mirror of
https://github.com/pjreddie/darknet.git
synced 2023-08-10 21:13:14 +03:00
New YOLO
This commit is contained in:
Joseph Redmon
@ -6,42 +6,32 @@
|
||||
#include "cuda.h"
|
||||
#include "utils.h"
|
||||
#include <stdio.h>
|
||||
#include <assert.h>
|
||||
#include <string.h>
|
||||
#include <stdlib.h>
|
||||
|
||||
int get_detection_layer_locations(detection_layer l)
|
||||
{
|
||||
return l.inputs / (l.classes+l.coords+l.joint+(l.background || l.objectness));
|
||||
}
|
||||
|
||||
int get_detection_layer_output_size(detection_layer l)
|
||||
{
|
||||
return get_detection_layer_locations(l)*((l.background || l.objectness) + l.classes + l.coords);
|
||||
}
|
||||
|
||||
detection_layer make_detection_layer(int batch, int inputs, int classes, int coords, int joint, int rescore, int background, int objectness)
|
||||
detection_layer make_detection_layer(int batch, int inputs, int n, int side, int classes, int coords, int rescore)
|
||||
{
|
||||
detection_layer l = {0};
|
||||
l.type = DETECTION;
|
||||
|
||||
|
||||
l.n = n;
|
||||
l.batch = batch;
|
||||
l.inputs = inputs;
|
||||
l.classes = classes;
|
||||
l.coords = coords;
|
||||
l.rescore = rescore;
|
||||
l.objectness = objectness;
|
||||
l.background = background;
|
||||
l.joint = joint;
|
||||
l.side = side;
|
||||
assert(side*side*((1 + l.coords)*l.n + l.classes) == inputs);
|
||||
l.cost = calloc(1, sizeof(float));
|
||||
l.does_cost=1;
|
||||
int outputs = get_detection_layer_output_size(l);
|
||||
l.outputs = outputs;
|
||||
l.output = calloc(batch*outputs, sizeof(float));
|
||||
l.delta = calloc(batch*outputs, sizeof(float));
|
||||
#ifdef GPU
|
||||
l.output_gpu = cuda_make_array(l.output, batch*outputs);
|
||||
l.delta_gpu = cuda_make_array(l.delta, batch*outputs);
|
||||
#endif
|
||||
l.outputs = l.inputs;
|
||||
l.truths = l.side*l.side*(1+l.coords+l.classes);
|
||||
l.output = calloc(batch*l.outputs, sizeof(float));
|
||||
l.delta = calloc(batch*l.outputs, sizeof(float));
|
||||
#ifdef GPU
|
||||
l.output_gpu = cuda_make_array(l.output, batch*l.outputs);
|
||||
l.delta_gpu = cuda_make_array(l.delta, batch*l.outputs);
|
||||
#endif
|
||||
|
||||
fprintf(stderr, "Detection Layer\n");
|
||||
srand(0);
|
||||
@ -51,124 +41,164 @@ detection_layer make_detection_layer(int batch, int inputs, int classes, int coo
|
||||
|
||||
void forward_detection_layer(const detection_layer l, network_state state)
|
||||
{
|
||||
int in_i = 0;
|
||||
int out_i = 0;
|
||||
int locations = get_detection_layer_locations(l);
|
||||
int locations = l.side*l.side;
|
||||
int i,j;
|
||||
for(i = 0; i < l.batch*locations; ++i){
|
||||
int mask = (!state.truth || state.truth[out_i + (l.background || l.objectness) + l.classes + 2]);
|
||||
float scale = 1;
|
||||
if(l.joint) scale = state.input[in_i++];
|
||||
else if(l.objectness){
|
||||
l.output[out_i++] = 1-state.input[in_i++];
|
||||
scale = mask;
|
||||
}
|
||||
else if(l.background) l.output[out_i++] = scale*state.input[in_i++];
|
||||
|
||||
for(j = 0; j < l.classes; ++j){
|
||||
l.output[out_i++] = scale*state.input[in_i++];
|
||||
}
|
||||
if(l.objectness){
|
||||
|
||||
}else if(l.background){
|
||||
softmax_array(l.output + out_i - l.classes-l.background, l.classes+l.background, l.output + out_i - l.classes-l.background);
|
||||
activate_array(state.input+in_i, l.coords, LOGISTIC);
|
||||
}
|
||||
for(j = 0; j < l.coords; ++j){
|
||||
l.output[out_i++] = mask*state.input[in_i++];
|
||||
memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));
|
||||
int b;
|
||||
if (l.softmax){
|
||||
for(b = 0; b < l.batch; ++b){
|
||||
int index = b*l.inputs;
|
||||
for (i = 0; i < locations; ++i) {
|
||||
int offset = i*l.classes;
|
||||
softmax_array(l.output + index + offset, l.classes,
|
||||
l.output + index + offset);
|
||||
}
|
||||
int offset = locations*l.classes;
|
||||
activate_array(l.output + index + offset, locations*l.n*(1+l.coords), LOGISTIC);
|
||||
}
|
||||
}
|
||||
float avg_iou = 0;
|
||||
int count = 0;
|
||||
if(l.does_cost && state.train){
|
||||
if(state.train){
|
||||
float avg_iou = 0;
|
||||
float avg_cat = 0;
|
||||
float avg_allcat = 0;
|
||||
float avg_obj = 0;
|
||||
float avg_anyobj = 0;
|
||||
int count = 0;
|
||||
*(l.cost) = 0;
|
||||
int size = get_detection_layer_output_size(l) * l.batch;
|
||||
int size = l.inputs * l.batch;
|
||||
memset(l.delta, 0, size * sizeof(float));
|
||||
for (i = 0; i < l.batch*locations; ++i) {
|
||||
int classes = (l.objectness || l.background)+l.classes;
|
||||
int offset = i*(classes+l.coords);
|
||||
for (j = offset; j < offset+classes; ++j) {
|
||||
*(l.cost) += pow(state.truth[j] - l.output[j], 2);
|
||||
l.delta[j] = state.truth[j] - l.output[j];
|
||||
if(l.background && j == offset) l.delta[j] *= .1;
|
||||
}
|
||||
|
||||
box truth;
|
||||
truth.x = state.truth[j+0]/7;
|
||||
truth.y = state.truth[j+1]/7;
|
||||
truth.w = pow(state.truth[j+2], 2);
|
||||
truth.h = pow(state.truth[j+3], 2);
|
||||
|
||||
box out;
|
||||
out.x = l.output[j+0]/7;
|
||||
out.y = l.output[j+1]/7;
|
||||
out.w = pow(l.output[j+2], 2);
|
||||
out.h = pow(l.output[j+3], 2);
|
||||
|
||||
if(!(truth.w*truth.h)) continue;
|
||||
float iou = box_iou(out, truth);
|
||||
avg_iou += iou;
|
||||
++count;
|
||||
|
||||
*(l.cost) += pow((1-iou), 2);
|
||||
l.delta[j+0] = 4 * (state.truth[j+0] - l.output[j+0]);
|
||||
l.delta[j+1] = 4 * (state.truth[j+1] - l.output[j+1]);
|
||||
l.delta[j+2] = 4 * (state.truth[j+2] - l.output[j+2]);
|
||||
l.delta[j+3] = 4 * (state.truth[j+3] - l.output[j+3]);
|
||||
if(l.rescore){
|
||||
if(l.objectness){
|
||||
state.truth[offset] = iou;
|
||||
l.delta[offset] = state.truth[offset] - l.output[offset];
|
||||
for (b = 0; b < l.batch; ++b){
|
||||
int index = b*l.inputs;
|
||||
for (i = 0; i < locations; ++i) {
|
||||
int truth_index = (b*locations + i)*(1+l.coords+l.classes);
|
||||
int is_obj = state.truth[truth_index];
|
||||
for (j = 0; j < l.n; ++j) {
|
||||
int p_index = index + locations*l.classes + i*l.n + j;
|
||||
l.delta[p_index] = l.noobject_scale*(0 - l.output[p_index]);
|
||||
*(l.cost) += l.noobject_scale*pow(l.output[p_index], 2);
|
||||
avg_anyobj += l.output[p_index];
|
||||
}
|
||||
else{
|
||||
for (j = offset; j < offset+classes; ++j) {
|
||||
if(state.truth[j]) state.truth[j] = iou;
|
||||
l.delta[j] = state.truth[j] - l.output[j];
|
||||
|
||||
int best_index = -1;
|
||||
float best_iou = 0;
|
||||
float best_rmse = 20;
|
||||
|
||||
if (!is_obj){
|
||||
continue;
|
||||
}
|
||||
|
||||
int class_index = index + i*l.classes;
|
||||
for(j = 0; j < l.classes; ++j) {
|
||||
l.delta[class_index+j] = l.class_scale * (state.truth[truth_index+1+j] - l.output[class_index+j]);
|
||||
*(l.cost) += l.class_scale * pow(state.truth[truth_index+1+j] - l.output[class_index+j], 2);
|
||||
if(state.truth[truth_index + 1 + j]) avg_cat += l.output[class_index+j];
|
||||
avg_allcat += l.output[class_index+j];
|
||||
}
|
||||
|
||||
box truth = float_to_box(state.truth + truth_index + 1 + l.classes);
|
||||
truth.x /= l.side;
|
||||
truth.y /= l.side;
|
||||
|
||||
for(j = 0; j < l.n; ++j){
|
||||
int box_index = index + locations*(l.classes + l.n) + (i*l.n + j) * l.coords;
|
||||
box out = float_to_box(l.output + box_index);
|
||||
out.x /= l.side;
|
||||
out.y /= l.side;
|
||||
|
||||
if (l.sqrt){
|
||||
out.w = out.w*out.w;
|
||||
out.h = out.h*out.h;
|
||||
}
|
||||
|
||||
float iou = box_iou(out, truth);
|
||||
//iou = 0;
|
||||
float rmse = box_rmse(out, truth);
|
||||
if(best_iou > 0 || iou > 0){
|
||||
if(iou > best_iou){
|
||||
best_iou = iou;
|
||||
best_index = j;
|
||||
}
|
||||
}else{
|
||||
if(rmse < best_rmse){
|
||||
best_rmse = rmse;
|
||||
best_index = j;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
if(l.forced){
|
||||
if(truth.w*truth.h < .1){
|
||||
best_index = 1;
|
||||
}else{
|
||||
best_index = 0;
|
||||
}
|
||||
}
|
||||
|
||||
int box_index = index + locations*(l.classes + l.n) + (i*l.n + best_index) * l.coords;
|
||||
int tbox_index = truth_index + 1 + l.classes;
|
||||
|
||||
box out = float_to_box(l.output + box_index);
|
||||
out.x /= l.side;
|
||||
out.y /= l.side;
|
||||
if (l.sqrt) {
|
||||
out.w = out.w*out.w;
|
||||
out.h = out.h*out.h;
|
||||
}
|
||||
float iou = box_iou(out, truth);
|
||||
|
||||
//printf("%d", best_index);
|
||||
int p_index = index + locations*l.classes + i*l.n + best_index;
|
||||
*(l.cost) -= l.noobject_scale * pow(l.output[p_index], 2);
|
||||
*(l.cost) += l.object_scale * pow(1-l.output[p_index], 2);
|
||||
avg_obj += l.output[p_index];
|
||||
l.delta[p_index] = l.object_scale * (1.-l.output[p_index]);
|
||||
|
||||
if(l.rescore){
|
||||
l.delta[p_index] = l.object_scale * (iou - l.output[p_index]);
|
||||
}
|
||||
|
||||
l.delta[box_index+0] = l.coord_scale*(state.truth[tbox_index + 0] - l.output[box_index + 0]);
|
||||
l.delta[box_index+1] = l.coord_scale*(state.truth[tbox_index + 1] - l.output[box_index + 1]);
|
||||
l.delta[box_index+2] = l.coord_scale*(state.truth[tbox_index + 2] - l.output[box_index + 2]);
|
||||
l.delta[box_index+3] = l.coord_scale*(state.truth[tbox_index + 3] - l.output[box_index + 3]);
|
||||
if(l.sqrt){
|
||||
l.delta[box_index+2] = l.coord_scale*(sqrt(state.truth[tbox_index + 2]) - l.output[box_index + 2]);
|
||||
l.delta[box_index+3] = l.coord_scale*(sqrt(state.truth[tbox_index + 3]) - l.output[box_index + 3]);
|
||||
}
|
||||
|
||||
*(l.cost) += pow(1-iou, 2);
|
||||
avg_iou += iou;
|
||||
++count;
|
||||
}
|
||||
if(l.softmax){
|
||||
gradient_array(l.output + index + locations*l.classes, locations*l.n*(1+l.coords),
|
||||
LOGISTIC, l.delta + index + locations*l.classes);
|
||||
}
|
||||
}
|
||||
printf("Avg IOU: %f\n", avg_iou/count);
|
||||
printf("Detection Avg IOU: %f, Pos Cat: %f, All Cat: %f, Pos Obj: %f, Any Obj: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_allcat/(count*l.classes), avg_obj/count, avg_anyobj/(l.batch*locations*l.n), count);
|
||||
}
|
||||
}
|
||||
|
||||
void backward_detection_layer(const detection_layer l, network_state state)
|
||||
{
|
||||
int locations = get_detection_layer_locations(l);
|
||||
int i,j;
|
||||
int in_i = 0;
|
||||
int out_i = 0;
|
||||
for(i = 0; i < l.batch*locations; ++i){
|
||||
float scale = 1;
|
||||
float latent_delta = 0;
|
||||
if(l.joint) scale = state.input[in_i++];
|
||||
else if (l.objectness) state.delta[in_i++] += -l.delta[out_i++];
|
||||
else if (l.background) state.delta[in_i++] += scale*l.delta[out_i++];
|
||||
for(j = 0; j < l.classes; ++j){
|
||||
latent_delta += state.input[in_i]*l.delta[out_i];
|
||||
state.delta[in_i++] += scale*l.delta[out_i++];
|
||||
}
|
||||
|
||||
if (l.objectness) {
|
||||
|
||||
}else if (l.background) gradient_array(l.output + out_i, l.coords, LOGISTIC, l.delta + out_i);
|
||||
for (j = 0; j < l.coords; ++j){
|
||||
state.delta[in_i++] += l.delta[out_i++];
|
||||
}
|
||||
if(l.joint) state.delta[in_i-l.coords-l.classes-l.joint] += latent_delta;
|
||||
}
|
||||
axpy_cpu(l.batch*l.inputs, 1, l.delta, 1, state.delta, 1);
|
||||
}
|
||||
|
||||
#ifdef GPU
|
||||
|
||||
void forward_detection_layer_gpu(const detection_layer l, network_state state)
|
||||
{
|
||||
int outputs = get_detection_layer_output_size(l);
|
||||
if(!state.train){
|
||||
copy_ongpu(l.batch*l.inputs, state.input, 1, l.output_gpu, 1);
|
||||
return;
|
||||
}
|
||||
|
||||
float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
|
||||
float *truth_cpu = 0;
|
||||
if(state.truth){
|
||||
truth_cpu = calloc(l.batch*outputs, sizeof(float));
|
||||
cuda_pull_array(state.truth, truth_cpu, l.batch*outputs);
|
||||
int num_truth = l.batch*l.side*l.side*(1+l.coords+l.classes);
|
||||
truth_cpu = calloc(num_truth, sizeof(float));
|
||||
cuda_pull_array(state.truth, truth_cpu, num_truth);
|
||||
}
|
||||
cuda_pull_array(state.input, in_cpu, l.batch*l.inputs);
|
||||
network_state cpu_state;
|
||||
@ -176,38 +206,16 @@ void forward_detection_layer_gpu(const detection_layer l, network_state state)
|
||||
cpu_state.truth = truth_cpu;
|
||||
cpu_state.input = in_cpu;
|
||||
forward_detection_layer(l, cpu_state);
|
||||
cuda_push_array(l.output_gpu, l.output, l.batch*outputs);
|
||||
cuda_push_array(l.delta_gpu, l.delta, l.batch*outputs);
|
||||
cuda_push_array(l.output_gpu, l.output, l.batch*l.outputs);
|
||||
cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs);
|
||||
free(cpu_state.input);
|
||||
if(cpu_state.truth) free(cpu_state.truth);
|
||||
}
|
||||
|
||||
void backward_detection_layer_gpu(detection_layer l, network_state state)
|
||||
{
|
||||
int outputs = get_detection_layer_output_size(l);
|
||||
|
||||
float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
|
||||
float *delta_cpu = calloc(l.batch*l.inputs, sizeof(float));
|
||||
float *truth_cpu = 0;
|
||||
if(state.truth){
|
||||
truth_cpu = calloc(l.batch*outputs, sizeof(float));
|
||||
cuda_pull_array(state.truth, truth_cpu, l.batch*outputs);
|
||||
}
|
||||
network_state cpu_state;
|
||||
cpu_state.train = state.train;
|
||||
cpu_state.input = in_cpu;
|
||||
cpu_state.truth = truth_cpu;
|
||||
cpu_state.delta = delta_cpu;
|
||||
|
||||
cuda_pull_array(state.input, in_cpu, l.batch*l.inputs);
|
||||
cuda_pull_array(state.delta, delta_cpu, l.batch*l.inputs);
|
||||
cuda_pull_array(l.delta_gpu, l.delta, l.batch*outputs);
|
||||
backward_detection_layer(l, cpu_state);
|
||||
cuda_push_array(state.delta, delta_cpu, l.batch*l.inputs);
|
||||
|
||||
if (truth_cpu) free(truth_cpu);
|
||||
free(in_cpu);
|
||||
free(delta_cpu);
|
||||
axpy_ongpu(l.batch*l.inputs, 1, l.delta_gpu, 1, state.delta, 1);
|
||||
//copy_ongpu(l.batch*l.inputs, l.delta_gpu, 1, state.delta, 1);
|
||||
}
|
||||
#endif
|
||||
|
||||
|
||||
Reference in New Issue
Block a user