From 741ada451cc7fee1b9a4c3deaec6af87a2af7497 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Mon, 17 Aug 2015 16:31:25 +0000
Subject: [PATCH] Added Darknet reference model
---
src/parser.c | 567 ++++++++++++++++++++++++++++++++++++++++++++++++++------
1 files changed, 505 insertions(+), 62 deletions(-)
diff --git a/src/parser.c b/src/parser.c
index 7541620..b373c01 100644
--- a/src/parser.c
+++ b/src/parser.c
@@ -4,9 +4,18 @@
#include "parser.h"
#include "activations.h"
+#include "crop_layer.h"
+#include "cost_layer.h"
#include "convolutional_layer.h"
+#include "normalization_layer.h"
+#include "deconvolutional_layer.h"
#include "connected_layer.h"
#include "maxpool_layer.h"
+#include "softmax_layer.h"
+#include "dropout_layer.h"
+#include "detection_layer.h"
+#include "avgpool_layer.h"
+#include "route_layer.h"
#include "list.h"
#include "option_list.h"
#include "utils.h"
@@ -16,92 +25,395 @@
list *options;
}section;
+int is_network(section *s);
int is_convolutional(section *s);
+int is_deconvolutional(section *s);
int is_connected(section *s);
int is_maxpool(section *s);
+int is_avgpool(section *s);
+int is_dropout(section *s);
+int is_softmax(section *s);
+int is_normalization(section *s);
+int is_crop(section *s);
+int is_cost(section *s);
+int is_detection(section *s);
+int is_route(section *s);
list *read_cfg(char *filename);
+void free_section(section *s)
+{
+ free(s->type);
+ node *n = s->options->front;
+ while(n){
+ kvp *pair = (kvp *)n->val;
+ free(pair->key);
+ free(pair);
+ node *next = n->next;
+ free(n);
+ n = next;
+ }
+ free(s->options);
+ free(s);
+}
+
+void parse_data(char *data, float *a, int n)
+{
+ int i;
+ if(!data) return;
+ char *curr = data;
+ char *next = data;
+ int done = 0;
+ for(i = 0; i < n && !done; ++i){
+ while(*++next !='\0' && *next != ',');
+ if(*next == '\0') done = 1;
+ *next = '\0';
+ sscanf(curr, "%g", &a[i]);
+ curr = next+1;
+ }
+}
+
+typedef struct size_params{
+ int batch;
+ int inputs;
+ int h;
+ int w;
+ int c;
+} size_params;
+
+deconvolutional_layer parse_deconvolutional(list *options, size_params params)
+{
+ int n = option_find_int(options, "filters",1);
+ int size = option_find_int(options, "size",1);
+ int stride = option_find_int(options, "stride",1);
+ char *activation_s = option_find_str(options, "activation", "logistic");
+ ACTIVATION activation = get_activation(activation_s);
+
+ int batch,h,w,c;
+ h = params.h;
+ w = params.w;
+ c = params.c;
+ batch=params.batch;
+ if(!(h && w && c)) error("Layer before deconvolutional layer must output image.");
+
+ deconvolutional_layer layer = make_deconvolutional_layer(batch,h,w,c,n,size,stride,activation);
+
+ char *weights = option_find_str(options, "weights", 0);
+ char *biases = option_find_str(options, "biases", 0);
+ parse_data(weights, layer.filters, c*n*size*size);
+ parse_data(biases, layer.biases, n);
+ #ifdef GPU
+ if(weights || biases) push_deconvolutional_layer(layer);
+ #endif
+ return layer;
+}
+
+convolutional_layer parse_convolutional(list *options, size_params params)
+{
+ int n = option_find_int(options, "filters",1);
+ int size = option_find_int(options, "size",1);
+ int stride = option_find_int(options, "stride",1);
+ int pad = option_find_int(options, "pad",0);
+ char *activation_s = option_find_str(options, "activation", "logistic");
+ ACTIVATION activation = get_activation(activation_s);
+
+ int batch,h,w,c;
+ h = params.h;
+ w = params.w;
+ c = params.c;
+ batch=params.batch;
+ if(!(h && w && c)) error("Layer before convolutional layer must output image.");
+
+ convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,pad,activation);
+
+ char *weights = option_find_str(options, "weights", 0);
+ char *biases = option_find_str(options, "biases", 0);
+ parse_data(weights, layer.filters, c*n*size*size);
+ parse_data(biases, layer.biases, n);
+ #ifdef GPU
+ if(weights || biases) push_convolutional_layer(layer);
+ #endif
+ return layer;
+}
+
+connected_layer parse_connected(list *options, size_params params)
+{
+ int output = option_find_int(options, "output",1);
+ char *activation_s = option_find_str(options, "activation", "logistic");
+ ACTIVATION activation = get_activation(activation_s);
+
+ connected_layer layer = make_connected_layer(params.batch, params.inputs, output, activation);
+
+ char *weights = option_find_str(options, "weights", 0);
+ char *biases = option_find_str(options, "biases", 0);
+ parse_data(biases, layer.biases, output);
+ parse_data(weights, layer.weights, params.inputs*output);
+ #ifdef GPU
+ if(weights || biases) push_connected_layer(layer);
+ #endif
+ return layer;
+}
+
+softmax_layer parse_softmax(list *options, size_params params)
+{
+ int groups = option_find_int(options, "groups",1);
+ softmax_layer layer = make_softmax_layer(params.batch, params.inputs, groups);
+ return layer;
+}
+
+detection_layer parse_detection(list *options, size_params params)
+{
+ int coords = option_find_int(options, "coords", 1);
+ int classes = option_find_int(options, "classes", 1);
+ int rescore = option_find_int(options, "rescore", 0);
+ int joint = option_find_int(options, "joint", 0);
+ int objectness = option_find_int(options, "objectness", 0);
+ int background = 0;
+ detection_layer layer = make_detection_layer(params.batch, params.inputs, classes, coords, joint, rescore, background, objectness);
+ return layer;
+}
+
+cost_layer parse_cost(list *options, size_params params)
+{
+ char *type_s = option_find_str(options, "type", "sse");
+ COST_TYPE type = get_cost_type(type_s);
+ cost_layer layer = make_cost_layer(params.batch, params.inputs, type);
+ return layer;
+}
+
+crop_layer parse_crop(list *options, size_params params)
+{
+ int crop_height = option_find_int(options, "crop_height",1);
+ int crop_width = option_find_int(options, "crop_width",1);
+ int flip = option_find_int(options, "flip",0);
+ float angle = option_find_float(options, "angle",0);
+ float saturation = option_find_float(options, "saturation",1);
+ float exposure = option_find_float(options, "exposure",1);
+
+ int batch,h,w,c;
+ h = params.h;
+ w = params.w;
+ c = params.c;
+ batch=params.batch;
+ if(!(h && w && c)) error("Layer before crop layer must output image.");
+
+ int noadjust = option_find_int_quiet(options, "noadjust",0);
+
+ crop_layer l = make_crop_layer(batch,h,w,c,crop_height,crop_width,flip, angle, saturation, exposure);
+ l.noadjust = noadjust;
+ return l;
+}
+
+maxpool_layer parse_maxpool(list *options, size_params params)
+{
+ int stride = option_find_int(options, "stride",1);
+ int size = option_find_int(options, "size",stride);
+
+ int batch,h,w,c;
+ h = params.h;
+ w = params.w;
+ c = params.c;
+ batch=params.batch;
+ if(!(h && w && c)) error("Layer before maxpool layer must output image.");
+
+ maxpool_layer layer = make_maxpool_layer(batch,h,w,c,size,stride);
+ return layer;
+}
+
+avgpool_layer parse_avgpool(list *options, size_params params)
+{
+ int batch,w,h,c;
+ w = params.w;
+ h = params.h;
+ c = params.c;
+ batch=params.batch;
+ if(!(h && w && c)) error("Layer before avgpool layer must output image.");
+
+ avgpool_layer layer = make_avgpool_layer(batch,w,h,c);
+ return layer;
+}
+
+dropout_layer parse_dropout(list *options, size_params params)
+{
+ float probability = option_find_float(options, "probability", .5);
+ dropout_layer layer = make_dropout_layer(params.batch, params.inputs, probability);
+ layer.out_w = params.w;
+ layer.out_h = params.h;
+ layer.out_c = params.c;
+ return layer;
+}
+
+layer parse_normalization(list *options, size_params params)
+{
+ float alpha = option_find_float(options, "alpha", .0001);
+ float beta = option_find_float(options, "beta" , .75);
+ float kappa = option_find_float(options, "kappa", 1);
+ int size = option_find_int(options, "size", 5);
+ layer l = make_normalization_layer(params.batch, params.w, params.h, params.c, size, alpha, beta, kappa);
+ return l;
+}
+
+route_layer parse_route(list *options, size_params params, network net)
+{
+ char *l = option_find(options, "layers");
+ int len = strlen(l);
+ if(!l) error("Route Layer must specify input layers");
+ int n = 1;
+ int i;
+ for(i = 0; i < len; ++i){
+ if (l[i] == ',') ++n;
+ }
+
+ int *layers = calloc(n, sizeof(int));
+ int *sizes = calloc(n, sizeof(int));
+ for(i = 0; i < n; ++i){
+ int index = atoi(l);
+ l = strchr(l, ',')+1;
+ layers[i] = index;
+ sizes[i] = net.layers[index].outputs;
+ }
+ int batch = params.batch;
+
+ route_layer layer = make_route_layer(batch, n, layers, sizes);
+
+ convolutional_layer first = net.layers[layers[0]];
+ layer.out_w = first.out_w;
+ layer.out_h = first.out_h;
+ layer.out_c = first.out_c;
+ for(i = 1; i < n; ++i){
+ int index = layers[i];
+ convolutional_layer next = net.layers[index];
+ if(next.out_w == first.out_w && next.out_h == first.out_h){
+ layer.out_c += next.out_c;
+ }else{
+ layer.out_h = layer.out_w = layer.out_c = 0;
+ }
+ }
+
+ return layer;
+}
+
+void parse_net_options(list *options, network *net)
+{
+ net->batch = option_find_int(options, "batch",1);
+ net->learning_rate = option_find_float(options, "learning_rate", .001);
+ net->momentum = option_find_float(options, "momentum", .9);
+ net->decay = option_find_float(options, "decay", .0001);
+ int subdivs = option_find_int(options, "subdivisions",1);
+ net->batch /= subdivs;
+ net->subdivisions = subdivs;
+
+ net->h = option_find_int_quiet(options, "height",0);
+ net->w = option_find_int_quiet(options, "width",0);
+ net->c = option_find_int_quiet(options, "channels",0);
+ net->inputs = option_find_int_quiet(options, "inputs", net->h * net->w * net->c);
+ if(!net->inputs && !(net->h && net->w && net->c)) error("No input parameters supplied");
+}
network parse_network_cfg(char *filename)
{
list *sections = read_cfg(filename);
- network net = make_network(sections->size);
-
node *n = sections->front;
+ if(!n) error("Config file has no sections");
+ network net = make_network(sections->size - 1);
+ size_params params;
+
+ section *s = (section *)n->val;
+ list *options = s->options;
+ if(!is_network(s)) error("First section must be [net] or [network]");
+ parse_net_options(options, &net);
+
+ params.h = net.h;
+ params.w = net.w;
+ params.c = net.c;
+ params.inputs = net.inputs;
+ params.batch = net.batch;
+
+ n = n->next;
int count = 0;
while(n){
- section *s = (section *)n->val;
- list *options = s->options;
+ fprintf(stderr, "%d: ", count);
+ s = (section *)n->val;
+ options = s->options;
+ layer l = {0};
if(is_convolutional(s)){
- int h,w,c;
- int n = option_find_int(options, "filters",1);
- int size = option_find_int(options, "size",1);
- int stride = option_find_int(options, "stride",1);
- char *activation_s = option_find_str(options, "activation", "sigmoid");
- ACTIVATION activation = get_activation(activation_s);
- if(count == 0){
- h = option_find_int(options, "height",1);
- w = option_find_int(options, "width",1);
- c = option_find_int(options, "channels",1);
- }else{
- image m = get_network_image_layer(net, count-1);
- h = m.h;
- w = m.w;
- c = m.c;
- if(h == 0) error("Layer before convolutional layer must output image.");
- }
- convolutional_layer *layer = make_convolutional_layer(h,w,c,n,size,stride, activation);
- net.types[count] = CONVOLUTIONAL;
- net.layers[count] = layer;
- option_unused(options);
- }
- else if(is_connected(s)){
- int input;
- int output = option_find_int(options, "output",1);
- char *activation_s = option_find_str(options, "activation", "sigmoid");
- ACTIVATION activation = get_activation(activation_s);
- if(count == 0){
- input = option_find_int(options, "input",1);
- }else{
- input = get_network_output_size_layer(net, count-1);
- }
- connected_layer *layer = make_connected_layer(input, output, activation);
- net.types[count] = CONNECTED;
- net.layers[count] = layer;
- option_unused(options);
+ l = parse_convolutional(options, params);
+ }else if(is_deconvolutional(s)){
+ l = parse_deconvolutional(options, params);
+ }else if(is_connected(s)){
+ l = parse_connected(options, params);
+ }else if(is_crop(s)){
+ l = parse_crop(options, params);
+ }else if(is_cost(s)){
+ l = parse_cost(options, params);
+ }else if(is_detection(s)){
+ l = parse_detection(options, params);
+ }else if(is_softmax(s)){
+ l = parse_softmax(options, params);
+ }else if(is_normalization(s)){
+ l = parse_normalization(options, params);
}else if(is_maxpool(s)){
- int h,w,c;
- int stride = option_find_int(options, "stride",1);
- //char *activation_s = option_find_str(options, "activation", "sigmoid");
- if(count == 0){
- h = option_find_int(options, "height",1);
- w = option_find_int(options, "width",1);
- c = option_find_int(options, "channels",1);
- }else{
- image m = get_network_image_layer(net, count-1);
- h = m.h;
- w = m.w;
- c = m.c;
- if(h == 0) error("Layer before convolutional layer must output image.");
- }
- maxpool_layer *layer = make_maxpool_layer(h,w,c,stride);
- net.types[count] = MAXPOOL;
- net.layers[count] = layer;
- option_unused(options);
+ l = parse_maxpool(options, params);
+ }else if(is_avgpool(s)){
+ l = parse_avgpool(options, params);
+ }else if(is_route(s)){
+ l = parse_route(options, params, net);
+ }else if(is_dropout(s)){
+ l = parse_dropout(options, params);
+ l.output = net.layers[count-1].output;
+ l.delta = net.layers[count-1].delta;
+ #ifdef GPU
+ l.output_gpu = net.layers[count-1].output_gpu;
+ l.delta_gpu = net.layers[count-1].delta_gpu;
+ #endif
}else{
fprintf(stderr, "Type not recognized: %s\n", s->type);
}
- ++count;
+ l.dontload = option_find_int_quiet(options, "dontload", 0);
+ option_unused(options);
+ net.layers[count] = l;
+ free_section(s);
n = n->next;
+ if(n){
+ params.h = l.out_h;
+ params.w = l.out_w;
+ params.c = l.out_c;
+ params.inputs = l.outputs;
+ }
+ ++count;
}
+ free_list(sections);
+ net.outputs = get_network_output_size(net);
+ net.output = get_network_output(net);
return net;
}
+int is_crop(section *s)
+{
+ return (strcmp(s->type, "[crop]")==0);
+}
+int is_cost(section *s)
+{
+ return (strcmp(s->type, "[cost]")==0);
+}
+int is_detection(section *s)
+{
+ return (strcmp(s->type, "[detection]")==0);
+}
+int is_deconvolutional(section *s)
+{
+ return (strcmp(s->type, "[deconv]")==0
+ || strcmp(s->type, "[deconvolutional]")==0);
+}
int is_convolutional(section *s)
{
return (strcmp(s->type, "[conv]")==0
|| strcmp(s->type, "[convolutional]")==0);
}
+int is_network(section *s)
+{
+ return (strcmp(s->type, "[net]")==0
+ || strcmp(s->type, "[network]")==0);
+}
int is_connected(section *s)
{
return (strcmp(s->type, "[conn]")==0
@@ -112,11 +424,36 @@
return (strcmp(s->type, "[max]")==0
|| strcmp(s->type, "[maxpool]")==0);
}
+int is_avgpool(section *s)
+{
+ return (strcmp(s->type, "[avg]")==0
+ || strcmp(s->type, "[avgpool]")==0);
+}
+int is_dropout(section *s)
+{
+ return (strcmp(s->type, "[dropout]")==0);
+}
+
+int is_normalization(section *s)
+{
+ return (strcmp(s->type, "[lrn]")==0
+ || strcmp(s->type, "[normalization]")==0);
+}
+
+int is_softmax(section *s)
+{
+ return (strcmp(s->type, "[soft]")==0
+ || strcmp(s->type, "[softmax]")==0);
+}
+int is_route(section *s)
+{
+ return (strcmp(s->type, "[route]")==0);
+}
int read_option(char *s, list *options)
{
- int i;
- int len = strlen(s);
+ size_t i;
+ size_t len = strlen(s);
char *val = 0;
for(i = 0; i < len; ++i){
if(s[i] == '='){
@@ -156,7 +493,7 @@
break;
default:
if(!read_option(line, current->options)){
- printf("Config file error line %d, could parse: %s\n", nu, line);
+ fprintf(stderr, "Config file error line %d, could parse: %s\n", nu, line);
free(line);
}
break;
@@ -166,3 +503,109 @@
return sections;
}
+void save_weights_upto(network net, char *filename, int cutoff)
+{
+ fprintf(stderr, "Saving weights to %s\n", filename);
+ FILE *fp = fopen(filename, "w");
+ if(!fp) file_error(filename);
+
+ fwrite(&net.learning_rate, sizeof(float), 1, fp);
+ fwrite(&net.momentum, sizeof(float), 1, fp);
+ fwrite(&net.decay, sizeof(float), 1, fp);
+ fwrite(&net.seen, sizeof(int), 1, fp);
+
+ int i;
+ for(i = 0; i < net.n && i < cutoff; ++i){
+ layer l = net.layers[i];
+ if(l.type == CONVOLUTIONAL){
+#ifdef GPU
+ if(gpu_index >= 0){
+ pull_convolutional_layer(l);
+ }
+#endif
+ int num = l.n*l.c*l.size*l.size;
+ fwrite(l.biases, sizeof(float), l.n, fp);
+ fwrite(l.filters, sizeof(float), num, fp);
+ }
+ if(l.type == DECONVOLUTIONAL){
+#ifdef GPU
+ if(gpu_index >= 0){
+ pull_deconvolutional_layer(l);
+ }
+#endif
+ int num = l.n*l.c*l.size*l.size;
+ fwrite(l.biases, sizeof(float), l.n, fp);
+ fwrite(l.filters, sizeof(float), num, fp);
+ }
+ if(l.type == CONNECTED){
+#ifdef GPU
+ if(gpu_index >= 0){
+ pull_connected_layer(l);
+ }
+#endif
+ fwrite(l.biases, sizeof(float), l.outputs, fp);
+ fwrite(l.weights, sizeof(float), l.outputs*l.inputs, fp);
+ }
+ }
+ fclose(fp);
+}
+void save_weights(network net, char *filename)
+{
+ save_weights_upto(net, filename, net.n);
+}
+
+void load_weights_upto(network *net, char *filename, int cutoff)
+{
+ fprintf(stderr, "Loading weights from %s...", filename);
+ fflush(stdout);
+ FILE *fp = fopen(filename, "r");
+ if(!fp) file_error(filename);
+
+ fread(&net->learning_rate, sizeof(float), 1, fp);
+ fread(&net->momentum, sizeof(float), 1, fp);
+ fread(&net->decay, sizeof(float), 1, fp);
+ fread(&net->seen, sizeof(int), 1, fp);
+
+ int i;
+ for(i = 0; i < net->n && i < cutoff; ++i){
+ layer l = net->layers[i];
+ if (l.dontload) continue;
+ if(l.type == CONVOLUTIONAL){
+ int num = l.n*l.c*l.size*l.size;
+ fread(l.biases, sizeof(float), l.n, fp);
+ fread(l.filters, sizeof(float), num, fp);
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_convolutional_layer(l);
+ }
+#endif
+ }
+ if(l.type == DECONVOLUTIONAL){
+ int num = l.n*l.c*l.size*l.size;
+ fread(l.biases, sizeof(float), l.n, fp);
+ fread(l.filters, sizeof(float), num, fp);
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_deconvolutional_layer(l);
+ }
+#endif
+ }
+ if(l.type == CONNECTED){
+ fread(l.biases, sizeof(float), l.outputs, fp);
+ fread(l.weights, sizeof(float), l.outputs*l.inputs, fp);
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_connected_layer(l);
+ }
+#endif
+ }
+ }
+ fprintf(stderr, "Done!\n");
+ fclose(fp);
+}
+
+void load_weights(network *net, char *filename)
+{
+ load_weights_upto(net, filename, net->n);
+}
+
--
Gitblit v1.10.0