| cfg/detection.cfg | ●●●●● patch | view | raw | blame | history | |
| cfg/rescore.cfg | ●●●●● patch | view | raw | blame | history | |
| src/connected_layer.c | ●●●●● patch | view | raw | blame | history | |
| src/convolutional_layer.c | ●●●●● patch | view | raw | blame | history | |
| src/data.c | ●●●●● patch | view | raw | blame | history | |
| src/detection.c | ●●●●● patch | view | raw | blame | history | |
| src/detection_layer.c | ●●●●● patch | view | raw | blame | history |
cfg/detection.cfg
New file @@ -0,0 +1,197 @@ [net] batch=64 subdivisions=4 height=448 width=448 channels=3 learning_rate=0.01 momentum=0.9 decay=0.0005 seen = 0 [crop] crop_width=448 crop_height=448 flip=0 angle=0 saturation = 2 exposure = 2 [convolutional] filters=64 size=7 stride=2 pad=1 activation=ramp [convolutional] filters=192 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=128 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=128 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=128 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=1024 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=512 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=1024 size=3 stride=1 pad=1 activation=ramp [convolutional] size=3 stride=1 pad=1 filters=1024 activation=ramp [convolutional] size=3 stride=2 pad=1 filters=1024 activation=ramp [convolutional] size=3 stride=1 pad=1 filters=1024 activation=ramp [connected] output=4096 activation=ramp [dropout] probability=.5 [connected] output=1225 activation=logistic [detection] classes=20 coords=4 rescore=0 nuisance = 1 background=1 cfg/rescore.cfg
New file @@ -0,0 +1,198 @@ [net] batch=64 subdivisions=4 height=448 width=448 channels=3 learning_rate=0.01 momentum=0.9 decay=0.0005 seen = 0 [crop] crop_width=448 crop_height=448 flip=0 angle=0 saturation = 2 exposure = 2 [convolutional] filters=64 size=7 stride=2 pad=1 activation=ramp [convolutional] filters=192 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=128 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=128 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=128 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=512 size=3 stride=1 pad=1 activation=ramp [convolutional] filters=256 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=1024 size=3 stride=2 pad=1 activation=ramp [convolutional] filters=512 size=1 stride=1 pad=1 activation=ramp [convolutional] filters=1024 size=3 stride=1 pad=1 activation=ramp [convolutional] size=3 stride=1 pad=1 filters=1024 activation=ramp [convolutional] size=3 stride=2 pad=1 filters=1024 activation=ramp [convolutional] size=3 stride=1 pad=1 filters=1024 activation=ramp [connected] output=4096 activation=ramp [dropout] probability=.5 [connected] output=1225 activation=logistic [detection] classes=20 coords=4 rescore=1 nuisance = 0 background=0
@@ -29,7 +29,8 @@ l.biases = calloc(outputs, sizeof(float)); float scale = 1./sqrt(inputs); //float scale = 1./sqrt(inputs); float scale = sqrt(2./inputs); for(i = 0; i < inputs*outputs; ++i){ l.weights[i] = 2*scale*rand_uniform() - scale; } src/convolutional_layer.c
@@ -61,7 +61,8 @@ l.biases = calloc(n, sizeof(float)); l.bias_updates = calloc(n, sizeof(float)); float scale = 1./sqrt(size*size*c); //float scale = 1./sqrt(size*size*c); float scale = sqrt(2./(size*size*c)); for(i = 0; i < c*n*size*size; ++i) l.filters[i] = 2*scale*rand_uniform() - scale; for(i = 0; i < n; ++i){ l.biases[i] = scale; src/data.c
@@ -174,7 +174,7 @@ } int index = (i+j*num_boxes)*(4+classes+background); //if(truth[index+classes+background+2]) continue; if(truth[index+classes+background+2]) continue; if(background) truth[index++] = 0; truth[index+id] = 1; index += classes; src/detection.c
@@ -47,6 +47,8 @@ int top = (y-h/2)*im.h; int bot = (y+h/2)*im.h; draw_box(im, left, top, right, bot, red, green, blue); draw_box(im, left+1, top+1, right+1, bot+1, red, green, blue); draw_box(im, left-1, top-1, right-1, bot-1, red, green, blue); } } } @@ -116,7 +118,11 @@ float loss = train_network(net, train); //TODO #ifdef GPU float *out = get_network_output_gpu(net); #else float *out = get_network_output(net); #endif image im = float_to_image(net.w, net.h, 3, train.X.vals[127]); image copy = copy_image(im); draw_localization(copy, &(out[63*80])); @@ -213,7 +219,7 @@ avg_loss = avg_loss*.9 + loss*.1; printf("%d: %f, %f avg, %lf seconds, %d images\n", i, loss, avg_loss, sec(clock()-time), i*imgs); if(i == 100){ net.learning_rate *= 10; //net.learning_rate *= 10; } if(i%100==0){ char buff[256]; @@ -309,8 +315,8 @@ float y = (pred.vals[j][ci + 1] + row)/num_boxes; float w = pred.vals[j][ci + 2]; //* distance_from_edge(row, num_boxes); float h = pred.vals[j][ci + 3]; //* distance_from_edge(col, num_boxes); w = pow(w, 1); h = pow(h, 1); w = pow(w, 2); h = pow(h, 2); float prob = scale*pred.vals[j][k+class+background+nuisance]; if(prob < threshold) continue; printf("%d %d %f %f %f %f %f\n", offset + j, class, prob, x, y, w, h); src/detection_layer.c
@@ -330,8 +330,9 @@ l.output[out_i++] = mask*state.input[in_i++]; } } float avg_iou = 0; int count = 0; if(l.does_cost && state.train){ int count = 0; *(l.cost) = 0; int size = get_detection_layer_output_size(l) * l.batch; memset(l.delta, 0, size * sizeof(float)); @@ -342,65 +343,54 @@ *(l.cost) += pow(state.truth[j] - l.output[j], 2); l.delta[j] = state.truth[j] - l.output[j]; } box truth; truth.x = state.truth[j+0]; truth.y = state.truth[j+1]; truth.w = state.truth[j+2]; truth.h = state.truth[j+3]; 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]; out.y = l.output[j+1]; out.w = l.output[j+2]; out.h = l.output[j+3]; 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; l.delta[j+0] = (truth.x - out.x); l.delta[j+1] = (truth.y - out.y); l.delta[j+2] = (truth.w - out.w); l.delta[j+3] = (truth.h - out.h); *(l.cost) += pow((out.x - truth.x), 2); *(l.cost) += pow((out.y - truth.y), 2); *(l.cost) += pow((out.w - truth.w), 2); *(l.cost) += pow((out.h - truth.h), 2); /* l.delta[j+0] = .1 * (truth.x - out.x) / (49 * truth.w * truth.w); l.delta[j+1] = .1 * (truth.y - out.y) / (49 * truth.h * truth.h); l.delta[j+2] = .1 * (truth.w - out.w) / ( truth.w * truth.w); l.delta[j+3] = .1 * (truth.h - out.h) / ( truth.h * truth.h); *(l.cost) += pow((out.x - truth.x)/truth.w/7., 2); *(l.cost) += pow((out.y - truth.y)/truth.h/7., 2); *(l.cost) += pow((out.w - truth.w)/truth.w, 2); *(l.cost) += pow((out.h - truth.h)/truth.h, 2); */ float iou = box_iou(out, truth); avg_iou += iou; ++count; dbox delta = diou(out, truth); l.delta[j+0] = 10 * delta.dx/7; l.delta[j+1] = 10 * delta.dy/7; l.delta[j+2] = 10 * delta.dw * 2 * sqrt(out.w); l.delta[j+3] = 10 * delta.dh * 2 * sqrt(out.h); *(l.cost) += pow((1-iou), 2); if(0){ l.delta[j+0] = (state.truth[j+0] - l.output[j+0]); l.delta[j+1] = (state.truth[j+1] - l.output[j+1]); l.delta[j+2] = (state.truth[j+2] - l.output[j+2]); l.delta[j+3] = (state.truth[j+3] - l.output[j+3]); }else{ l.delta[j+0] = 4 * (state.truth[j+0] - l.output[j+0]) / 7; l.delta[j+1] = 4 * (state.truth[j+1] - l.output[j+1]) / 7; 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(0){ for (j = offset; j < offset+classes; ++j) { if(state.truth[j]) state.truth[j] = iou; l.delta[j] = state.truth[j] - l.output[j]; } } /* */ } printf("Avg IOU: %f\n", avg_iou/count); } /* int count = 0; for(i = 0; i < l.batch*locations; ++i){ for(j = 0; j < l.classes+l.background; ++j){ printf("%f, ", l.output[count++]); } printf("\n"); for(j = 0; j < l.coords; ++j){ printf("%f, ", l.output[count++]); } printf("\n"); } */ /* if(l.background || 1){ for(i = 0; i < l.batch*locations; ++i){ int index = i*(l.classes+l.coords+l.background); for(j= 0; j < l.classes; ++j){ if(state.truth[index+j+l.background]){ //dark_zone(l, j, index, state); } } } } */ } void backward_detection_layer(const detection_layer l, network_state state)
