From 08c7cf9c88befd845f00c00d85e40a9eead4b1b3 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 19 Jun 2016 21:28:15 +0000
Subject: [PATCH] no mean on input binarization
---
src/parser.c | 409 +++++++++++++++++++++++++++++++++++++++++++++++++++-------
1 files changed, 358 insertions(+), 51 deletions(-)
diff --git a/src/parser.c b/src/parser.c
index 218fd27..71f54cc 100644
--- a/src/parser.c
+++ b/src/parser.c
@@ -9,8 +9,12 @@
#include "convolutional_layer.h"
#include "activation_layer.h"
#include "normalization_layer.h"
+#include "batchnorm_layer.h"
#include "deconvolutional_layer.h"
#include "connected_layer.h"
+#include "rnn_layer.h"
+#include "gru_layer.h"
+#include "crnn_layer.h"
#include "maxpool_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
@@ -34,11 +38,15 @@
int is_local(section *s);
int is_deconvolutional(section *s);
int is_connected(section *s);
+int is_rnn(section *s);
+int is_gru(section *s);
+int is_crnn(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_batchnorm(section *s);
int is_crop(section *s);
int is_shortcut(section *s);
int is_cost(section *s);
@@ -85,6 +93,7 @@
int w;
int c;
int index;
+ int time_steps;
} size_params;
deconvolutional_layer parse_deconvolutional(list *options, size_params params)
@@ -151,9 +160,12 @@
batch=params.batch;
if(!(h && w && c)) error("Layer before convolutional layer must output image.");
int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
+ int binary = option_find_int_quiet(options, "binary", 0);
+ int xnor = option_find_int_quiet(options, "xnor", 0);
- convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,pad,activation, batch_normalize);
+ convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,pad,activation, batch_normalize, binary, xnor);
layer.flipped = option_find_int_quiet(options, "flipped", 0);
+ layer.dot = option_find_float_quiet(options, "dot", 0);
char *weights = option_find_str(options, "weights", 0);
char *biases = option_find_str(options, "biases", 0);
@@ -165,13 +177,55 @@
return layer;
}
+layer parse_crnn(list *options, size_params params)
+{
+ int output_filters = option_find_int(options, "output_filters",1);
+ int hidden_filters = option_find_int(options, "hidden_filters",1);
+ char *activation_s = option_find_str(options, "activation", "logistic");
+ ACTIVATION activation = get_activation(activation_s);
+ int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
+
+ layer l = make_crnn_layer(params.batch, params.w, params.h, params.c, hidden_filters, output_filters, params.time_steps, activation, batch_normalize);
+
+ l.shortcut = option_find_int_quiet(options, "shortcut", 0);
+
+ return l;
+}
+
+layer parse_rnn(list *options, size_params params)
+{
+ int output = option_find_int(options, "output",1);
+ int hidden = option_find_int(options, "hidden",1);
+ char *activation_s = option_find_str(options, "activation", "logistic");
+ ACTIVATION activation = get_activation(activation_s);
+ int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
+ int logistic = option_find_int_quiet(options, "logistic", 0);
+
+ layer l = make_rnn_layer(params.batch, params.inputs, hidden, output, params.time_steps, activation, batch_normalize, logistic);
+
+ l.shortcut = option_find_int_quiet(options, "shortcut", 0);
+
+ return l;
+}
+
+layer parse_gru(list *options, size_params params)
+{
+ int output = option_find_int(options, "output",1);
+ int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
+
+ layer l = make_gru_layer(params.batch, params.inputs, output, params.time_steps, batch_normalize);
+
+ return l;
+}
+
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);
+ int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
- connected_layer layer = make_connected_layer(params.batch, params.inputs, output, activation);
+ connected_layer layer = make_connected_layer(params.batch, params.inputs, output, activation, batch_normalize);
char *weights = option_find_str(options, "weights", 0);
char *biases = option_find_str(options, "biases", 0);
@@ -185,8 +239,9 @@
softmax_layer parse_softmax(list *options, size_params params)
{
- int groups = option_find_int(options, "groups",1);
+ int groups = option_find_int_quiet(options, "groups",1);
softmax_layer layer = make_softmax_layer(params.batch, params.inputs, groups);
+ layer.temperature = option_find_float_quiet(options, "temperature", 1);
return layer;
}
@@ -202,6 +257,7 @@
layer.softmax = option_find_int(options, "softmax", 0);
layer.sqrt = option_find_int(options, "sqrt", 0);
+ layer.max_boxes = option_find_int_quiet(options, "max",30);
layer.coord_scale = option_find_float(options, "coord_scale", 1);
layer.forced = option_find_int(options, "forced", 0);
layer.object_scale = option_find_float(options, "object_scale", 1);
@@ -293,6 +349,12 @@
return l;
}
+layer parse_batchnorm(list *options, size_params params)
+{
+ layer l = make_batchnorm_layer(params.batch, params.w, params.h, params.c);
+ return l;
+}
+
layer parse_shortcut(list *options, size_params params, network net)
{
char *l = option_find(options, "from");
@@ -371,6 +433,7 @@
learning_rate_policy get_policy(char *s)
{
+ if (strcmp(s, "random")==0) return RANDOM;
if (strcmp(s, "poly")==0) return POLY;
if (strcmp(s, "constant")==0) return CONSTANT;
if (strcmp(s, "step")==0) return STEP;
@@ -388,13 +451,17 @@
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->time_steps = option_find_int_quiet(options, "time_steps",1);
net->batch /= subdivs;
+ net->batch *= net->time_steps;
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);
+ net->max_crop = option_find_int_quiet(options, "max_crop",net->w*2);
+ net->min_crop = option_find_int_quiet(options, "min_crop",net->w);
if(!net->inputs && !(net->h && net->w && net->c)) error("No input parameters supplied");
@@ -432,7 +499,7 @@
} else if (net->policy == SIG){
net->gamma = option_find_float(options, "gamma", 1);
net->step = option_find_int(options, "step", 1);
- } else if (net->policy == POLY){
+ } else if (net->policy == POLY || net->policy == RANDOM){
net->power = option_find_float(options, "power", 1);
}
net->max_batches = option_find_int(options, "max_batches", 0);
@@ -456,7 +523,9 @@
params.c = net.c;
params.inputs = net.inputs;
params.batch = net.batch;
+ params.time_steps = net.time_steps;
+ size_t workspace_size = 0;
n = n->next;
int count = 0;
free_section(s);
@@ -474,6 +543,12 @@
l = parse_activation(options, params);
}else if(is_deconvolutional(s)){
l = parse_deconvolutional(options, params);
+ }else if(is_rnn(s)){
+ l = parse_rnn(options, params);
+ }else if(is_gru(s)){
+ l = parse_gru(options, params);
+ }else if(is_crnn(s)){
+ l = parse_crnn(options, params);
}else if(is_connected(s)){
l = parse_connected(options, params);
}else if(is_crop(s)){
@@ -486,6 +561,8 @@
l = parse_softmax(options, params);
}else if(is_normalization(s)){
l = parse_normalization(options, params);
+ }else if(is_batchnorm(s)){
+ l = parse_batchnorm(options, params);
}else if(is_maxpool(s)){
l = parse_maxpool(options, params);
}else if(is_avgpool(s)){
@@ -509,6 +586,7 @@
l.dontloadscales = option_find_int_quiet(options, "dontloadscales", 0);
option_unused(options);
net.layers[count] = l;
+ if (l.workspace_size > workspace_size) workspace_size = l.workspace_size;
free_section(s);
n = n->next;
++count;
@@ -522,9 +600,51 @@
free_list(sections);
net.outputs = get_network_output_size(net);
net.output = get_network_output(net);
+ if(workspace_size){
+ //printf("%ld\n", workspace_size);
+#ifdef GPU
+ net.workspace = cuda_make_array(0, (workspace_size-1)/sizeof(float)+1);
+#else
+ net.workspace = calloc(1, workspace_size);
+#endif
+ }
return net;
}
+LAYER_TYPE string_to_layer_type(char * type)
+{
+
+ if (strcmp(type, "[shortcut]")==0) return SHORTCUT;
+ if (strcmp(type, "[crop]")==0) return CROP;
+ if (strcmp(type, "[cost]")==0) return COST;
+ if (strcmp(type, "[detection]")==0) return DETECTION;
+ if (strcmp(type, "[local]")==0) return LOCAL;
+ if (strcmp(type, "[deconv]")==0
+ || strcmp(type, "[deconvolutional]")==0) return DECONVOLUTIONAL;
+ if (strcmp(type, "[conv]")==0
+ || strcmp(type, "[convolutional]")==0) return CONVOLUTIONAL;
+ if (strcmp(type, "[activation]")==0) return ACTIVE;
+ if (strcmp(type, "[net]")==0
+ || strcmp(type, "[network]")==0) return NETWORK;
+ if (strcmp(type, "[crnn]")==0) return CRNN;
+ if (strcmp(type, "[gru]")==0) return GRU;
+ if (strcmp(type, "[rnn]")==0) return RNN;
+ if (strcmp(type, "[conn]")==0
+ || strcmp(type, "[connected]")==0) return CONNECTED;
+ if (strcmp(type, "[max]")==0
+ || strcmp(type, "[maxpool]")==0) return MAXPOOL;
+ if (strcmp(type, "[avg]")==0
+ || strcmp(type, "[avgpool]")==0) return AVGPOOL;
+ if (strcmp(type, "[dropout]")==0) return DROPOUT;
+ if (strcmp(type, "[lrn]")==0
+ || strcmp(type, "[normalization]")==0) return NORMALIZATION;
+ if (strcmp(type, "[batchnorm]")==0) return BATCHNORM;
+ if (strcmp(type, "[soft]")==0
+ || strcmp(type, "[softmax]")==0) return SOFTMAX;
+ if (strcmp(type, "[route]")==0) return ROUTE;
+ return BLANK;
+}
+
int is_shortcut(section *s)
{
return (strcmp(s->type, "[shortcut]")==0);
@@ -564,6 +684,18 @@
return (strcmp(s->type, "[net]")==0
|| strcmp(s->type, "[network]")==0);
}
+int is_crnn(section *s)
+{
+ return (strcmp(s->type, "[crnn]")==0);
+}
+int is_gru(section *s)
+{
+ return (strcmp(s->type, "[gru]")==0);
+}
+int is_rnn(section *s)
+{
+ return (strcmp(s->type, "[rnn]")==0);
+}
int is_connected(section *s)
{
return (strcmp(s->type, "[conn]")==0
@@ -590,6 +722,11 @@
|| strcmp(s->type, "[normalization]")==0);
}
+int is_batchnorm(section *s)
+{
+ return (strcmp(s->type, "[batchnorm]")==0);
+}
+
int is_softmax(section *s)
{
return (strcmp(s->type, "[soft]")==0
@@ -674,6 +811,87 @@
fclose(fp);
}
+void save_convolutional_weights_binary(layer l, FILE *fp)
+{
+#ifdef GPU
+ if(gpu_index >= 0){
+ pull_convolutional_layer(l);
+ }
+#endif
+ binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
+ int size = l.c*l.size*l.size;
+ int i, j, k;
+ fwrite(l.biases, sizeof(float), l.n, fp);
+ if (l.batch_normalize){
+ fwrite(l.scales, sizeof(float), l.n, fp);
+ fwrite(l.rolling_mean, sizeof(float), l.n, fp);
+ fwrite(l.rolling_variance, sizeof(float), l.n, fp);
+ }
+ for(i = 0; i < l.n; ++i){
+ float mean = l.binary_filters[i*size];
+ if(mean < 0) mean = -mean;
+ fwrite(&mean, sizeof(float), 1, fp);
+ for(j = 0; j < size/8; ++j){
+ int index = i*size + j*8;
+ unsigned char c = 0;
+ for(k = 0; k < 8; ++k){
+ if (j*8 + k >= size) break;
+ if (l.binary_filters[index + k] > 0) c = (c | 1<<k);
+ }
+ fwrite(&c, sizeof(char), 1, fp);
+ }
+ }
+}
+
+void save_convolutional_weights(layer l, FILE *fp)
+{
+ if(l.binary){
+ //save_convolutional_weights_binary(l, fp);
+ //return;
+ }
+#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);
+ if (l.batch_normalize){
+ fwrite(l.scales, sizeof(float), l.n, fp);
+ fwrite(l.rolling_mean, sizeof(float), l.n, fp);
+ fwrite(l.rolling_variance, sizeof(float), l.n, fp);
+ }
+ fwrite(l.filters, sizeof(float), num, fp);
+}
+
+void save_batchnorm_weights(layer l, FILE *fp)
+{
+#ifdef GPU
+ if(gpu_index >= 0){
+ pull_batchnorm_layer(l);
+ }
+#endif
+ fwrite(l.scales, sizeof(float), l.c, fp);
+ fwrite(l.rolling_mean, sizeof(float), l.c, fp);
+ fwrite(l.rolling_variance, sizeof(float), l.c, fp);
+}
+
+void save_connected_weights(layer l, FILE *fp)
+{
+#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);
+ if (l.batch_normalize){
+ fwrite(l.scales, sizeof(float), l.outputs, fp);
+ fwrite(l.rolling_mean, sizeof(float), l.outputs, fp);
+ fwrite(l.rolling_variance, sizeof(float), l.outputs, fp);
+ }
+}
+
void save_weights_upto(network net, char *filename, int cutoff)
{
fprintf(stderr, "Saving weights to %s\n", filename);
@@ -692,27 +910,26 @@
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);
- if (l.batch_normalize){
- fwrite(l.scales, sizeof(float), l.n, fp);
- fwrite(l.rolling_mean, sizeof(float), l.n, fp);
- fwrite(l.rolling_variance, sizeof(float), l.n, fp);
- }
- fwrite(l.filters, sizeof(float), num, fp);
+ save_convolutional_weights(l, 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);
+ save_connected_weights(l, fp);
+ } if(l.type == BATCHNORM){
+ save_batchnorm_weights(l, fp);
+ } if(l.type == RNN){
+ save_connected_weights(*(l.input_layer), fp);
+ save_connected_weights(*(l.self_layer), fp);
+ save_connected_weights(*(l.output_layer), fp);
+ } if(l.type == GRU){
+ save_connected_weights(*(l.input_z_layer), fp);
+ save_connected_weights(*(l.input_r_layer), fp);
+ save_connected_weights(*(l.input_h_layer), fp);
+ save_connected_weights(*(l.state_z_layer), fp);
+ save_connected_weights(*(l.state_r_layer), fp);
+ save_connected_weights(*(l.state_h_layer), fp);
+ } if(l.type == CRNN){
+ save_convolutional_weights(*(l.input_layer), fp);
+ save_convolutional_weights(*(l.self_layer), fp);
+ save_convolutional_weights(*(l.output_layer), fp);
} if(l.type == LOCAL){
#ifdef GPU
if(gpu_index >= 0){
@@ -745,11 +962,103 @@
free(transpose);
}
+void load_connected_weights(layer l, FILE *fp, int transpose)
+{
+ fread(l.biases, sizeof(float), l.outputs, fp);
+ fread(l.weights, sizeof(float), l.outputs*l.inputs, fp);
+ if(transpose){
+ transpose_matrix(l.weights, l.inputs, l.outputs);
+ }
+ //printf("Biases: %f mean %f variance\n", mean_array(l.biases, l.outputs), variance_array(l.biases, l.outputs));
+ //printf("Weights: %f mean %f variance\n", mean_array(l.weights, l.outputs*l.inputs), variance_array(l.weights, l.outputs*l.inputs));
+ if (l.batch_normalize && (!l.dontloadscales)){
+ fread(l.scales, sizeof(float), l.outputs, fp);
+ fread(l.rolling_mean, sizeof(float), l.outputs, fp);
+ fread(l.rolling_variance, sizeof(float), l.outputs, fp);
+ //printf("Scales: %f mean %f variance\n", mean_array(l.scales, l.outputs), variance_array(l.scales, l.outputs));
+ //printf("rolling_mean: %f mean %f variance\n", mean_array(l.rolling_mean, l.outputs), variance_array(l.rolling_mean, l.outputs));
+ //printf("rolling_variance: %f mean %f variance\n", mean_array(l.rolling_variance, l.outputs), variance_array(l.rolling_variance, l.outputs));
+ }
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_connected_layer(l);
+ }
+#endif
+}
+
+void load_batchnorm_weights(layer l, FILE *fp)
+{
+ fread(l.scales, sizeof(float), l.c, fp);
+ fread(l.rolling_mean, sizeof(float), l.c, fp);
+ fread(l.rolling_variance, sizeof(float), l.c, fp);
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_batchnorm_layer(l);
+ }
+#endif
+}
+
+void load_convolutional_weights_binary(layer l, FILE *fp)
+{
+ fread(l.biases, sizeof(float), l.n, fp);
+ if (l.batch_normalize && (!l.dontloadscales)){
+ fread(l.scales, sizeof(float), l.n, fp);
+ fread(l.rolling_mean, sizeof(float), l.n, fp);
+ fread(l.rolling_variance, sizeof(float), l.n, fp);
+ }
+ int size = l.c*l.size*l.size;
+ int i, j, k;
+ for(i = 0; i < l.n; ++i){
+ float mean = 0;
+ fread(&mean, sizeof(float), 1, fp);
+ for(j = 0; j < size/8; ++j){
+ int index = i*size + j*8;
+ unsigned char c = 0;
+ fread(&c, sizeof(char), 1, fp);
+ for(k = 0; k < 8; ++k){
+ if (j*8 + k >= size) break;
+ l.filters[index + k] = (c & 1<<k) ? mean : -mean;
+ }
+ }
+ }
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_convolutional_layer(l);
+ }
+#endif
+}
+
+void load_convolutional_weights(layer l, FILE *fp)
+{
+ if(l.binary){
+ //load_convolutional_weights_binary(l, fp);
+ //return;
+ }
+ int num = l.n*l.c*l.size*l.size;
+ fread(l.biases, sizeof(float), l.n, fp);
+ if (l.batch_normalize && (!l.dontloadscales)){
+ fread(l.scales, sizeof(float), l.n, fp);
+ fread(l.rolling_mean, sizeof(float), l.n, fp);
+ fread(l.rolling_variance, sizeof(float), l.n, fp);
+ }
+ fread(l.filters, sizeof(float), num, fp);
+ if (l.flipped) {
+ transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
+ }
+ //if (l.binary) binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.filters);
+#ifdef GPU
+ if(gpu_index >= 0){
+ push_convolutional_layer(l);
+ }
+#endif
+}
+
+
void load_weights_upto(network *net, char *filename, int cutoff)
{
fprintf(stderr, "Loading weights from %s...", filename);
fflush(stdout);
- FILE *fp = fopen(filename, "r");
+ FILE *fp = fopen(filename, "rb");
if(!fp) file_error(filename);
int major;
@@ -759,28 +1068,14 @@
fread(&minor, sizeof(int), 1, fp);
fread(&revision, sizeof(int), 1, fp);
fread(net->seen, sizeof(int), 1, fp);
+ int transpose = (major > 1000) || (minor > 1000);
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);
- if (l.batch_normalize && (!l.dontloadscales)){
- fread(l.scales, sizeof(float), l.n, fp);
- fread(l.rolling_mean, sizeof(float), l.n, fp);
- fread(l.rolling_variance, sizeof(float), l.n, fp);
- }
- fread(l.filters, sizeof(float), num, fp);
- if (l.flipped) {
- transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
- }
-#ifdef GPU
- if(gpu_index >= 0){
- push_convolutional_layer(l);
- }
-#endif
+ load_convolutional_weights(l, fp);
}
if(l.type == DECONVOLUTIONAL){
int num = l.n*l.c*l.size*l.size;
@@ -793,16 +1088,28 @@
#endif
}
if(l.type == CONNECTED){
- fread(l.biases, sizeof(float), l.outputs, fp);
- fread(l.weights, sizeof(float), l.outputs*l.inputs, fp);
- if(major > 1000 || minor > 1000){
- transpose_matrix(l.weights, l.inputs, l.outputs);
- }
-#ifdef GPU
- if(gpu_index >= 0){
- push_connected_layer(l);
- }
-#endif
+ load_connected_weights(l, fp, transpose);
+ }
+ if(l.type == BATCHNORM){
+ load_batchnorm_weights(l, fp);
+ }
+ if(l.type == CRNN){
+ load_convolutional_weights(*(l.input_layer), fp);
+ load_convolutional_weights(*(l.self_layer), fp);
+ load_convolutional_weights(*(l.output_layer), fp);
+ }
+ if(l.type == RNN){
+ load_connected_weights(*(l.input_layer), fp, transpose);
+ load_connected_weights(*(l.self_layer), fp, transpose);
+ load_connected_weights(*(l.output_layer), fp, transpose);
+ }
+ if(l.type == GRU){
+ load_connected_weights(*(l.input_z_layer), fp, transpose);
+ load_connected_weights(*(l.input_r_layer), fp, transpose);
+ load_connected_weights(*(l.input_h_layer), fp, transpose);
+ load_connected_weights(*(l.state_z_layer), fp, transpose);
+ load_connected_weights(*(l.state_r_layer), fp, transpose);
+ load_connected_weights(*(l.state_h_layer), fp, transpose);
}
if(l.type == LOCAL){
int locations = l.out_w*l.out_h;
--
Gitblit v1.10.0