From 481b57a96a9ef29b112caec1bb3e17ffb043ceae Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sun, 25 Sep 2016 06:12:54 +0000
Subject: [PATCH] So I have this new programming paradigm.......
---
src/parser.c | 460 +++++++++++++++++++++++++++++----------------------------
1 files changed, 235 insertions(+), 225 deletions(-)
diff --git a/src/parser.c b/src/parser.c
index c109a14..a27d245 100644
--- a/src/parser.c
+++ b/src/parser.c
@@ -2,21 +2,26 @@
#include <string.h>
#include <stdlib.h>
+#include "blas.h"
#include "parser.h"
+#include "assert.h"
#include "activations.h"
#include "crop_layer.h"
#include "cost_layer.h"
#include "convolutional_layer.h"
#include "activation_layer.h"
#include "normalization_layer.h"
-#include "deconvolutional_layer.h"
+#include "batchnorm_layer.h"
#include "connected_layer.h"
#include "rnn_layer.h"
+#include "gru_layer.h"
#include "crnn_layer.h"
#include "maxpool_layer.h"
+#include "reorg_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "detection_layer.h"
+#include "region_layer.h"
#include "avgpool_layer.h"
#include "local_layer.h"
#include "route_layer.h"
@@ -30,26 +35,42 @@
list *options;
}section;
-int is_network(section *s);
-int is_convolutional(section *s);
-int is_activation(section *s);
-int is_local(section *s);
-int is_deconvolutional(section *s);
-int is_connected(section *s);
-int is_rnn(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_crop(section *s);
-int is_shortcut(section *s);
-int is_cost(section *s);
-int is_detection(section *s);
-int is_route(section *s);
list *read_cfg(char *filename);
+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, "[region]")==0) return REGION;
+ if (strcmp(type, "[local]")==0) return LOCAL;
+ 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, "[reorg]")==0) return REORG;
+ 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;
+}
+
void free_section(section *s)
{
free(s->type);
@@ -92,33 +113,6 @@
int time_steps;
} 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;
-}
-
local_layer parse_local(list *options, size_params params)
{
int n = option_find_int(options, "filters",1);
@@ -145,7 +139,10 @@
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);
+ int pad = option_find_int_quiet(options, "pad",0);
+ int padding = option_find_int_quiet(options, "padding",0);
+ if(pad) padding = size/2;
+
char *activation_s = option_find_str(options, "activation", "logistic");
ACTIVATION activation = get_activation(activation_s);
@@ -157,18 +154,12 @@
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, binary);
+ convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,padding,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);
- 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;
}
@@ -203,6 +194,16 @@
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);
@@ -212,13 +213,6 @@
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);
- 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;
}
@@ -230,6 +224,32 @@
return layer;
}
+layer parse_region(list *options, size_params params)
+{
+ int coords = option_find_int(options, "coords", 4);
+ int classes = option_find_int(options, "classes", 20);
+ int num = option_find_int(options, "num", 1);
+
+ params.w = option_find_int(options, "side", params.w);
+ params.h = option_find_int(options, "side", params.h);
+
+ layer l = make_region_layer(params.batch, params.w, params.h, num, classes, coords);
+ assert(l.outputs == params.inputs);
+
+ l.log = option_find_int_quiet(options, "log", 0);
+ l.sqrt = option_find_int_quiet(options, "sqrt", 0);
+
+ l.softmax = option_find_int(options, "softmax", 0);
+ l.max_boxes = option_find_int_quiet(options, "max",30);
+ l.jitter = option_find_float(options, "jitter", .2);
+ l.rescore = option_find_int_quiet(options, "rescore",0);
+
+ l.coord_scale = option_find_float(options, "coord_scale", 1);
+ l.object_scale = option_find_float(options, "object_scale", 1);
+ l.noobject_scale = option_find_float(options, "noobject_scale", 1);
+ l.class_scale = option_find_float(options, "class_scale", 1);
+ return l;
+}
detection_layer parse_detection(list *options, size_params params)
{
int coords = option_find_int(options, "coords", 1);
@@ -242,12 +262,15 @@
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);
layer.noobject_scale = option_find_float(options, "noobject_scale", 1);
layer.class_scale = option_find_float(options, "class_scale", 1);
layer.jitter = option_find_float(options, "jitter", .2);
+ layer.random = option_find_int_quiet(options, "random", 0);
+ layer.reorg = option_find_int_quiet(options, "reorg", 0);
return layer;
}
@@ -257,6 +280,7 @@
COST_TYPE type = get_cost_type(type_s);
float scale = option_find_float_quiet(options, "scale",1);
cost_layer layer = make_cost_layer(params.batch, params.inputs, type, scale);
+ layer.ratio = option_find_float_quiet(options, "ratio",0);
return layer;
}
@@ -284,10 +308,26 @@
return l;
}
+layer parse_reorg(list *options, size_params params)
+{
+ int stride = option_find_int(options, "stride",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 reorg layer must output image.");
+
+ layer layer = make_reorg_layer(batch,w,h,c,stride);
+ return layer;
+}
+
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 padding = option_find_int_quiet(options, "padding", (size-1)/2);
int batch,h,w,c;
h = params.h;
@@ -296,7 +336,7 @@
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);
+ maxpool_layer layer = make_maxpool_layer(batch,h,w,c,size,stride,padding);
return layer;
}
@@ -333,6 +373,12 @@
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");
@@ -411,6 +457,7 @@
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;
@@ -438,11 +485,19 @@
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);
+
+ net->angle = option_find_float_quiet(options, "angle", 0);
+ net->aspect = option_find_float_quiet(options, "aspect", 1);
+ net->saturation = option_find_float_quiet(options, "saturation", 1);
+ net->exposure = option_find_float_quiet(options, "exposure", 1);
+ net->hue = option_find_float_quiet(options, "hue", 0);
if(!net->inputs && !(net->h && net->w && net->c)) error("No input parameters supplied");
char *policy_s = option_find_str(options, "policy", "constant");
net->policy = get_policy(policy_s);
+ net->burn_in = option_find_int_quiet(options, "burn_in", 0);
if(net->policy == STEP){
net->step = option_find_int(options, "step", 1);
net->scale = option_find_float(options, "scale", 1);
@@ -475,18 +530,25 @@
} 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);
}
+int is_network(section *s)
+{
+ return (strcmp(s->type, "[net]")==0
+ || strcmp(s->type, "[network]")==0);
+}
+
network parse_network_cfg(char *filename)
{
list *sections = read_cfg(filename);
node *n = sections->front;
if(!n) error("Config file has no sections");
network net = make_network(sections->size - 1);
+ net.gpu_index = gpu_index;
size_params params;
section *s = (section *)n->val;
@@ -501,6 +563,7 @@
params.batch = net.batch;
params.time_steps = net.time_steps;
+ size_t workspace_size = 0;
n = n->next;
int count = 0;
free_section(s);
@@ -510,39 +573,46 @@
s = (section *)n->val;
options = s->options;
layer l = {0};
- if(is_convolutional(s)){
+ LAYER_TYPE lt = string_to_layer_type(s->type);
+ if(lt == CONVOLUTIONAL){
l = parse_convolutional(options, params);
- }else if(is_local(s)){
+ }else if(lt == LOCAL){
l = parse_local(options, params);
- }else if(is_activation(s)){
+ }else if(lt == ACTIVE){
l = parse_activation(options, params);
- }else if(is_deconvolutional(s)){
- l = parse_deconvolutional(options, params);
- }else if(is_rnn(s)){
+ }else if(lt == RNN){
l = parse_rnn(options, params);
- }else if(is_crnn(s)){
+ }else if(lt == GRU){
+ l = parse_gru(options, params);
+ }else if(lt == CRNN){
l = parse_crnn(options, params);
- }else if(is_connected(s)){
+ }else if(lt == CONNECTED){
l = parse_connected(options, params);
- }else if(is_crop(s)){
+ }else if(lt == CROP){
l = parse_crop(options, params);
- }else if(is_cost(s)){
+ }else if(lt == COST){
l = parse_cost(options, params);
- }else if(is_detection(s)){
+ }else if(lt == REGION){
+ l = parse_region(options, params);
+ }else if(lt == DETECTION){
l = parse_detection(options, params);
- }else if(is_softmax(s)){
+ }else if(lt == SOFTMAX){
l = parse_softmax(options, params);
- }else if(is_normalization(s)){
+ }else if(lt == NORMALIZATION){
l = parse_normalization(options, params);
- }else if(is_maxpool(s)){
+ }else if(lt == BATCHNORM){
+ l = parse_batchnorm(options, params);
+ }else if(lt == MAXPOOL){
l = parse_maxpool(options, params);
- }else if(is_avgpool(s)){
+ }else if(lt == REORG){
+ l = parse_reorg(options, params);
+ }else if(lt == AVGPOOL){
l = parse_avgpool(options, params);
- }else if(is_route(s)){
+ }else if(lt == ROUTE){
l = parse_route(options, params, net);
- }else if(is_shortcut(s)){
+ }else if(lt == SHORTCUT){
l = parse_shortcut(options, params, net);
- }else if(is_dropout(s)){
+ }else if(lt == DROPOUT){
l = parse_dropout(options, params);
l.output = net.layers[count-1].output;
l.delta = net.layers[count-1].delta;
@@ -557,6 +627,7 @@
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;
@@ -570,92 +641,21 @@
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
+ if(gpu_index >= 0){
+ net.workspace = cuda_make_array(0, (workspace_size-1)/sizeof(float)+1);
+ }else {
+ net.workspace = calloc(1, workspace_size);
+ }
+#else
+ net.workspace = calloc(1, workspace_size);
+#endif
+ }
return net;
}
-int is_shortcut(section *s)
-{
- return (strcmp(s->type, "[shortcut]")==0);
-}
-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_local(section *s)
-{
- return (strcmp(s->type, "[local]")==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_activation(section *s)
-{
- return (strcmp(s->type, "[activation]")==0);
-}
-int is_network(section *s)
-{
- return (strcmp(s->type, "[net]")==0
- || strcmp(s->type, "[network]")==0);
-}
-int is_crnn(section *s)
-{
- return (strcmp(s->type, "[crnn]")==0);
-}
-int is_rnn(section *s)
-{
- return (strcmp(s->type, "[rnn]")==0);
-}
-int is_connected(section *s)
-{
- return (strcmp(s->type, "[conn]")==0
- || strcmp(s->type, "[connected]")==0);
-}
-int is_maxpool(section *s)
-{
- 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);
-}
-
list *read_cfg(char *filename)
{
FILE *file = fopen(filename, "r");
@@ -691,45 +691,6 @@
return sections;
}
-void save_weights_double(network net, char *filename)
-{
- fprintf(stderr, "Saving doubled 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,j,k;
- for(i = 0; i < net.n; ++i){
- layer l = net.layers[i];
- if(l.type == CONVOLUTIONAL){
-#ifdef GPU
- if(gpu_index >= 0){
- pull_convolutional_layer(l);
- }
-#endif
- float zero = 0;
- fwrite(l.biases, sizeof(float), l.n, fp);
- fwrite(l.biases, sizeof(float), l.n, fp);
-
- for (j = 0; j < l.n; ++j){
- int index = j*l.c*l.size*l.size;
- fwrite(l.filters+index, sizeof(float), l.c*l.size*l.size, fp);
- for (k = 0; k < l.c*l.size*l.size; ++k) fwrite(&zero, sizeof(float), 1, fp);
- }
- for (j = 0; j < l.n; ++j){
- int index = j*l.c*l.size*l.size;
- for (k = 0; k < l.c*l.size*l.size; ++k) fwrite(&zero, sizeof(float), 1, fp);
- fwrite(l.filters+index, sizeof(float), l.c*l.size*l.size, fp);
- }
- }
- }
- fclose(fp);
-}
-
void save_convolutional_weights_binary(layer l, FILE *fp)
{
#ifdef GPU
@@ -737,7 +698,7 @@
pull_convolutional_layer(l);
}
#endif
- binarize_filters(l.filters, l.n, l.c*l.size*l.size, l.binary_filters);
+ binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
int size = l.c*l.size*l.size;
int i, j, k;
fwrite(l.biases, sizeof(float), l.n, fp);
@@ -747,7 +708,7 @@
fwrite(l.rolling_variance, sizeof(float), l.n, fp);
}
for(i = 0; i < l.n; ++i){
- float mean = l.binary_filters[i*size];
+ float mean = l.binary_weights[i*size];
if(mean < 0) mean = -mean;
fwrite(&mean, sizeof(float), 1, fp);
for(j = 0; j < size/8; ++j){
@@ -755,7 +716,7 @@
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);
+ if (l.binary_weights[index + k] > 0) c = (c | 1<<k);
}
fwrite(&c, sizeof(char), 1, fp);
}
@@ -780,7 +741,19 @@
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);
+ fwrite(l.weights, 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)
@@ -801,6 +774,11 @@
void save_weights_upto(network net, char *filename, int cutoff)
{
+#ifdef GPU
+ if(net.gpu_index >= 0){
+ cuda_set_device(net.gpu_index);
+ }
+#endif
fprintf(stderr, "Saving weights to %s\n", filename);
FILE *fp = fopen(filename, "w");
if(!fp) file_error(filename);
@@ -820,10 +798,19 @@
save_convolutional_weights(l, fp);
} if(l.type == CONNECTED){
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);
@@ -837,7 +824,7 @@
int locations = l.out_w*l.out_h;
int size = l.size*l.size*l.c*l.n*locations;
fwrite(l.biases, sizeof(float), l.outputs, fp);
- fwrite(l.filters, sizeof(float), size, fp);
+ fwrite(l.weights, sizeof(float), size, fp);
}
}
fclose(fp);
@@ -867,10 +854,15 @@
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){
@@ -879,6 +871,18 @@
#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);
@@ -898,11 +902,10 @@
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;
+ l.weights[index + k] = (c & 1<<k) ? mean : -mean;
}
}
}
- binarize_filters2(l.filters, l.n, l.c*l.size*l.size, l.cfilters, l.scales);
#ifdef GPU
if(gpu_index >= 0){
push_convolutional_layer(l);
@@ -923,11 +926,12 @@
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);
+ fread(l.weights, sizeof(float), num, fp);
+ //if(l.c == 3) scal_cpu(num, 1./256, l.weights, 1);
if (l.flipped) {
- transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
+ transpose_matrix(l.weights, 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);
+ //if (l.binary) binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.weights);
#ifdef GPU
if(gpu_index >= 0){
push_convolutional_layer(l);
@@ -938,6 +942,11 @@
void load_weights_upto(network *net, char *filename, int cutoff)
{
+#ifdef GPU
+ if(net->gpu_index >= 0){
+ cuda_set_device(net->gpu_index);
+ }
+#endif
fprintf(stderr, "Loading weights from %s...", filename);
fflush(stdout);
FILE *fp = fopen(filename, "rb");
@@ -959,19 +968,12 @@
if(l.type == CONVOLUTIONAL){
load_convolutional_weights(l, fp);
}
- 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){
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);
@@ -982,11 +984,19 @@
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;
int size = l.size*l.size*l.c*l.n*locations;
fread(l.biases, sizeof(float), l.outputs, fp);
- fread(l.filters, sizeof(float), size, fp);
+ fread(l.weights, sizeof(float), size, fp);
#ifdef GPU
if(gpu_index >= 0){
push_local_layer(l);
--
Gitblit v1.10.0