr, " Done!\n");
return 0;
}
int get_network_output_size(network net)
{
int i;
- for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
- return get_network_output_size_layer(net, i);
+ for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
+ return net.layers[i].outputs;
}
int get_network_input_size(network net)
{
- return get_network_input_size_layer(net, 0);
+ return net.layers[0].inputs;
}
-detection_layer *get_network_detection_layer(network net)
+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;
+ if(net.layers[i].type == DETECTION){
+ return net.layers[i];
}
}
- return 0;
+ fprintf(stderr, "Detection layer not found!!\n");
+ detection_layer l = {0};
+ return l;
}
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);
+ layer l = net.layers[i];
+ if (l.out_w && l.out_h && l.out_c){
+ return float_to_image(l.out_w, l.out_h, l.out_c, l.output);
}
- 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);
- }
- else if(net.types[i] == NORMALIZATION){
- 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);
- }
- return make_empty_image(0,0,0);
+ image def = {0};
+ return def;
}
image get_network_image(network net)
@@ -550,7 +500,8 @@
image m = get_network_image_layer(net, i);
if(m.h != 0) return m;
}
- return make_empty_image(0,0,0);
+ image def = {0};
+ return def;
}
void visualize_network(network net)
@@ -558,16 +509,11 @@
image *prev = 0;
int i;
char buff[256];
- //show_image(get_network_image_layer(net, 0), "Crop");
for(i = 0; i < net.n; ++i){
sprintf(buff, "Layer %d", i);
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- prev = visualize_convolutional_layer(layer, buff, prev);
- }
- if(net.types[i] == NORMALIZATION){
- normalization_layer layer = *(normalization_layer *)net.layers[i];
- visualize_normalization_layer(layer, buff);
+ layer l = net.layers[i];
+ if(l.type == CONVOLUTIONAL){
+ prev = visualize_convolutional_layer(l, buff, prev);
}
}
}
@@ -587,6 +533,8 @@
#endif
network_state state;
+ state.net = net;
+ state.index = 0;
state.input = input;
state.truth = 0;
state.train = 0;
@@ -596,6 +544,112 @@
return out;
}
+int num_detections(network *net, float thresh)
+{
+ int i;
+ int s = 0;
+ for (i = 0; i < net->n; ++i) {
+ layer l = net->layers[i];
+ if (l.type == YOLO) {
+ s += yolo_num_detections(l, thresh);
+ }
+ if (l.type == DETECTION || l.type == REGION) {
+ s += l.w*l.h*l.n;
+ }
+ }
+ return s;
+}
+
+detection *make_network_boxes(network *net, float thresh, int *num)
+{
+ layer l = net->layers[net->n - 1];
+ int i;
+ int nboxes = num_detections(net, thresh);
+ if (num) *num = nboxes;
+ detection *dets = calloc(nboxes, sizeof(detection));
+ for (i = 0; i < nboxes; ++i) {
+ dets[i].prob = calloc(l.classes, sizeof(float));
+ if (l.coords > 4) {
+ dets[i].mask = calloc(l.coords - 4, sizeof(float));
+ }
+ }
+ return dets;
+}
+
+
+void custom_get_region_detections(layer l, int w, int h, int net_w, int net_h, float thresh, int *map, float hier, int relative, detection *dets, int letter)
+{
+ box *boxes = calloc(l.w*l.h*l.n, sizeof(box));
+ float **probs = calloc(l.w*l.h*l.n, sizeof(float *));
+ int i, j;
+ for (j = 0; j < l.w*l.h*l.n; ++j) probs[j] = calloc(l.classes, sizeof(float));
+ get_region_boxes(l, 1, 1, thresh, probs, boxes, 0, map);
+ for (j = 0; j < l.w*l.h*l.n; ++j) {
+ dets[j].classes = l.classes;
+ dets[j].bbox = boxes[j];
+ dets[j].objectness = 1;
+ for (i = 0; i < l.classes; ++i) {
+ dets[j].prob[i] = probs[j][i];
+ }
+ }
+
+ free(boxes);
+ free_ptrs((void **)probs, l.w*l.h*l.n);
+
+ //correct_region_boxes(dets, l.w*l.h*l.n, w, h, net_w, net_h, relative);
+ correct_yolo_boxes(dets, l.w*l.h*l.n, w, h, net_w, net_h, relative, letter);
+}
+
+void fill_network_boxes(network *net, int w, int h, float thresh, float hier, int *map, int relative, detection *dets, int letter)
+{
+ int j;
+ for (j = 0; j < net->n; ++j) {
+ layer l = net->layers[j];
+ if (l.type == YOLO) {
+ int count = get_yolo_detections(l, w, h, net->w, net->h, thresh, map, relative, dets, letter);
+ dets += count;
+ }
+ if (l.type == REGION) {
+ custom_get_region_detections(l, w, h, net->w, net->h, thresh, map, hier, relative, dets, letter);
+ //get_region_detections(l, w, h, net->w, net->h, thresh, map, hier, relative, dets);
+ dets += l.w*l.h*l.n;
+ }
+ if (l.type == DETECTION) {
+ get_detection_detections(l, w, h, thresh, dets);
+ dets += l.w*l.h*l.n;
+ }
+ }
+}
+
+detection *get_network_boxes(network *net, int w, int h, float thresh, float hier, int *map, int relative, int *num, int letter)
+{
+ detection *dets = make_network_boxes(net, thresh, num);
+ fill_network_boxes(net, w, h, thresh, hier, map, relative, dets, letter);
+ return dets;
+}
+
+void free_detections(detection *dets, int n)
+{
+ int i;
+ for (i = 0; i < n; ++i) {
+ free(dets[i].prob);
+ if (dets[i].mask) free(dets[i].mask);
+ }
+ free(dets);
+}
+
+float *network_predict_image(network *net, image im)
+{
+ image imr = letterbox_image(im, net->w, net->h);
+ set_batch_network(net, 1);
+ float *p = network_predict(*net, imr.data);
+ free_image(imr);
+ return p;
+}
+
+int network_width(network *net) { return net->w; }
+int network_height(network *net) { return net->h; }
+
matrix network_predict_data_multi(network net, data test, int n)
{
int i,j,b,m;
@@ -648,36 +702,9 @@
{
int i,j;
for(i = 0; i < net.n; ++i){
- float *output = 0;
- int n = 0;
- if(net.types[i] == CONVOLUTIONAL){
- convolutional_layer layer = *(convolutional_layer *)net.layers[i];
- output = layer.output;
- image m = get_convolutional_image(layer);
- n = m.h*m.w*m.c;
- }
- else if(net.types[i] == MAXPOOL){
- maxpool_layer layer = *(maxpool_layer *)net.layers[i];
- output = layer.output;
- image m = get_maxpool_image(layer);
- n = m.h*m.w*m.c;
- }
- else if(net.types[i] == CROP){
- crop_layer layer = *(crop_layer *)net.layers[i];
- output = layer.output;
- image m = get_crop_image(layer);
- n = m.h*m.w*m.c;
- }
- else if(net.types[i] == CONNECTED){
- connected_layer layer = *(connected_layer *)net.layers[i];
- output = layer.output;
- n = layer.outputs;
- }
- else if(net.types[i] == SOFTMAX){
- softmax_layer layer = *(softmax_layer *)net.layers[i];
- output = layer.output;
- n = layer.inputs;
- }
+ layer l = net.layers[i];
+ float *output = l.output;
+ int n = l.outputs;
float mean = mean_array(output, n);
float vari = variance_array(output, n);
fprintf(stderr, "Layer %d - Mean: %f, Variance: %f\n",i,mean, vari);
@@ -721,17 +748,16 @@
return acc;
}
-float *network_accuracies(network net, data d)
+float *network_accuracies(network net, data d, int n)
{
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);
+ acc[0] = matrix_topk_accuracy(d.y, guess, 1);
+ acc[1] = matrix_topk_accuracy(d.y, guess, n);
free_matrix(guess);
return acc;
}
-
float network_accuracy_multi(network net, data d, int n)
{
matrix guess = network_predict_data_multi(net, d, n);
@@ -740,4 +766,72 @@
return acc;
}
+void free_network(network net)
+{
+ int i;
+ for (i = 0; i < net.n; ++i) {
+ free_layer(net.layers[i]);
+ }
+ free(net.layers);
+ free(net.scales);
+ free(net.steps);
+ free(net.seen);
+
+#ifdef GPU
+ if (gpu_index >= 0) cuda_free(net.workspace);
+ else free(net.workspace);
+ if (*net.input_gpu) cuda_free(*net.input_gpu);
+ if (*net.truth_gpu) cuda_free(*net.truth_gpu);
+ if (net.input_gpu) free(net.input_gpu);
+ if (net.truth_gpu) free(net.truth_gpu);
+
+ if (*net.input16_gpu) cuda_free(*net.input16_gpu);
+ if (*net.output16_gpu) cuda_free(*net.output16_gpu);
+ if (net.input16_gpu) free(net.input16_gpu);
+ if (net.output16_gpu) free(net.output16_gpu);
+ if (net.max_input16_size) free(net.max_input16_size);
+ if (net.max_output16_size) free(net.max_output16_size);
+#else
+ free(net.workspace);
+#endif
+}
+
+
+void fuse_conv_batchnorm(network net)
+{
+ int j;
+ for (j = 0; j < net.n; ++j) {
+ layer *l = &net.layers[j];
+
+ if (l->type == CONVOLUTIONAL) {
+ //printf(" Merges Convolutional-%d and batch_norm \n", j);
+
+ if (l->batch_normalize) {
+ int f;
+ for (f = 0; f < l->n; ++f)
+ {
+ l->biases[f] = l->biases[f] - (double)l->scales[f] * l->rolling_mean[f] / (sqrt((double)l->rolling_variance[f]) + .000001f);
+
+ const size_t filter_size = l->size*l->size*l->c;
+ int i;
+ for (i = 0; i < filter_size; ++i) {
+ int w_index = f*filter_size + i;
+
+ l->weights[w_index] = (double)l->weights[w_index] * l->scales[f] / (sqrt((double)l->rolling_variance[f]) + .000001f);
+ }
+ }
+
+ l->batch_normalize = 0;
+#ifdef GPU
+ if (gpu_index >= 0) {
+ push_convolutional_layer(*l);
+ }
+#endif
+ }
+ }
+ else {
+ //printf(" Fusion skip layer type: %d \n", l->type);
+ }
+ }
+}
--
Gitblit v1.10.0