void forward_detection_layer_gpu(const detection_layer l, network_state state)
{
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){
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 = state;
cpu_state.train = state.train;
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*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 pull_connected_layer(connected_layer l)
{
cuda_pull_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
}
void pull_local_layer(local_layer l)
{
int locations = l.out_w*l.out_h;
int size = l.size*l.size*l.c*l.n*locations;
cuda_pull_array(l.weights_gpu, l.weights, size);
cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
}
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);
}
void pull_weights(layer l)
{
if(l.type == CONVOLUTIONAL){
cuda_pull_array(l.biases_gpu, l.biases, l.n);
cuda_pull_array(l.weights_gpu, l.weights, l.n*l.size*l.size*l.c);
if(l.scales) cuda_pull_array(l.scales_gpu, l.scales, l.n);
} else if(l.type == CONNECTED){
cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
cuda_pull_array(l.weights_gpu, l.weights, l.outputs*l.inputs);
}
}
void test_resize(char *filename)
{
image im = load_image(filename, 0,0, 3);
float mag = mag_array(im.data, im.w*im.h*im.c);
printf("L2 Norm: %f\n", mag);
image gray = grayscale_image(im);
image sat2 = copy_image(im);
saturate_image(sat2, 2);
image sat5 = copy_image(im);
saturate_image(sat5, .5);
image exp2 = copy_image(im);
exposure_image(exp2, 2);
image exp5 = copy_image(im);
exposure_image(exp5, .5);
#ifdef GPU
image r = resize_image(im, im.w, im.h);
image black = make_image(im.w*2 + 3, im.h*2 + 3, 9);
image black2 = make_image(im.w, im.h, 3);
float *r_gpu = cuda_make_array(r.data, r.w*r.h*r.c);
float *black_gpu = cuda_make_array(black.data, black.w*black.h*black.c);
float *black2_gpu = cuda_make_array(black2.data, black2.w*black2.h*black2.c);
shortcut_gpu(3, r.w, r.h, 1, r_gpu, black.w, black.h, 3, black_gpu);
//flip_image(r);
//shortcut_gpu(3, r.w, r.h, 1, r.data, black.w, black.h, 3, black.data);
shortcut_gpu(3, black.w, black.h, 3, black_gpu, black2.w, black2.h, 1, black2_gpu);
cuda_pull_array(black_gpu, black.data, black.w*black.h*black.c);
cuda_pull_array(black2_gpu, black2.data, black2.w*black2.h*black2.c);
show_image_layers(black, "Black");
show_image(black2, "Recreate");
#endif
show_image(im, "Original");
show_image(gray, "Gray");
show_image(sat2, "Saturation-2");
show_image(sat5, "Saturation-.5");
show_image(exp2, "Exposure-2");
show_image(exp5, "Exposure-.5");
#ifdef OPENCV
cvWaitKey(0);
#endif
}
float cuda_mag_array(float *x_gpu, size_t n) {
float *temp = calloc(n, sizeof(float));
cuda_pull_array(x_gpu, temp, n);
float m = mag_array(temp, n);
free(temp);
return m;
}
void forward_region_layer_gpu(const region_layer l, network_state state)
{
float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
float *truth_cpu = 0;
if(state.truth){
truth_cpu = calloc(l.batch*l.outputs, sizeof(float));
cuda_pull_array(state.truth, truth_cpu, l.batch*l.outputs);
}
cuda_pull_array(state.input, in_cpu, l.batch*l.inputs);
network_state cpu_state;
cpu_state.train = state.train;
cpu_state.truth = truth_cpu;
cpu_state.input = in_cpu;
forward_region_layer(l, cpu_state);
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 forward_cost_layer_gpu(cost_layer l, network *net)
{
if (!net->truth) return;
if(l.smooth){
scal_gpu(l.batch*l.inputs, (1-l.smooth), net->truth_gpu, 1);
add_gpu(l.batch*l.inputs, l.smooth * 1./l.inputs, net->truth_gpu, 1);
}
if(l.cost_type == SMOOTH){
smooth_l1_gpu(l.batch*l.inputs, net->input_gpu, net->truth_gpu, l.delta_gpu, l.output_gpu);
} else if (l.cost_type == L1){
l1_gpu(l.batch*l.inputs, net->input_gpu, net->truth_gpu, l.delta_gpu, l.output_gpu);
} else if (l.cost_type == WGAN){
wgan_gpu(l.batch*l.inputs, net->input_gpu, net->truth_gpu, l.delta_gpu, l.output_gpu);
} else {
l2_gpu(l.batch*l.inputs, net->input_gpu, net->truth_gpu, l.delta_gpu, l.output_gpu);
}
if (l.cost_type == SEG && l.noobject_scale != 1) {
scale_mask_gpu(l.batch*l.inputs, l.delta_gpu, 0, net->truth_gpu, l.noobject_scale);
scale_mask_gpu(l.batch*l.inputs, l.output_gpu, 0, net->truth_gpu, l.noobject_scale);
}
if (l.cost_type == MASKED) {
mask_gpu(l.batch*l.inputs, net->delta_gpu, SECRET_NUM, net->truth_gpu, 0);
}
if(l.ratio){
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
qsort(l.delta, l.batch*l.inputs, sizeof(float), float_abs_compare);
int n = (1-l.ratio) * l.batch*l.inputs;
float thresh = l.delta[n];
thresh = 0;
printf("%f\n", thresh);
supp_gpu(l.batch*l.inputs, thresh, l.delta_gpu, 1);
}
if(l.thresh){
supp_gpu(l.batch*l.inputs, l.thresh*1./l.inputs, l.delta_gpu, 1);
}
cuda_pull_array(l.output_gpu, l.output, l.batch*l.inputs);
l.cost[0] = sum_array(l.output, l.batch*l.inputs);
}
void forward_region_layer_gpu(const region_layer l, network_state state)
{
/*
if(!state.train){
copy_ongpu(l.batch*l.inputs, state.input, 1, l.output_gpu, 1);
return;
}
*/
flatten_ongpu(state.input, l.h*l.w, l.n*(l.coords + l.classes + 1), l.batch, 1, l.output_gpu);
if(l.softmax_tree){
int i;
int count = 5;
for (i = 0; i < l.softmax_tree->groups; ++i) {
int group_size = l.softmax_tree->group_size[i];
softmax_gpu(l.output_gpu+count, group_size, l.classes + 5, l.w*l.h*l.n*l.batch, 1, l.output_gpu + count);
count += group_size;
}
}else if (l.softmax){
softmax_gpu(l.output_gpu+5, l.classes, l.classes + 5, l.w*l.h*l.n*l.batch, 1, l.output_gpu + 5);
}
float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
float *truth_cpu = 0;
if(state.truth){
int num_truth = l.batch*l.truths;
truth_cpu = calloc(num_truth, sizeof(float));
cuda_pull_array(state.truth, truth_cpu, num_truth);
}
cuda_pull_array(l.output_gpu, in_cpu, l.batch*l.inputs);
//cudaStreamSynchronize(get_cuda_stream());
network_state cpu_state = state;
cpu_state.train = state.train;
cpu_state.truth = truth_cpu;
cpu_state.input = in_cpu;
forward_region_layer(l, cpu_state);
//cuda_push_array(l.output_gpu, l.output, l.batch*l.outputs);
free(cpu_state.input);
if(!state.train) return;
cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
//cudaStreamSynchronize(get_cuda_stream());
if(cpu_state.truth) free(cpu_state.truth);
}
float cuda_compare(float *x_gpu, float *x, size_t n, char *s) {
float *tmp = calloc(n, sizeof(float));
cuda_pull_array(x_gpu, tmp, n);
//int i;
//for(i = 0; i < n; ++i) printf("%f %f\n", tmp[i], x[i]);
axpy_cpu(n, -1, x, 1, tmp, 1);
float err = dot_cpu(n, tmp, 1, tmp, 1);
printf("Error %s: %f\n", s, sqrt(err / n));
free(tmp);
return err;
}
image get_network_image_layer(network net, int i)
{
layer l = net.layers[i];
#ifdef GPU
cuda_pull_array(l.output_gpu, l.output, l.outputs);
#endif
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);
}
image def = {0};
return def;
}
void forward_cost_layer_gpu(cost_layer l, network_state state)
{
if (!state.truth) return;
if (l.cost_type == MASKED) {
mask_ongpu(l.batch*l.inputs, state.input, state.truth);
}
copy_ongpu(l.batch*l.inputs, state.truth, 1, l.delta_gpu, 1);
axpy_ongpu(l.batch*l.inputs, -1, state.input, 1, l.delta_gpu, 1);
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
*(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
}
void forward_iseg_layer_gpu(const layer l, network net) {
copy_gpu(l.batch * l.inputs, net.input_gpu, 1, l.output_gpu, 1, net.st);
int b;
for (b = 0; b < l.batch; ++b) {
activate_array_gpu(l.output_gpu + b * l.outputs, l.classes * l.w * l.h,
LOGISTIC, net.st);
//if(l.extra) activate_array_gpu(l.output_gpu + b*l.outputs + l.classes*l.w*l.h, l.extra*l.w*l.h, LOGISTIC);
}
cuda_pull_array(l.output_gpu, net.input, l.batch * l.inputs);
forward_iseg_layer(l, net);
cuda_push_array(l.delta_gpu, l.delta, l.batch * l.outputs);
}
void pull_connected_layer(connected_layer l)
{
cuda_pull_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
if (l.batch_normalize){
cuda_pull_array(l.scales_gpu, l.scales, l.outputs);
cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
}
}
void forward_cost_layer_gpu(cost_layer l, network_state state)
{
if (!state.truth) return;
if(l.smooth){
scal_ongpu(l.batch*l.inputs, (1-l.smooth), state.truth, 1);
add_ongpu(l.batch*l.inputs, l.smooth * 1./l.inputs, state.truth, 1);
}
if (l.cost_type == MASKED) {
mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth);
}
if(l.cost_type == SMOOTH){
smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
} else if (l.cost_type == L1){
l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
} else {
l2_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
}
if(l.ratio){
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
qsort(l.delta, l.batch*l.inputs, sizeof(float), float_abs_compare);
int n = (1-l.ratio) * l.batch*l.inputs;
float thresh = l.delta[n];
thresh = 0;
printf("%f\n", thresh);
supp_ongpu(l.batch*l.inputs, thresh, l.delta_gpu, 1);
}
if(l.thresh){
supp_ongpu(l.batch*l.inputs, l.thresh*1./l.inputs, l.delta_gpu, 1);
}
cuda_pull_array(l.output_gpu, l.output, l.batch*l.inputs);
l.cost[0] = sum_array(l.output, l.batch*l.inputs);
}
void vec_char_rnn(char *cfgfile, char *weightfile, char *seed)
{
char *base = basecfg(cfgfile);
fprintf(stderr, "%s\n", base);
network net = parse_network_cfg(cfgfile);
if(weightfile){
load_weights(&net, weightfile);
}
int inputs = get_network_input_size(net);
int c;
int seed_len = strlen(seed);
float *input = calloc(inputs, sizeof(float));
int i;
char *line;
while((line=fgetl(stdin)) != 0){
reset_rnn_state(net, 0);
for(i = 0; i < seed_len; ++i){
c = seed[i];
input[(int)c] = 1;
network_predict(net, input);
input[(int)c] = 0;
}
strip(line);
int str_len = strlen(line);
for(i = 0; i < str_len; ++i){
c = line[i];
input[(int)c] = 1;
network_predict(net, input);
input[(int)c] = 0;
}
c = ' ';
input[(int)c] = 1;
network_predict(net, input);
input[(int)c] = 0;
layer l = net.layers[0];
#ifdef GPU
cuda_pull_array(l.output_gpu, l.output, l.outputs);
#endif
printf("%s", line);
for(i = 0; i < l.outputs; ++i){
printf(",%g", l.output[i]);
}
printf("\n");
}
}
void forward_cost_layer_gpu(cost_layer l, network_state state)
{
if (!state.truth) return;
if (l.cost_type == MASKED) {
mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth);
}
if(l.cost_type == SMOOTH){
smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu);
} else {
copy_ongpu(l.batch*l.inputs, state.truth, 1, l.delta_gpu, 1);
axpy_ongpu(l.batch*l.inputs, -1, state.input, 1, l.delta_gpu, 1);
}
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
*(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1);
}
void gemm_gpu(int TA, int TB, int M, int N, int K, float ALPHA,
float *A, int lda,
float *B, int ldb,
float BETA,
float *C, int ldc)
{
float *A_gpu = cuda_make_array(A, (TA ? lda*K:lda*M));
float *B_gpu = cuda_make_array(B, (TB ? ldb*N : ldb*K));
float *C_gpu = cuda_make_array(C, ldc*M);
gemm_ongpu(TA, TB, M, N, K, ALPHA, A_gpu, lda, B_gpu, ldb, BETA, C_gpu, ldc);
cuda_pull_array(C_gpu, C, ldc*M);
cuda_free(A_gpu);
cuda_free(B_gpu);
cuda_free(C_gpu);
}
void reconstruct_picture(network net, float *features, image recon, image update, float rate, float momentum, float lambda, int smooth_size, int iters)
{
int iter = 0;
for (iter = 0; iter < iters; ++iter) {
image delta = make_image(recon.w, recon.h, recon.c);
NETWORK_STATE(state);
#ifdef GPU
state.input = cuda_make_array(recon.data, recon.w*recon.h*recon.c);
state.delta = cuda_make_array(delta.data, delta.w*delta.h*delta.c);
state.truth = cuda_make_array(features, get_network_output_size(net));
forward_network_gpu(net, state);
backward_network_gpu(net, state);
cuda_pull_array(state.delta, delta.data, delta.w*delta.h*delta.c);
cuda_free(state.input);
cuda_free(state.delta);
cuda_free(state.truth);
#else
state.input = recon.data;
state.delta = delta.data;
state.truth = features;
forward_network(net, state);
backward_network(net, state);
#endif
fltadd(update.data, delta.data, recon.w * recon.h * recon.c);
smooth(recon, update, lambda, smooth_size);
fltaddmul(recon.data, update.data, recon.w * recon.h * recon.c, rate);
scal_cpu(recon.w*recon.h*recon.c, momentum, update.data, 1);
//float mag = mag_array(recon.data, recon.w*recon.h*recon.c);
//scal_cpu(recon.w*recon.h*recon.c, 600/mag, recon.data, 1);
constrain_image(recon);
free_image(delta);
}
}
void reconstruct_picture(network net, float *features, image recon, image update, float rate, float momentum, float lambda, int smooth_size)
{
scale_image(recon, 2);
translate_image(recon, -1);
image delta = make_image(recon.w, recon.h, recon.c);
network_state state = {0};
#ifdef GPU
state.input = cuda_make_array(recon.data, recon.w*recon.h*recon.c);
state.delta = cuda_make_array(delta.data, delta.w*delta.h*delta.c);
state.truth = cuda_make_array(features, get_network_output_size(net));
forward_network_gpu(net, state);
backward_network_gpu(net, state);
cuda_pull_array(state.delta, delta.data, delta.w*delta.h*delta.c);
cuda_free(state.input);
cuda_free(state.delta);
cuda_free(state.truth);
#else
state.input = recon.data;
state.delta = delta.data;
state.truth = features;
forward_network(net, state);
backward_network(net, state);
#endif
axpy_cpu(recon.w*recon.h*recon.c, 1, delta.data, 1, update.data, 1);
smooth(recon, update, lambda, smooth_size);
axpy_cpu(recon.w*recon.h*recon.c, rate, update.data, 1, recon.data, 1);
scal_cpu(recon.w*recon.h*recon.c, momentum, update.data, 1);
translate_image(recon, 1);
scale_image(recon, .5);
constrain_image(recon);
free_image(delta);
}
void pull_softmax_layer_output(const softmax_layer layer)
{
cuda_pull_array(layer.output_gpu, layer.output, layer.inputs*layer.batch);
}
void optimize_picture(network *net, image orig, int max_layer, float scale, float rate, float thresh, int norm)
{
//scale_image(orig, 2);
//translate_image(orig, -1);
net->n = max_layer + 1;
int dx = rand()%16 - 8;
int dy = rand()%16 - 8;
int flip = rand()%2;
image crop = crop_image(orig, dx, dy, orig.w, orig.h);
image im = resize_image(crop, (int)(orig.w * scale), (int)(orig.h * scale));
if(flip) flip_image(im);
resize_network(net, im.w, im.h);
layer_t last = net->layers[net->n-1];
//net->layers[net->n - 1].activation = LINEAR;
image delta = make_image(im.w, im.h, im.c);
NETWORK_STATE(state);
#ifdef GPU
state.input = cuda_make_array(im.data, im.w*im.h*im.c);
state.delta = cuda_make_array(im.data, im.w*im.h*im.c);
forward_network_gpu(*net, state);
copy_ongpu(last.outputs, last.output_gpu, 1, last.delta_gpu, 1);
cuda_pull_array(last.delta_gpu, last.delta, last.outputs);
calculate_loss(last.delta, last.delta, last.outputs, thresh);
cuda_push_array(last.delta_gpu, last.delta, last.outputs);
backward_network_gpu(*net, state);
cuda_pull_array(state.delta, delta.data, im.w*im.h*im.c);
cuda_free(state.input);
cuda_free(state.delta);
#else
state.input = im.data;
state.delta = delta.data;
forward_network(*net, state);
fltcpy(last.delta, last.output, last.outputs);
calculate_loss(last.output, last.delta, last.outputs, thresh);
backward_network(*net, state);
#endif
if(flip) flip_image(delta);
//normalize_array(delta.data, delta.w*delta.h*delta.c);
image resized = resize_image(delta, orig.w, orig.h);
image out = crop_image(resized, -dx, -dy, orig.w, orig.h);
/*
image g = grayscale_image(out);
free_image(out);
out = g;
*/
//rate = rate / abs_mean(out.data, out.w*out.h*out.c);
if(norm) normalize_array(out.data, out.w*out.h*out.c);
fltaddmul(orig.data, out.data, orig.w * orig.h * orig.c, rate);
/*
normalize_array(orig.data, orig.w*orig.h*orig.c);
scale_image(orig, sqrt(var));
translate_image(orig, mean);
*/
//translate_image(orig, 1);
//scale_image(orig, .5);
//normalize_image(orig);
constrain_image(orig);
free_image(crop);
free_image(im);
free_image(delta);
free_image(resized);
free_image(out);
}
void pull_network_output(network net)
{
layer l = get_network_output_layer(net);
cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
}
void train_lsd3(char *fcfg, char *fweight, char *gcfg, char *gweight, char *acfg, char *aweight, int clear)
{
#ifdef GPU
//char *train_images = "/home/pjreddie/data/coco/trainvalno5k.txt";
char *train_images = "/home/pjreddie/data/imagenet/imagenet1k.train.list";
//char *style_images = "/home/pjreddie/data/coco/trainvalno5k.txt";
char *style_images = "/home/pjreddie/zelda.txt";
char *backup_directory = "/home/pjreddie/backup/";
srand(time(0));
network fnet = load_network(fcfg, fweight, clear);
network gnet = load_network(gcfg, gweight, clear);
network anet = load_network(acfg, aweight, clear);
char *gbase = basecfg(gcfg);
char *abase = basecfg(acfg);
printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", gnet.learning_rate, gnet.momentum, gnet.decay);
int imgs = gnet.batch*gnet.subdivisions;
int i = *gnet.seen/imgs;
data train, tbuffer;
data style, sbuffer;
list *slist = get_paths(style_images);
char **spaths = (char **)list_to_array(slist);
list *tlist = get_paths(train_images);
char **tpaths = (char **)list_to_array(tlist);
load_args targs= get_base_args(gnet);
targs.paths = tpaths;
targs.n = imgs;
targs.m = tlist->size;
targs.d = &tbuffer;
targs.type = CLASSIFICATION_DATA;
targs.classes = 1;
char *ls[1] = {"zelda"};
targs.labels = ls;
load_args sargs = get_base_args(gnet);
sargs.paths = spaths;
sargs.n = imgs;
sargs.m = slist->size;
sargs.d = &sbuffer;
sargs.type = CLASSIFICATION_DATA;
sargs.classes = 1;
sargs.labels = ls;
pthread_t tload_thread = load_data_in_thread(targs);
pthread_t sload_thread = load_data_in_thread(sargs);
clock_t time;
float aloss_avg = -1;
float floss_avg = -1;
network_state fstate = {};
fstate.index = 0;
fstate.net = fnet;
int x_size = get_network_input_size(fnet)*fnet.batch;
int y_size = get_network_output_size(fnet)*fnet.batch;
fstate.input = cuda_make_array(0, x_size);
fstate.truth = cuda_make_array(0, y_size);
fstate.delta = cuda_make_array(0, x_size);
fstate.train = 1;
float *X = (float*)calloc(x_size, sizeof(float));
float *y = (float*)calloc(y_size, sizeof(float));
float *ones = cuda_make_array(0, anet.batch);
float *zeros = cuda_make_array(0, anet.batch);
fill_ongpu(anet.batch, .99, ones, 1);
fill_ongpu(anet.batch, .01, zeros, 1);
network_state astate = {};
astate.index = 0;
astate.net = anet;
int ax_size = get_network_input_size(anet)*anet.batch;
int ay_size = get_network_output_size(anet)*anet.batch;
astate.input = 0;
astate.truth = ones;
astate.delta = cuda_make_array(0, ax_size);
astate.train = 1;
network_state gstate = {};
gstate.index = 0;
gstate.net = gnet;
int gx_size = get_network_input_size(gnet)*gnet.batch;
int gy_size = get_network_output_size(gnet)*gnet.batch;
gstate.input = cuda_make_array(0, gx_size);
gstate.truth = 0;
gstate.delta = 0;
gstate.train = 1;
while (get_current_batch(gnet) < gnet.max_batches) {
i += 1;
time=clock();
pthread_join(tload_thread, 0);
pthread_join(sload_thread, 0);
train = tbuffer;
style = sbuffer;
tload_thread = load_data_in_thread(targs);
sload_thread = load_data_in_thread(sargs);
printf("Loaded: %lf seconds\n", sec(clock()-time));
data generated = copy_data(train);
time=clock();
int j, k;
float floss = 0;
for(j = 0; j < fnet.subdivisions; ++j){
layer imlayer = gnet.layers[gnet.n - 1];
get_next_batch(train, fnet.batch, j*fnet.batch, X, y);
cuda_push_array(fstate.input, X, x_size);
cuda_push_array(gstate.input, X, gx_size);
*gnet.seen += gnet.batch;
forward_network_gpu(fnet, fstate);
float *feats = fnet.layers[fnet.n - 2].output_gpu;
copy_ongpu(y_size, feats, 1, fstate.truth, 1);
forward_network_gpu(gnet, gstate);
float *gen = gnet.layers[gnet.n-1].output_gpu;
copy_ongpu(x_size, gen, 1, fstate.input, 1);
fill_ongpu(x_size, 0, fstate.delta, 1);
forward_network_gpu(fnet, fstate);
backward_network_gpu(fnet, fstate);
//HERE
astate.input = gen;
fill_ongpu(ax_size, 0, astate.delta, 1);
forward_network_gpu(anet, astate);
backward_network_gpu(anet, astate);
float *delta = imlayer.delta_gpu;
fill_ongpu(x_size, 0, delta, 1);
scal_ongpu(x_size, 100, astate.delta, 1);
scal_ongpu(x_size, .00001, fstate.delta, 1);
axpy_ongpu(x_size, 1, fstate.delta, 1, delta, 1);
axpy_ongpu(x_size, 1, astate.delta, 1, delta, 1);
//fill_ongpu(x_size, 0, delta, 1);
//cuda_push_array(delta, X, x_size);
//axpy_ongpu(x_size, -1, imlayer.output_gpu, 1, delta, 1);
//printf("pix error: %f\n", cuda_mag_array(delta, x_size));
printf("fea error: %f\n", cuda_mag_array(fstate.delta, x_size));
printf("adv error: %f\n", cuda_mag_array(astate.delta, x_size));
//axpy_ongpu(x_size, 1, astate.delta, 1, delta, 1);
backward_network_gpu(gnet, gstate);
floss += get_network_cost(fnet) /(fnet.subdivisions*fnet.batch);
cuda_pull_array(imlayer.output_gpu, imlayer.output, x_size);
for(k = 0; k < gnet.batch; ++k){
int index = j*gnet.batch + k;
copy_cpu(imlayer.outputs, imlayer.output + k*imlayer.outputs, 1, generated.X.vals[index], 1);
generated.y.vals[index][0] = .01;
}
}
/*
image sim = float_to_image(anet.w, anet.h, anet.c, style.X.vals[j]);
show_image(sim, "style");
cvWaitKey(0);
*/
harmless_update_network_gpu(anet);
data merge = concat_data(style, generated);
randomize_data(merge);
float aloss = train_network(anet, merge);
update_network_gpu(gnet);
free_data(merge);
free_data(train);
free_data(generated);
free_data(style);
if (aloss_avg < 0) aloss_avg = aloss;
if (floss_avg < 0) floss_avg = floss;
aloss_avg = aloss_avg*.9 + aloss*.1;
floss_avg = floss_avg*.9 + floss*.1;
printf("%d: gen: %f, adv: %f | gen_avg: %f, adv_avg: %f, %f rate, %lf seconds, %d images\n", i, floss, aloss, floss_avg, aloss_avg, get_current_rate(gnet), sec(clock()-time), i*imgs);
if(i%1000==0){
char buff[256];
sprintf(buff, "%s/%s_%d.weights", backup_directory, gbase, i);
save_weights(gnet, buff);
sprintf(buff, "%s/%s_%d.weights", backup_directory, abase, i);
save_weights(anet, buff);
}
if(i%100==0){
char buff[256];
sprintf(buff, "%s/%s.backup", backup_directory, gbase);
save_weights(gnet, buff);
sprintf(buff, "%s/%s.backup", backup_directory, abase);
save_weights(anet, buff);
}
}
#endif
}
void train_lsd2(char *cfgfile, char *weightfile, char *acfgfile, char *aweightfile, int clear)
{
#ifdef GPU
char *train_images = "/home/pjreddie/data/coco/trainvalno5k.txt";
char *backup_directory = "/home/pjreddie/backup/";
srand(time(0));
char *base = basecfg(cfgfile);
printf("%s\n", base);
network net = parse_network_cfg(cfgfile);
if(weightfile){
load_weights(&net, weightfile);
}
if(clear) *net.seen = 0;
char *abase = basecfg(acfgfile);
network anet = parse_network_cfg(acfgfile);
if(aweightfile){
load_weights(&anet, aweightfile);
}
if(clear) *anet.seen = 0;
int i, j, k;
layer imlayer = {};
for (i = 0; i < net.n; ++i) {
if (net.layers[i].out_c == 3) {
imlayer = net.layers[i];
break;
}
}
printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
int imgs = net.batch*net.subdivisions;
i = *net.seen/imgs;
data train, buffer;
list *plist = get_paths(train_images);
//int N = plist->size;
char **paths = (char **)list_to_array(plist);
load_args args = {};
args.w = net.w;
args.h = net.h;
args.paths = paths;
args.n = imgs;
args.m = plist->size;
args.d = &buffer;
args.min = net.min_crop;
args.max = net.max_crop;
args.angle = net.angle;
args.aspect = net.aspect;
args.exposure = net.exposure;
args.saturation = net.saturation;
args.hue = net.hue;
args.size = net.w;
args.type = CLASSIFICATION_DATA;
args.classes = 1;
char *ls[1] = {"coco"};
args.labels = ls;
pthread_t load_thread = load_data_in_thread(args);
clock_t time;
network_state gstate = {};
gstate.index = 0;
gstate.net = net;
int x_size = get_network_input_size(net)*net.batch;
int y_size = 1*net.batch;
gstate.input = cuda_make_array(0, x_size);
gstate.truth = 0;
gstate.delta = 0;
gstate.train = 1;
float *X = (float*)calloc(x_size, sizeof(float));
float *y = (float*)calloc(y_size, sizeof(float));
network_state astate = {};
astate.index = 0;
astate.net = anet;
int ay_size = get_network_output_size(anet)*anet.batch;
astate.input = 0;
astate.truth = 0;
astate.delta = 0;
astate.train = 1;
float *imerror = cuda_make_array(0, imlayer.outputs);
float *ones_gpu = cuda_make_array(0, ay_size);
fill_ongpu(ay_size, 1, ones_gpu, 1);
float aloss_avg = -1;
float gloss_avg = -1;
//data generated = copy_data(train);
while (get_current_batch(net) < net.max_batches) {
i += 1;
time=clock();
pthread_join(load_thread, 0);
train = buffer;
load_thread = load_data_in_thread(args);
printf("Loaded: %lf seconds\n", sec(clock()-time));
data generated = copy_data(train);
time=clock();
float gloss = 0;
for(j = 0; j < net.subdivisions; ++j){
get_next_batch(train, net.batch, j*net.batch, X, y);
cuda_push_array(gstate.input, X, x_size);
*net.seen += net.batch;
forward_network_gpu(net, gstate);
fill_ongpu(imlayer.outputs, 0, imerror, 1);
astate.input = imlayer.output_gpu;
astate.delta = imerror;
astate.truth = ones_gpu;
forward_network_gpu(anet, astate);
backward_network_gpu(anet, astate);
scal_ongpu(imlayer.outputs, 1, imerror, 1);
axpy_ongpu(imlayer.outputs, 1, imerror, 1, imlayer.delta_gpu, 1);
backward_network_gpu(net, gstate);
printf("features %f\n", cuda_mag_array(imlayer.delta_gpu, imlayer.outputs));
printf("realness %f\n", cuda_mag_array(imerror, imlayer.outputs));
gloss += get_network_cost(net) /(net.subdivisions*net.batch);
cuda_pull_array(imlayer.output_gpu, imlayer.output, x_size);
for(k = 0; k < net.batch; ++k){
int index = j*net.batch + k;
copy_cpu(imlayer.outputs, imlayer.output + k*imlayer.outputs, 1, generated.X.vals[index], 1);
generated.y.vals[index][0] = 0;
}
}
harmless_update_network_gpu(anet);
data merge = concat_data(train, generated);
randomize_data(merge);
float aloss = train_network(anet, merge);
update_network_gpu(net);
update_network_gpu(anet);
free_data(merge);
free_data(train);
free_data(generated);
if (aloss_avg < 0) aloss_avg = aloss;
aloss_avg = aloss_avg*.9 + aloss*.1;
gloss_avg = gloss_avg*.9 + gloss*.1;
printf("%d: gen: %f, adv: %f | gen_avg: %f, adv_avg: %f, %f rate, %lf seconds, %d images\n", i, gloss, aloss, gloss_avg, aloss_avg, get_current_rate(net), sec(clock()-time), i*imgs);
if(i%1000==0){
char buff[256];
sprintf(buff, "%s/%s_%d.weights", backup_directory, base, i);
save_weights(net, buff);
sprintf(buff, "%s/%s_%d.weights", backup_directory, abase, i);
save_weights(anet, buff);
}
if(i%100==0){
char buff[256];
sprintf(buff, "%s/%s.backup", backup_directory, base);
save_weights(net, buff);
sprintf(buff, "%s/%s.backup", backup_directory, abase);
save_weights(anet, buff);
}
}
char buff[256];
sprintf(buff, "%s/%s_final.weights", backup_directory, base);
save_weights(net, buff);
#endif
}
void pull_batchnorm_layer(layer l)
{
cuda_pull_array(l.scales_gpu, l.scales, l.c);
cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.c);
cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.c);
}
float * ofxDarknet::get_network_output_layer_gpu(int i)
{
layer l = net.layers[i];
if(l.type != REGION) cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
return l.output;
}
void pull_cost_layer(cost_layer l)
{
cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
}
void try_classifier(char *datacfg, char *cfgfile, char *weightfile, char *filename, int layer_num)
{
network *net = load_network(cfgfile, weightfile, 0);
set_batch_network(net, 1);
srand(2222222);
list *options = read_data_cfg(datacfg);
char *name_list = option_find_str(options, "names", 0);
if(!name_list) name_list = option_find_str(options, "labels", "data/labels.list");
int top = option_find_int(options, "top", 1);
int i = 0;
char **names = get_labels(name_list);
clock_t time;
int *indexes = calloc(top, sizeof(int));
char buff[256];
char *input = buff;
while(1){
if(filename){
strncpy(input, filename, 256);
}else{
printf("Enter Image Path: ");
fflush(stdout);
input = fgets(input, 256, stdin);
if(!input) return;
strtok(input, "\n");
}
image orig = load_image_color(input, 0, 0);
image r = resize_min(orig, 256);
image im = crop_image(r, (r.w - 224 - 1)/2 + 1, (r.h - 224 - 1)/2 + 1, 224, 224);
float mean[] = {0.48263312050943, 0.45230225481413, 0.40099074308742};
float std[] = {0.22590347483426, 0.22120921437787, 0.22103996251583};
float var[3];
var[0] = std[0]*std[0];
var[1] = std[1]*std[1];
var[2] = std[2]*std[2];
normalize_cpu(im.data, mean, var, 1, 3, im.w*im.h);
float *X = im.data;
time=clock();
float *predictions = network_predict(net, X);
layer l = net->layers[layer_num];
for(i = 0; i < l.c; ++i){
if(l.rolling_mean) printf("%f %f %f\n", l.rolling_mean[i], l.rolling_variance[i], l.scales[i]);
}
#ifdef GPU
cuda_pull_array(l.output_gpu, l.output, l.outputs);
#endif
for(i = 0; i < l.outputs; ++i){
printf("%f\n", l.output[i]);
}
/*
printf("\n\nWeights\n");
for(i = 0; i < l.n*l.size*l.size*l.c; ++i){
printf("%f\n", l.filters[i]);
}
printf("\n\nBiases\n");
for(i = 0; i < l.n; ++i){
printf("%f\n", l.biases[i]);
}
*/
top_predictions(net, top, indexes);
printf("%s: Predicted in %f seconds.\n", input, sec(clock()-time));
for(i = 0; i < top; ++i){
int index = indexes[i];
printf("%s: %f\n", names[index], predictions[index]);
}
free_image(im);
if (filename) break;
}
}
