diff --git a/Makefile b/Makefile index 9c3043b0..877fc7f0 100644 --- a/Makefile +++ b/Makefile @@ -1,6 +1,6 @@ CC=gcc GPU=0 -COMMON=-Wall -Werror -Wfatal-errors `pkg-config --cflags opencv` -I/usr/local/cuda/include/ +COMMON=-Wall -Wfatal-errors `pkg-config --cflags opencv` -I/usr/local/cuda/include/ ifeq ($(GPU), 1) COMMON+=-DGPU else @@ -19,13 +19,13 @@ LDFLAGS= -lOpenCL endif endif CFLAGS= $(COMMON) $(OPTS) -#CFLAGS= $(COMMON) -O0 -g +#CFLAGS= $(COMMON) -O0 -g LDFLAGS+=`pkg-config --libs opencv` -lm VPATH=./src/ EXEC=cnn OBJDIR=./obj/ -OBJ=network.o image.o cnn.o connected_layer.o maxpool_layer.o activations.o list.o option_list.o parser.o utils.o data.o matrix.o softmax_layer.o mini_blas.o convolutional_layer.o gemm.o normalization_layer.o opencl.o im2col.o col2im.o axpy.o +OBJ=network.o image.o cnn.o connected_layer.o maxpool_layer.o activations.o list.o option_list.o parser.o utils.o data.o matrix.o softmax_layer.o mini_blas.o convolutional_layer.o gemm.o normalization_layer.o opencl.o im2col.o col2im.o axpy.o dropout_layer.o OBJS = $(addprefix $(OBJDIR), $(OBJ)) all: $(EXEC) diff --git a/src/activations.c b/src/activations.c index 3b117166..04b27c92 100644 --- a/src/activations.c +++ b/src/activations.c @@ -41,29 +41,28 @@ float relu_activate(float x){return x*(x>0);} float ramp_activate(float x){return x*(x>0)+.1*x;} float tanh_activate(float x){return (exp(2*x)-1)/(exp(2*x)+1);} -float activate(float x, ACTIVATION a, float dropout) +float activate(float x, ACTIVATION a) { - if(dropout && (float)rand()/RAND_MAX < dropout) return 0; switch(a){ case LINEAR: - return linear_activate(x)/(1-dropout); + return linear_activate(x); case SIGMOID: - return sigmoid_activate(x)/(1-dropout); + return sigmoid_activate(x); case RELU: - return relu_activate(x)/(1-dropout); + return relu_activate(x); case RAMP: - return ramp_activate(x)/(1-dropout); + return ramp_activate(x); case TANH: - return tanh_activate(x)/(1-dropout); + return tanh_activate(x); } return 0; } -void activate_array(float *x, const int n, const ACTIVATION a, float dropout) +void activate_array(float *x, const int n, const ACTIVATION a) { int i; for(i = 0; i < n; ++i){ - x[i] = activate(x[i], a, dropout); + x[i] = activate(x[i], a); } } @@ -109,7 +108,7 @@ cl_kernel get_activation_kernel() } -void activate_array_ongpu(cl_mem x, int n, ACTIVATION a, float dropout) +void activate_array_ongpu(cl_mem x, int n, ACTIVATION a) { cl_setup(); cl_kernel kernel = get_activation_kernel(); @@ -119,8 +118,6 @@ void activate_array_ongpu(cl_mem x, int n, ACTIVATION a, float dropout) cl.error = clSetKernelArg(kernel, i++, sizeof(x), (void*) &x); cl.error = clSetKernelArg(kernel, i++, sizeof(n), (void*) &n); cl.error = clSetKernelArg(kernel, i++, sizeof(a), (void*) &a); - cl.error = clSetKernelArg(kernel, i++, sizeof(dropout), - (void*) &dropout); check_error(cl); size_t gsize = n; diff --git a/src/activations.cl b/src/activations.cl index 6ab135a1..65131c55 100644 --- a/src/activations.cl +++ b/src/activations.cl @@ -8,27 +8,26 @@ float relu_activate(float x){return x*(x>0);} float ramp_activate(float x){return x*(x>0)+.1*x;} float tanh_activate(float x){return (exp(2*x)-1)/(exp(2*x)+1);} -float activate(float x, ACTIVATION a, float dropout) +float activate(float x, ACTIVATION a) { - //if((float)rand()/RAND_MAX < dropout) return 0; switch(a){ case LINEAR: - return linear_activate(x)/(1-dropout); + return linear_activate(x); case SIGMOID: - return sigmoid_activate(x)/(1-dropout); + return sigmoid_activate(x); case RELU: - return relu_activate(x)/(1-dropout); + return relu_activate(x); case RAMP: - return ramp_activate(x)/(1-dropout); + return ramp_activate(x); case TANH: - return tanh_activate(x)/(1-dropout); + return tanh_activate(x); } return 0; } __kernel void activate_array(__global float *x, - const int n, const ACTIVATION a, const float dropout) + const int n, const ACTIVATION a) { int i = get_global_id(0); - x[i] = activate(x[i], a, dropout); + x[i] = activate(x[i], a); } diff --git a/src/activations.h b/src/activations.h index e47914c5..8c4287e0 100644 --- a/src/activations.h +++ b/src/activations.h @@ -9,12 +9,12 @@ typedef enum{ ACTIVATION get_activation(char *s); char *get_activation_string(ACTIVATION a); -float activate(float x, ACTIVATION a, float dropout); +float activate(float x, ACTIVATION a); float gradient(float x, ACTIVATION a); void gradient_array(const float *x, const int n, const ACTIVATION a, float *delta); -void activate_array(float *x, const int n, const ACTIVATION a, float dropout); +void activate_array(float *x, const int n, const ACTIVATION a); #ifdef GPU -void activate_array_ongpu(cl_mem x, int n, ACTIVATION a, float dropout); +void activate_array_ongpu(cl_mem x, int n, ACTIVATION a); #endif #endif diff --git a/src/cnn.c b/src/cnn.c index cac11494..f8661942 100644 --- a/src/cnn.c +++ b/src/cnn.c @@ -51,7 +51,7 @@ void test_convolve_matrix() int i; clock_t start = clock(), end; for(i = 0; i < 1000; ++i){ - im2col_cpu(dog.data, dog.c, dog.h, dog.w, size, stride, 0, matrix); + im2col_cpu(dog.data,1, dog.c, dog.h, dog.w, size, stride, 0, matrix); gemm(0,0,n,mw,mh,1,filters,mh,matrix,mw,1,edge.data,mw); } end = clock(); @@ -75,7 +75,7 @@ void verify_convolutional_layer() int size = 3; float eps = .00000001; image test = make_random_image(5,5, 1); - convolutional_layer layer = *make_convolutional_layer(1,test.h,test.w,test.c, n, size, stride, 0, RELU); + convolutional_layer layer = *make_convolutional_layer(1,test.h,test.w,test.c, n, size, stride, 0, RELU,0,0,0); image out = get_convolutional_image(layer); float **jacobian = calloc(test.h*test.w*test.c, sizeof(float)); @@ -158,25 +158,10 @@ void test_rotate() void test_parser() { - network net = parse_network_cfg("test_parser.cfg"); - float input[1]; - int count = 0; - - float avgerr = 0; - while(++count < 100000000){ - float v = ((float)rand()/RAND_MAX); - float truth = v*v; - input[0] = v; - forward_network(net, input, 1); - float *out = get_network_output(net); - float *delta = get_network_delta(net); - float err = pow((out[0]-truth),2.); - avgerr = .99 * avgerr + .01 * err; - if(count % 1000000 == 0) printf("%f %f :%f AVG %f \n", truth, out[0], err, avgerr); - delta[0] = truth - out[0]; - backward_network(net, input, &truth); - update_network(net, .001,0,0); - } + network net = parse_network_cfg("cfg/test_parser.cfg"); + save_network(net, "cfg/test_parser_1.cfg"); + network net2 = parse_network_cfg("cfg/test_parser_1.cfg"); + save_network(net2, "cfg/test_parser_2.cfg"); } void test_data() @@ -206,7 +191,7 @@ void train_full() //scale_data_rows(train, 1./255.); normalize_data_rows(train); clock_t start = clock(), end; - float loss = train_network_sgd(net, train, 1000, lr, momentum, decay); + float loss = train_network_sgd(net, train, 1000); end = clock(); printf("%d: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", i, loss, (float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay); free_data(train); @@ -255,558 +240,567 @@ void test_full() void test_cifar10() { - data test = load_cifar10_data("images/cifar10/test_batch.bin"); - scale_data_rows(test, 1./255); - network net = parse_network_cfg("cfg/cifar10.cfg"); - int count = 0; - float lr = .000005; - float momentum = .99; - float decay = 0.001; - decay = 0; - int batch = 10000; - while(++count <= 10000){ - char buff[256]; - sprintf(buff, "images/cifar10/data_batch_%d.bin", rand()%5+1); - data train = load_cifar10_data(buff); - scale_data_rows(train, 1./255); - train_network_sgd(net, train, batch, lr, momentum, decay); - //printf("%5f %5f\n",(double)count*batch/train.X.rows, loss); - - float test_acc = network_accuracy(net, test); - printf("%5f %5f\n",(double)count*batch/train.X.rows/5, 1-test_acc); - free_data(train); - } + srand(222222); + network net = parse_network_cfg("cfg/cifar10.cfg"); + //data test = load_cifar10_data("data/cifar10/test_batch.bin"); + int count = 0; + int iters = 10000/net.batch; + data train = load_all_cifar10(); + while(++count <= 10000){ + clock_t start = clock(), end; + float loss = train_network_sgd(net, train, iters); + end = clock(); + //visualize_network(net); + //cvWaitKey(1000); + //float test_acc = network_accuracy(net, test); + //printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay); + printf("%d: Loss: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, (float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay); + } + free_data(train); } void test_vince() { - network net = parse_network_cfg("cfg/vince.cfg"); - data train = load_categorical_data_csv("images/vince.txt", 144, 2); - normalize_data_rows(train); + network net = parse_network_cfg("cfg/vince.cfg"); + data train = load_categorical_data_csv("images/vince.txt", 144, 2); + normalize_data_rows(train); + + int count = 0; + //float lr = .00005; + //float momentum = .9; + //float decay = 0.0001; + //decay = 0; + int batch = 10000; + while(++count <= 10000){ + float loss = train_network_sgd(net, train, batch); + printf("%5f %5f\n",(double)count*batch/train.X.rows, loss); + } +} + +void test_nist_single() +{ + srand(222222); + network net = parse_network_cfg("cfg/nist.cfg"); + data train = load_categorical_data_csv("data/mnist/mnist_tiny.csv", 0, 10); + normalize_data_rows(train); + float loss = train_network_sgd(net, train, 5); + printf("Loss: %f, LR: %f, Momentum: %f, Decay: %f\n", loss, net.learning_rate, net.momentum, net.decay); - int count = 0; - float lr = .00005; - float momentum = .9; - float decay = 0.0001; - decay = 0; - int batch = 10000; - while(++count <= 10000){ - float loss = train_network_sgd(net, train, batch, lr, momentum, decay); - printf("%5f %5f\n",(double)count*batch/train.X.rows, loss); - } } void test_nist() { - srand(222222); - network net = parse_network_cfg("cfg/nist.cfg"); - data train = load_categorical_data_csv("data/mnist/mnist_train.csv", 0, 10); - data test = load_categorical_data_csv("data/mnist/mnist_test.csv",0,10); - normalize_data_rows(train); - normalize_data_rows(test); - //randomize_data(train); - int count = 0; - float lr = .0001; - float momentum = .9; - float decay = 0.0001; - //clock_t start = clock(), end; - int iters = 1000; - while(++count <= 10){ - clock_t start = clock(), end; - float loss = train_network_sgd(net, train, iters, lr, momentum, decay); - end = clock(); - float test_acc = network_accuracy(net, test); + srand(222222); + network net = parse_network_cfg("cfg/nist.cfg"); + data train = load_categorical_data_csv("data/mnist/mnist_train.csv", 0, 10); + data test = load_categorical_data_csv("data/mnist/mnist_test.csv",0,10); + translate_data_rows(train, -144); + scale_data_rows(train, 1./128); + translate_data_rows(test, -144); + scale_data_rows(test, 1./128); + //randomize_data(train); + int count = 0; + //clock_t start = clock(), end; + int iters = 10000/net.batch; + while(++count <= 100){ + clock_t start = clock(), end; + float loss = train_network_sgd(net, train, iters); + end = clock(); + float test_acc = network_accuracy(net, test); //float test_acc = 0; - printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay); + printf("%d: Loss: %f, Test Acc: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", count, loss, test_acc,(float)(end-start)/CLOCKS_PER_SEC, net.learning_rate, net.momentum, net.decay); + //save_network(net, "cfg/nist_basic_trained.cfg"); - //printf("%5d Training Loss: %lf, Params: %f %f %f, ",count*1000, loss, lr, momentum, decay); - //end = clock(); - //printf("Time: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC); - //start=end; - //lr *= .5; - } - //save_network(net, "cfg/nist_basic_trained.cfg"); + //printf("%5d Training Loss: %lf, Params: %f %f %f, ",count*1000, loss, lr, momentum, decay); + //end = clock(); + //printf("Time: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC); + //start=end; + //lr *= .5; + } + //save_network(net, "cfg/nist_basic_trained.cfg"); } void test_ensemble() { - int i; - srand(888888); - data d = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10); - normalize_data_rows(d); - data test = load_categorical_data_csv("mnist/mnist_test.csv", 0,10); - normalize_data_rows(test); - data train = d; - // data *split = split_data(d, 1, 10); - // data train = split[0]; - // data test = split[1]; - matrix prediction = make_matrix(test.y.rows, test.y.cols); - int n = 30; - for(i = 0; i < n; ++i){ - int count = 0; - float lr = .0005; - float momentum = .9; - float decay = .01; - network net = parse_network_cfg("nist.cfg"); - while(++count <= 15){ - float acc = train_network_sgd(net, train, train.X.rows, lr, momentum, decay); - printf("Training Accuracy: %lf Learning Rate: %f Momentum: %f Decay: %f\n", acc, lr, momentum, decay ); - lr /= 2; - } - matrix partial = network_predict_data(net, test); - float acc = matrix_accuracy(test.y, partial); - printf("Model Accuracy: %lf\n", acc); - matrix_add_matrix(partial, prediction); - acc = matrix_accuracy(test.y, prediction); - printf("Current Ensemble Accuracy: %lf\n", acc); - free_matrix(partial); - } - float acc = matrix_accuracy(test.y, prediction); - printf("Full Ensemble Accuracy: %lf\n", acc); + int i; + srand(888888); + data d = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10); + normalize_data_rows(d); + data test = load_categorical_data_csv("mnist/mnist_test.csv", 0,10); + normalize_data_rows(test); + data train = d; + // data *split = split_data(d, 1, 10); + // data train = split[0]; + // data test = split[1]; + matrix prediction = make_matrix(test.y.rows, test.y.cols); + int n = 30; + for(i = 0; i < n; ++i){ + int count = 0; + float lr = .0005; + float momentum = .9; + float decay = .01; + network net = parse_network_cfg("nist.cfg"); + while(++count <= 15){ + float acc = train_network_sgd(net, train, train.X.rows); + printf("Training Accuracy: %lf Learning Rate: %f Momentum: %f Decay: %f\n", acc, lr, momentum, decay ); + lr /= 2; + } + matrix partial = network_predict_data(net, test); + float acc = matrix_accuracy(test.y, partial); + printf("Model Accuracy: %lf\n", acc); + matrix_add_matrix(partial, prediction); + acc = matrix_accuracy(test.y, prediction); + printf("Current Ensemble Accuracy: %lf\n", acc); + free_matrix(partial); + } + float acc = matrix_accuracy(test.y, prediction); + printf("Full Ensemble Accuracy: %lf\n", acc); } void test_random_classify() { - network net = parse_network_cfg("connected.cfg"); - matrix m = csv_to_matrix("train.csv"); - //matrix ho = hold_out_matrix(&m, 2500); - float *truth = pop_column(&m, 0); - //float *ho_truth = pop_column(&ho, 0); - int i; - clock_t start = clock(), end; - int count = 0; - while(++count <= 300){ - for(i = 0; i < m.rows; ++i){ - int index = rand()%m.rows; - //image p = float_to_image(1690,1,1,m.vals[index]); - //normalize_image(p); - forward_network(net, m.vals[index], 1); - float *out = get_network_output(net); - float *delta = get_network_delta(net); - //printf("%f\n", out[0]); - delta[0] = truth[index] - out[0]; - // printf("%f\n", delta[0]); - //printf("%f %f\n", truth[index], out[0]); - //backward_network(net, m.vals[index], ); - update_network(net, .00001, 0,0); - } - //float test_acc = error_network(net, m, truth); - //float valid_acc = error_network(net, ho, ho_truth); - //printf("%f, %f\n", test_acc, valid_acc); - //fprintf(stderr, "%5d: %f Valid: %f\n",count, test_acc, valid_acc); - //if(valid_acc > .70) break; - } - end = clock(); - FILE *fp = fopen("submission/out.txt", "w"); - matrix test = csv_to_matrix("test.csv"); - truth = pop_column(&test, 0); - for(i = 0; i < test.rows; ++i){ - forward_network(net, test.vals[i], 0); - float *out = get_network_output(net); - if(fabs(out[0]) < .5) fprintf(fp, "0\n"); - else fprintf(fp, "1\n"); - } - fclose(fp); - printf("Neural Net Learning: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC); + network net = parse_network_cfg("connected.cfg"); + matrix m = csv_to_matrix("train.csv"); + //matrix ho = hold_out_matrix(&m, 2500); + float *truth = pop_column(&m, 0); + //float *ho_truth = pop_column(&ho, 0); + int i; + clock_t start = clock(), end; + int count = 0; + while(++count <= 300){ + for(i = 0; i < m.rows; ++i){ + int index = rand()%m.rows; + //image p = float_to_image(1690,1,1,m.vals[index]); + //normalize_image(p); + forward_network(net, m.vals[index], 1); + float *out = get_network_output(net); + float *delta = get_network_delta(net); + //printf("%f\n", out[0]); + delta[0] = truth[index] - out[0]; + // printf("%f\n", delta[0]); + //printf("%f %f\n", truth[index], out[0]); + //backward_network(net, m.vals[index], ); + update_network(net); + } + //float test_acc = error_network(net, m, truth); + //float valid_acc = error_network(net, ho, ho_truth); + //printf("%f, %f\n", test_acc, valid_acc); + //fprintf(stderr, "%5d: %f Valid: %f\n",count, test_acc, valid_acc); + //if(valid_acc > .70) break; + } + end = clock(); + FILE *fp = fopen("submission/out.txt", "w"); + matrix test = csv_to_matrix("test.csv"); + truth = pop_column(&test, 0); + for(i = 0; i < test.rows; ++i){ + forward_network(net, test.vals[i], 0); + float *out = get_network_output(net); + if(fabs(out[0]) < .5) fprintf(fp, "0\n"); + else fprintf(fp, "1\n"); + } + fclose(fp); + printf("Neural Net Learning: %lf seconds\n", (float)(end-start)/CLOCKS_PER_SEC); } void test_split() { - data train = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10); - data *split = split_data(train, 0, 13); - printf("%d, %d, %d\n", train.X.rows, split[0].X.rows, split[1].X.rows); + data train = load_categorical_data_csv("mnist/mnist_train.csv", 0, 10); + data *split = split_data(train, 0, 13); + printf("%d, %d, %d\n", train.X.rows, split[0].X.rows, split[1].X.rows); } void test_im2row() { - int h = 20; - int w = 20; - int c = 3; - int stride = 1; - int size = 11; - image test = make_random_image(h,w,c); - int mc = 1; - int mw = ((h-size)/stride+1)*((w-size)/stride+1); - int mh = (size*size*c); - int msize = mc*mw*mh; - float *matrix = calloc(msize, sizeof(float)); - int i; - for(i = 0; i < 1000; ++i){ - im2col_cpu(test.data, c, h, w, size, stride, 0, matrix); - //image render = float_to_image(mh, mw, mc, matrix); - } + int h = 20; + int w = 20; + int c = 3; + int stride = 1; + int size = 11; + image test = make_random_image(h,w,c); + int mc = 1; + int mw = ((h-size)/stride+1)*((w-size)/stride+1); + int mh = (size*size*c); + int msize = mc*mw*mh; + float *matrix = calloc(msize, sizeof(float)); + int i; + for(i = 0; i < 1000; ++i){ + im2col_cpu(test.data,1, c, h, w, size, stride, 0, matrix); + //image render = float_to_image(mh, mw, mc, matrix); + } } void flip_network() { - network net = parse_network_cfg("cfg/voc_imagenet_orig.cfg"); - save_network(net, "cfg/voc_imagenet_rev.cfg"); + network net = parse_network_cfg("cfg/voc_imagenet_orig.cfg"); + save_network(net, "cfg/voc_imagenet_rev.cfg"); } void tune_VOC() { - network net = parse_network_cfg("cfg/voc_start.cfg"); - srand(2222222); - int i = 20; - char *labels[] = {"aeroplane","bicycle","bird","boat","bottle","bus","car","cat","chair","cow","diningtable","dog","horse","motorbike","person","pottedplant","sheep","sofa","train","tvmonitor"}; - float lr = .000005; - float momentum = .9; - float decay = 0.0001; - while(i++ < 1000 || 1){ - data train = load_data_image_pathfile_random("/home/pjreddie/VOC2012/trainval_paths.txt", 10, labels, 20, 256, 256); + network net = parse_network_cfg("cfg/voc_start.cfg"); + srand(2222222); + int i = 20; + char *labels[] = {"aeroplane","bicycle","bird","boat","bottle","bus","car","cat","chair","cow","diningtable","dog","horse","motorbike","person","pottedplant","sheep","sofa","train","tvmonitor"}; + float lr = .000005; + float momentum = .9; + float decay = 0.0001; + while(i++ < 1000 || 1){ + data train = load_data_image_pathfile_random("/home/pjreddie/VOC2012/trainval_paths.txt", 10, labels, 20, 256, 256); - image im = float_to_image(256, 256, 3,train.X.vals[0]); - show_image(im, "input"); - visualize_network(net); - cvWaitKey(100); + image im = float_to_image(256, 256, 3,train.X.vals[0]); + show_image(im, "input"); + visualize_network(net); + cvWaitKey(100); - translate_data_rows(train, -144); - clock_t start = clock(), end; - float loss = train_network_sgd(net, train, 10, lr, momentum, decay); - end = clock(); - printf("%d: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", i, loss, (float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay); - free_data(train); + translate_data_rows(train, -144); + clock_t start = clock(), end; + float loss = train_network_sgd(net, train, 10); + end = clock(); + printf("%d: %f, Time: %lf seconds, LR: %f, Momentum: %f, Decay: %f\n", i, loss, (float)(end-start)/CLOCKS_PER_SEC, lr, momentum, decay); + free_data(train); /* - if(i%10==0){ - char buff[256]; - sprintf(buff, "/home/pjreddie/voc_cfg/voc_ramp_%d.cfg", i); - save_network(net, buff); - } - */ - //lr *= .99; - } + if(i%10==0){ + char buff[256]; + sprintf(buff, "/home/pjreddie/voc_cfg/voc_ramp_%d.cfg", i); + save_network(net, buff); + } + */ + //lr *= .99; + } } int voc_size(int x) { - x = x-1+3; - x = x-1+3; - x = x-1+3; - x = (x-1)*2+1; - x = x-1+5; - x = (x-1)*2+1; - x = (x-1)*4+11; - return x; + x = x-1+3; + x = x-1+3; + x = x-1+3; + x = (x-1)*2+1; + x = x-1+5; + x = (x-1)*2+1; + x = (x-1)*4+11; + return x; } image features_output_size(network net, IplImage *src, int outh, int outw) { - int h = voc_size(outh); - int w = voc_size(outw); - fprintf(stderr, "%d %d\n", h, w); + int h = voc_size(outh); + int w = voc_size(outw); + fprintf(stderr, "%d %d\n", h, w); - IplImage *sized = cvCreateImage(cvSize(w,h), src->depth, src->nChannels); - cvResize(src, sized, CV_INTER_LINEAR); - image im = ipl_to_image(sized); - //normalize_array(im.data, im.h*im.w*im.c); - translate_image(im, -144); - resize_network(net, im.h, im.w, im.c); - forward_network(net, im.data, 0); - image out = get_network_image(net); - free_image(im); - cvReleaseImage(&sized); - return copy_image(out); + IplImage *sized = cvCreateImage(cvSize(w,h), src->depth, src->nChannels); + cvResize(src, sized, CV_INTER_LINEAR); + image im = ipl_to_image(sized); + //normalize_array(im.data, im.h*im.w*im.c); + translate_image(im, -144); + resize_network(net, im.h, im.w, im.c); + forward_network(net, im.data, 0); + image out = get_network_image(net); + free_image(im); + cvReleaseImage(&sized); + return copy_image(out); } void features_VOC_image_size(char *image_path, int h, int w) { - int j; - network net = parse_network_cfg("cfg/voc_imagenet.cfg"); - fprintf(stderr, "%s\n", image_path); + int j; + network net = parse_network_cfg("cfg/voc_imagenet.cfg"); + fprintf(stderr, "%s\n", image_path); - IplImage* src = 0; - if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path); - image out = features_output_size(net, src, h, w); - for(j = 0; j < out.c*out.h*out.w; ++j){ - if(j != 0) printf(","); - printf("%g", out.data[j]); - } - printf("\n"); - free_image(out); - cvReleaseImage(&src); + IplImage* src = 0; + if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path); + image out = features_output_size(net, src, h, w); + for(j = 0; j < out.c*out.h*out.w; ++j){ + if(j != 0) printf(","); + printf("%g", out.data[j]); + } + printf("\n"); + free_image(out); + cvReleaseImage(&src); } void visualize_imagenet_topk(char *filename) { - int i,j,k,l; - int topk = 10; - network net = parse_network_cfg("cfg/voc_imagenet.cfg"); - list *plist = get_paths(filename); - node *n = plist->front; - int h = voc_size(1), w = voc_size(1); - int num = get_network_image(net).c; - image **vizs = calloc(num, sizeof(image*)); - float **score = calloc(num, sizeof(float *)); - for(i = 0; i < num; ++i){ - vizs[i] = calloc(topk, sizeof(image)); - for(j = 0; j < topk; ++j) vizs[i][j] = make_image(h,w,3); - score[i] = calloc(topk, sizeof(float)); - } + int i,j,k,l; + int topk = 10; + network net = parse_network_cfg("cfg/voc_imagenet.cfg"); + list *plist = get_paths(filename); + node *n = plist->front; + int h = voc_size(1), w = voc_size(1); + int num = get_network_image(net).c; + image **vizs = calloc(num, sizeof(image*)); + float **score = calloc(num, sizeof(float *)); + for(i = 0; i < num; ++i){ + vizs[i] = calloc(topk, sizeof(image)); + for(j = 0; j < topk; ++j) vizs[i][j] = make_image(h,w,3); + score[i] = calloc(topk, sizeof(float)); + } - int count = 0; - while(n){ - ++count; - char *image_path = (char *)n->val; - image im = load_image(image_path, 0, 0); - n = n->next; - if(im.h < 200 || im.w < 200) continue; - printf("Processing %dx%d image\n", im.h, im.w); - resize_network(net, im.h, im.w, im.c); - //scale_image(im, 1./255); - translate_image(im, -144); - forward_network(net, im.data, 0); - image out = get_network_image(net); + int count = 0; + while(n){ + ++count; + char *image_path = (char *)n->val; + image im = load_image(image_path, 0, 0); + n = n->next; + if(im.h < 200 || im.w < 200) continue; + printf("Processing %dx%d image\n", im.h, im.w); + resize_network(net, im.h, im.w, im.c); + //scale_image(im, 1./255); + translate_image(im, -144); + forward_network(net, im.data, 0); + image out = get_network_image(net); - int dh = (im.h - h)/(out.h-1); - int dw = (im.w - w)/(out.w-1); - //printf("%d %d\n", dh, dw); - for(k = 0; k < out.c; ++k){ - float topv = 0; - int topi = -1; - int topj = -1; - for(i = 0; i < out.h; ++i){ - for(j = 0; j < out.w; ++j){ - float val = get_pixel(out, i, j, k); - if(val > topv){ - topv = val; - topi = i; - topj = j; - } - } - } - if(topv){ - image sub = get_sub_image(im, dh*topi, dw*topj, h, w); - for(l = 0; l < topk; ++l){ - if(topv > score[k][l]){ - float swap = score[k][l]; - score[k][l] = topv; - topv = swap; + int dh = (im.h - h)/(out.h-1); + int dw = (im.w - w)/(out.w-1); + //printf("%d %d\n", dh, dw); + for(k = 0; k < out.c; ++k){ + float topv = 0; + int topi = -1; + int topj = -1; + for(i = 0; i < out.h; ++i){ + for(j = 0; j < out.w; ++j){ + float val = get_pixel(out, i, j, k); + if(val > topv){ + topv = val; + topi = i; + topj = j; + } + } + } + if(topv){ + image sub = get_sub_image(im, dh*topi, dw*topj, h, w); + for(l = 0; l < topk; ++l){ + if(topv > score[k][l]){ + float swap = score[k][l]; + score[k][l] = topv; + topv = swap; - image swapi = vizs[k][l]; - vizs[k][l] = sub; - sub = swapi; - } - } - free_image(sub); - } - } - free_image(im); - if(count%50 == 0){ - image grid = grid_images(vizs, num, topk); - //show_image(grid, "IMAGENET Visualization"); - save_image(grid, "IMAGENET Grid Single Nonorm"); - free_image(grid); - } - } - //cvWaitKey(0); + image swapi = vizs[k][l]; + vizs[k][l] = sub; + sub = swapi; + } + } + free_image(sub); + } + } + free_image(im); + if(count%50 == 0){ + image grid = grid_images(vizs, num, topk); + //show_image(grid, "IMAGENET Visualization"); + save_image(grid, "IMAGENET Grid Single Nonorm"); + free_image(grid); + } + } + //cvWaitKey(0); } void visualize_imagenet_features(char *filename) { - int i,j,k; - network net = parse_network_cfg("cfg/voc_imagenet.cfg"); - list *plist = get_paths(filename); - node *n = plist->front; - int h = voc_size(1), w = voc_size(1); - int num = get_network_image(net).c; - image *vizs = calloc(num, sizeof(image)); - for(i = 0; i < num; ++i) vizs[i] = make_image(h, w, 3); - while(n){ - char *image_path = (char *)n->val; - image im = load_image(image_path, 0, 0); - printf("Processing %dx%d image\n", im.h, im.w); - resize_network(net, im.h, im.w, im.c); - forward_network(net, im.data, 0); - image out = get_network_image(net); + int i,j,k; + network net = parse_network_cfg("cfg/voc_imagenet.cfg"); + list *plist = get_paths(filename); + node *n = plist->front; + int h = voc_size(1), w = voc_size(1); + int num = get_network_image(net).c; + image *vizs = calloc(num, sizeof(image)); + for(i = 0; i < num; ++i) vizs[i] = make_image(h, w, 3); + while(n){ + char *image_path = (char *)n->val; + image im = load_image(image_path, 0, 0); + printf("Processing %dx%d image\n", im.h, im.w); + resize_network(net, im.h, im.w, im.c); + forward_network(net, im.data, 0); + image out = get_network_image(net); - int dh = (im.h - h)/h; - int dw = (im.w - w)/w; - for(i = 0; i < out.h; ++i){ - for(j = 0; j < out.w; ++j){ - image sub = get_sub_image(im, dh*i, dw*j, h, w); - for(k = 0; k < out.c; ++k){ - float val = get_pixel(out, i, j, k); - //printf("%f, ", val); - image sub_c = copy_image(sub); - scale_image(sub_c, val); - add_into_image(sub_c, vizs[k], 0, 0); - free_image(sub_c); - } - free_image(sub); - } - } - //printf("\n"); - show_images(vizs, 10, "IMAGENET Visualization"); - cvWaitKey(1000); - n = n->next; - } - cvWaitKey(0); + int dh = (im.h - h)/h; + int dw = (im.w - w)/w; + for(i = 0; i < out.h; ++i){ + for(j = 0; j < out.w; ++j){ + image sub = get_sub_image(im, dh*i, dw*j, h, w); + for(k = 0; k < out.c; ++k){ + float val = get_pixel(out, i, j, k); + //printf("%f, ", val); + image sub_c = copy_image(sub); + scale_image(sub_c, val); + add_into_image(sub_c, vizs[k], 0, 0); + free_image(sub_c); + } + free_image(sub); + } + } + //printf("\n"); + show_images(vizs, 10, "IMAGENET Visualization"); + cvWaitKey(1000); + n = n->next; + } + cvWaitKey(0); } void visualize_cat() { - network net = parse_network_cfg("cfg/voc_imagenet.cfg"); - image im = load_image("data/cat.png", 0, 0); - printf("Processing %dx%d image\n", im.h, im.w); - resize_network(net, im.h, im.w, im.c); - forward_network(net, im.data, 0); + network net = parse_network_cfg("cfg/voc_imagenet.cfg"); + image im = load_image("data/cat.png", 0, 0); + printf("Processing %dx%d image\n", im.h, im.w); + resize_network(net, im.h, im.w, im.c); + forward_network(net, im.data, 0); - visualize_network(net); - cvWaitKey(0); + visualize_network(net); + cvWaitKey(0); } void features_VOC_image(char *image_file, char *image_dir, char *out_dir, int flip, int interval) { - int i,j; - network net = parse_network_cfg("cfg/voc_imagenet.cfg"); - char image_path[1024]; - sprintf(image_path, "%s/%s",image_dir, image_file); - char out_path[1024]; - if (flip)sprintf(out_path, "%s%d/%s_r.txt",out_dir, interval, image_file); - else sprintf(out_path, "%s%d/%s.txt",out_dir, interval, image_file); - printf("%s\n", image_file); + int i,j; + network net = parse_network_cfg("cfg/voc_imagenet.cfg"); + char image_path[1024]; + sprintf(image_path, "%s/%s",image_dir, image_file); + char out_path[1024]; + if (flip)sprintf(out_path, "%s%d/%s_r.txt",out_dir, interval, image_file); + else sprintf(out_path, "%s%d/%s.txt",out_dir, interval, image_file); + printf("%s\n", image_file); - IplImage* src = 0; - if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path); - if(flip)cvFlip(src, 0, 1); - int w = src->width; - int h = src->height; - int sbin = 8; - double scale = pow(2., 1./interval); - int m = (w= interval"); - image *ims = calloc(max_scale+interval, sizeof(image)); + IplImage* src = 0; + if( (src = cvLoadImage(image_path,-1)) == 0 ) file_error(image_path); + if(flip)cvFlip(src, 0, 1); + int w = src->width; + int h = src->height; + int sbin = 8; + double scale = pow(2., 1./interval); + int m = (w= interval"); + image *ims = calloc(max_scale+interval, sizeof(image)); - for(i = 0; i < interval; ++i){ - double factor = 1./pow(scale, i); - double ih = round(h*factor); - double iw = round(w*factor); - int ex_h = round(ih/4.) - 2; - int ex_w = round(iw/4.) - 2; - ims[i] = features_output_size(net, src, ex_h, ex_w); + for(i = 0; i < interval; ++i){ + double factor = 1./pow(scale, i); + double ih = round(h*factor); + double iw = round(w*factor); + int ex_h = round(ih/4.) - 2; + int ex_w = round(iw/4.) - 2; + ims[i] = features_output_size(net, src, ex_h, ex_w); - ih = round(h*factor); - iw = round(w*factor); - ex_h = round(ih/8.) - 2; - ex_w = round(iw/8.) - 2; - ims[i+interval] = features_output_size(net, src, ex_h, ex_w); - for(j = i+interval; j < max_scale; j += interval){ - factor /= 2.; - ih = round(h*factor); - iw = round(w*factor); - ex_h = round(ih/8.) - 2; - ex_w = round(iw/8.) - 2; - ims[j+interval] = features_output_size(net, src, ex_h, ex_w); - } - } - FILE *fp = fopen(out_path, "w"); - if(fp == 0) file_error(out_path); - for(i = 0; i < max_scale+interval; ++i){ - image out = ims[i]; - fprintf(fp, "%d, %d, %d\n",out.c, out.h, out.w); - for(j = 0; j < out.c*out.h*out.w; ++j){ - if(j != 0)fprintf(fp, ","); - float o = out.data[j]; - if(o < 0) o = 0; - fprintf(fp, "%g", o); - } - fprintf(fp, "\n"); - free_image(out); - } - free(ims); - fclose(fp); - cvReleaseImage(&src); + ih = round(h*factor); + iw = round(w*factor); + ex_h = round(ih/8.) - 2; + ex_w = round(iw/8.) - 2; + ims[i+interval] = features_output_size(net, src, ex_h, ex_w); + for(j = i+interval; j < max_scale; j += interval){ + factor /= 2.; + ih = round(h*factor); + iw = round(w*factor); + ex_h = round(ih/8.) - 2; + ex_w = round(iw/8.) - 2; + ims[j+interval] = features_output_size(net, src, ex_h, ex_w); + } + } + FILE *fp = fopen(out_path, "w"); + if(fp == 0) file_error(out_path); + for(i = 0; i < max_scale+interval; ++i){ + image out = ims[i]; + fprintf(fp, "%d, %d, %d\n",out.c, out.h, out.w); + for(j = 0; j < out.c*out.h*out.w; ++j){ + if(j != 0)fprintf(fp, ","); + float o = out.data[j]; + if(o < 0) o = 0; + fprintf(fp, "%g", o); + } + fprintf(fp, "\n"); + free_image(out); + } + free(ims); + fclose(fp); + cvReleaseImage(&src); } void test_distribution() { - IplImage* img = 0; - if( (img = cvLoadImage("im_small.jpg",-1)) == 0 ) file_error("im_small.jpg"); - network net = parse_network_cfg("cfg/voc_features.cfg"); - int h = img->height/8-2; - int w = img->width/8-2; - image out = features_output_size(net, img, h, w); - int c = out.c; - out.c = 1; - show_image(out, "output"); - out.c = c; - image input = ipl_to_image(img); - show_image(input, "input"); - CvScalar s; - int i,j; - image affects = make_image(input.h, input.w, 1); - int count = 0; - for(i = 0; iheight; i += 1){ - for(j = 0; j < img->width; j += 1){ - IplImage *copy = cvCloneImage(img); - s=cvGet2D(copy,i,j); // get the (i,j) pixel value - printf("%d/%d\n", count++, img->height*img->width); - s.val[0]=0; - s.val[1]=0; - s.val[2]=0; - cvSet2D(copy,i,j,s); // set the (i,j) pixel value - image mod = features_output_size(net, copy, h, w); - image dist = image_distance(out, mod); - show_image(affects, "affects"); - cvWaitKey(1); - cvReleaseImage(©); - //affects.data[i*affects.w + j] += dist.data[3*dist.w+5]; - affects.data[i*affects.w + j] += dist.data[1*dist.w+1]; - free_image(mod); - free_image(dist); - } - } - show_image(affects, "Origins"); - cvWaitKey(0); - cvWaitKey(0); + IplImage* img = 0; + if( (img = cvLoadImage("im_small.jpg",-1)) == 0 ) file_error("im_small.jpg"); + network net = parse_network_cfg("cfg/voc_features.cfg"); + int h = img->height/8-2; + int w = img->width/8-2; + image out = features_output_size(net, img, h, w); + int c = out.c; + out.c = 1; + show_image(out, "output"); + out.c = c; + image input = ipl_to_image(img); + show_image(input, "input"); + CvScalar s; + int i,j; + image affects = make_image(input.h, input.w, 1); + int count = 0; + for(i = 0; iheight; i += 1){ + for(j = 0; j < img->width; j += 1){ + IplImage *copy = cvCloneImage(img); + s=cvGet2D(copy,i,j); // get the (i,j) pixel value + printf("%d/%d\n", count++, img->height*img->width); + s.val[0]=0; + s.val[1]=0; + s.val[2]=0; + cvSet2D(copy,i,j,s); // set the (i,j) pixel value + image mod = features_output_size(net, copy, h, w); + image dist = image_distance(out, mod); + show_image(affects, "affects"); + cvWaitKey(1); + cvReleaseImage(©); + //affects.data[i*affects.w + j] += dist.data[3*dist.w+5]; + affects.data[i*affects.w + j] += dist.data[1*dist.w+1]; + free_image(mod); + free_image(dist); + } + } + show_image(affects, "Origins"); + cvWaitKey(0); + cvWaitKey(0); } int main(int argc, char *argv[]) { - //train_full(); - //test_distribution(); - //feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW); + //train_full(); + //test_distribution(); + //feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW); - //test_blas(); - //test_visualize(); - //test_gpu_blas(); - //test_blas(); - //test_convolve_matrix(); - // test_im2row(); - //test_split(); - //test_ensemble(); - test_nist(); - //test_cifar10(); - //test_vince(); - //test_full(); - //tune_VOC(); - //features_VOC_image(argv[1], argv[2], argv[3], 0); - //features_VOC_image(argv[1], argv[2], argv[3], 1); - //train_VOC(); - //features_VOC_image(argv[1], argv[2], argv[3], 0, 4); - //features_VOC_image(argv[1], argv[2], argv[3], 1, 4); - //features_VOC_image_size(argv[1], atoi(argv[2]), atoi(argv[3])); - //visualize_imagenet_features("data/assira/train.list"); - //visualize_imagenet_topk("data/VOC2012.list"); - //visualize_cat(); - //flip_network(); - //test_visualize(); - fprintf(stderr, "Success!\n"); - //test_random_preprocess(); - //test_random_classify(); - //test_parser(); - //test_backpropagate(); - //test_ann(); - //test_convolve(); - //test_upsample(); - //test_rotate(); - //test_load(); - //test_network(); - //test_convolutional_layer(); - //verify_convolutional_layer(); - //test_color(); - //cvWaitKey(0); - return 0; + //test_blas(); + //test_visualize(); + //test_gpu_blas(); + //test_blas(); + //test_convolve_matrix(); + // test_im2row(); + //test_split(); + //test_ensemble(); + //test_nist_single(); + test_nist(); + //test_cifar10(); + //test_vince(); + //test_full(); + //tune_VOC(); + //features_VOC_image(argv[1], argv[2], argv[3], 0); + //features_VOC_image(argv[1], argv[2], argv[3], 1); + //train_VOC(); + //features_VOC_image(argv[1], argv[2], argv[3], 0, 4); + //features_VOC_image(argv[1], argv[2], argv[3], 1, 4); + //features_VOC_image_size(argv[1], atoi(argv[2]), atoi(argv[3])); + //visualize_imagenet_features("data/assira/train.list"); + //visualize_imagenet_topk("data/VOC2012.list"); + //visualize_cat(); + //flip_network(); + //test_visualize(); + //test_parser(); + fprintf(stderr, "Success!\n"); + //test_random_preprocess(); + //test_random_classify(); + //test_parser(); + //test_backpropagate(); + //test_ann(); + //test_convolve(); + //test_upsample(); + //test_rotate(); + //test_load(); + //test_network(); + //test_convolutional_layer(); + //verify_convolutional_layer(); + //test_color(); + //cvWaitKey(0); + return 0; } diff --git a/src/col2im.c b/src/col2im.c index bc15b7bd..fd7de4fa 100644 --- a/src/col2im.c +++ b/src/col2im.c @@ -1,4 +1,6 @@ -inline void col2im_set_pixel(float *im, int height, int width, int channels, +#include +#include +inline void col2im_add_pixel(float *im, int height, int width, int channels, int row, int col, int channel, int pad, float val) { row -= pad; @@ -6,7 +8,7 @@ inline void col2im_set_pixel(float *im, int height, int width, int channels, if (row < 0 || col < 0 || row >= height || col >= width) return; - im[col + width*(row + channel*height)] = val; + im[col + width*(row + channel*height)] += val; } //This one might be too, can't remember. void col2im_cpu(float* data_col, @@ -31,7 +33,7 @@ void col2im_cpu(float* data_col, int im_row = h_offset + h * stride; int im_col = w_offset + w * stride; double val = data_col[(c * height_col + h) * width_col + w]; - col2im_set_pixel(data_im, height, width, channels, + col2im_add_pixel(data_im, height, width, channels, im_row, im_col, c_im, pad, val); } } diff --git a/src/connected_layer.c b/src/connected_layer.c index bebf2d9d..368fb63c 100644 --- a/src/connected_layer.c +++ b/src/connected_layer.c @@ -7,15 +7,19 @@ #include #include -connected_layer *make_connected_layer(int batch, int inputs, int outputs, float dropout, ACTIVATION activation) +connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay) { fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs); int i; connected_layer *layer = calloc(1, sizeof(connected_layer)); + + layer->learning_rate = learning_rate; + layer->momentum = momentum; + layer->decay = decay; + layer->inputs = inputs; layer->outputs = outputs; layer->batch=batch; - layer->dropout = dropout; layer->output = calloc(batch*outputs, sizeof(float*)); layer->delta = calloc(batch*outputs, sizeof(float*)); @@ -25,8 +29,9 @@ connected_layer *make_connected_layer(int batch, int inputs, int outputs, float layer->weight_momentum = calloc(inputs*outputs, sizeof(float)); layer->weights = calloc(inputs*outputs, sizeof(float)); float scale = 1./inputs; + //scale = .01; for(i = 0; i < inputs*outputs; ++i) - layer->weights[i] = scale*(rand_uniform()); + layer->weights[i] = scale*(rand_uniform()-.5); layer->bias_updates = calloc(outputs, sizeof(float)); layer->bias_adapt = calloc(outputs, sizeof(float)); @@ -40,25 +45,24 @@ connected_layer *make_connected_layer(int batch, int inputs, int outputs, float return layer; } -void update_connected_layer(connected_layer layer, float step, float momentum, float decay) +void update_connected_layer(connected_layer layer) { int i; for(i = 0; i < layer.outputs; ++i){ - layer.bias_momentum[i] = step*(layer.bias_updates[i]) + momentum*layer.bias_momentum[i]; + layer.bias_momentum[i] = layer.learning_rate*(layer.bias_updates[i]) + layer.momentum*layer.bias_momentum[i]; layer.biases[i] += layer.bias_momentum[i]; } for(i = 0; i < layer.outputs*layer.inputs; ++i){ - layer.weight_momentum[i] = step*(layer.weight_updates[i] - decay*layer.weights[i]) + momentum*layer.weight_momentum[i]; + layer.weight_momentum[i] = layer.learning_rate*(layer.weight_updates[i] - layer.decay*layer.weights[i]) + layer.momentum*layer.weight_momentum[i]; layer.weights[i] += layer.weight_momentum[i]; } memset(layer.bias_updates, 0, layer.outputs*sizeof(float)); memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(float)); } -void forward_connected_layer(connected_layer layer, float *input, int train) +void forward_connected_layer(connected_layer layer, float *input) { int i; - if(!train) layer.dropout = 0; for(i = 0; i < layer.batch; ++i){ memcpy(layer.output+i*layer.outputs, layer.biases, layer.outputs*sizeof(float)); } @@ -69,7 +73,7 @@ void forward_connected_layer(connected_layer layer, float *input, int train) float *b = layer.weights; float *c = layer.output; gemm(0,0,m,n,k,1,a,k,b,n,1,c,n); - activate_array(layer.output, layer.outputs*layer.batch, layer.activation, layer.dropout); + activate_array(layer.output, layer.outputs*layer.batch, layer.activation); } void backward_connected_layer(connected_layer layer, float *input, float *delta) diff --git a/src/connected_layer.h b/src/connected_layer.h index ff5a0ce4..e9e461c5 100644 --- a/src/connected_layer.h +++ b/src/connected_layer.h @@ -4,6 +4,10 @@ #include "activations.h" typedef struct{ + float learning_rate; + float momentum; + float decay; + int batch; int inputs; int outputs; @@ -22,17 +26,15 @@ typedef struct{ float *output; float *delta; - float dropout; - ACTIVATION activation; } connected_layer; -connected_layer *make_connected_layer(int batch, int inputs, int outputs, float dropout, ACTIVATION activation); +connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay); -void forward_connected_layer(connected_layer layer, float *input, int train); +void forward_connected_layer(connected_layer layer, float *input); void backward_connected_layer(connected_layer layer, float *input, float *delta); -void update_connected_layer(connected_layer layer, float step, float momentum, float decay); +void update_connected_layer(connected_layer layer); #endif diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c index 44e92442..6c7f9470 100644 --- a/src/convolutional_layer.c +++ b/src/convolutional_layer.c @@ -37,11 +37,16 @@ image get_convolutional_delta(convolutional_layer layer) return float_to_image(h,w,c,layer.delta); } -convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation) +convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, float learning_rate, float momentum, float decay) { int i; size = 2*(size/2)+1; //HA! And you thought you'd use an even sized filter... convolutional_layer *layer = calloc(1, sizeof(convolutional_layer)); + + layer->learning_rate = learning_rate; + layer->momentum = momentum; + layer->decay = decay; + layer->h = h; layer->w = w; layer->c = c; @@ -59,7 +64,8 @@ convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, in layer->bias_updates = calloc(n, sizeof(float)); layer->bias_momentum = calloc(n, sizeof(float)); float scale = 1./(size*size*c); - for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform()); + //scale = .0001; + for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform()-.5); for(i = 0; i < n; ++i){ //layer->biases[i] = rand_normal()*scale + scale; layer->biases[i] = .5; @@ -79,7 +85,7 @@ convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, in layer->bias_updates_cl = cl_make_array(layer->bias_updates, n); layer->bias_momentum_cl = cl_make_array(layer->bias_momentum, n); - layer->col_image_cl = cl_make_array(layer->col_image, layer.batch*out_h*out_w*size*size*c); + layer->col_image_cl = cl_make_array(layer->col_image, layer->batch*out_h*out_w*size*size*c); layer->delta_cl = cl_make_array(layer->delta, layer->batch*out_h*out_w*n); layer->output_cl = cl_make_array(layer->output, layer->batch*out_h*out_w*n); #endif @@ -136,9 +142,10 @@ void forward_convolutional_layer(const convolutional_layer layer, float *in) float *b = layer.col_image; float *c = layer.output; + im2col_cpu(in, layer.batch, layer.c, layer.h, layer.w, + layer.size, layer.stride, layer.pad, b); + for(i = 0; i < layer.batch; ++i){ - im2col_cpu(in, layer.c, layer.h, layer.w, - layer.size, layer.stride, layer.pad, b); gemm(0,0,m,n,k,1,a,k,b,n,1,c,n); c += n*m; in += layer.h*layer.w*layer.c; @@ -149,29 +156,9 @@ void forward_convolutional_layer(const convolutional_layer layer, float *in) for(i = 0; i < m*n; ++i) printf("%f, ", layer.output[i]); printf("\n"); */ - activate_array(layer.output, m*n*layer.batch, layer.activation, 0.); + activate_array(layer.output, m*n*layer.batch, layer.activation); } -#ifdef GPU -void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in) -{ - int m = layer.n; - int k = layer.size*layer.size*layer.c; - int n = convolutional_out_height(layer)* - convolutional_out_width(layer)* - layer.batch; - - cl_write_array(layer.filters_cl, layer.filters, m*k); - cl_mem a = layer.filters_cl; - cl_mem b = layer.col_image_cl; - cl_mem c = layer.output_cl; - im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, b); - gemm_ongpu(0,0,m,n,k,1,a,k,b,n,0,c,n); - activate_array_ongpu(layer.output_cl, m*n, layer.activation, 0.); - cl_read_array(layer.output_cl, layer.output, m*n); -} -#endif - void learn_bias_convolutional_layer(convolutional_layer layer) { int i,b; @@ -225,15 +212,15 @@ void backward_convolutional_layer(convolutional_layer layer, float *delta) } } -void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay) +void update_convolutional_layer(convolutional_layer layer) { int size = layer.size*layer.size*layer.c*layer.n; - axpy_cpu(layer.n, step, layer.bias_updates, 1, layer.biases, 1); - scal_cpu(layer.n, momentum, layer.bias_updates, 1); + axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1); + scal_cpu(layer.n,layer.momentum, layer.bias_updates, 1); - scal_cpu(size, 1.-step*decay, layer.filters, 1); - axpy_cpu(size, step, layer.filter_updates, 1, layer.filters, 1); - scal_cpu(size, momentum, layer.filter_updates, 1); + scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1); + axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1); + scal_cpu(size, layer.momentum, layer.filter_updates, 1); } @@ -284,9 +271,29 @@ image *visualize_convolutional_layer(convolutional_layer layer, char *window, im image dc = collapse_image_layers(delta, 1); char buff[256]; sprintf(buff, "%s: Output", window); - show_image(dc, buff); - save_image(dc, buff); + //show_image(dc, buff); + //save_image(dc, buff); free_image(dc); return single_filters; } +#ifdef GPU +void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in) +{ + int m = layer.n; + int k = layer.size*layer.size*layer.c; + int n = convolutional_out_height(layer)* + convolutional_out_width(layer)* + layer.batch; + + cl_write_array(layer.filters_cl, layer.filters, m*k); + cl_mem a = layer.filters_cl; + cl_mem b = layer.col_image_cl; + cl_mem c = layer.output_cl; + im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, b); + gemm_ongpu(0,0,m,n,k,1,a,k,b,n,0,c,n); + activate_array_ongpu(layer.output_cl, m*n, layer.activation); + cl_read_array(layer.output_cl, layer.output, m*n); +} +#endif + diff --git a/src/convolutional_layer.h b/src/convolutional_layer.h index e0722f8d..f876e8b4 100644 --- a/src/convolutional_layer.h +++ b/src/convolutional_layer.h @@ -9,6 +9,10 @@ #include "activations.h" typedef struct { + float learning_rate; + float momentum; + float decay; + int batch; int h,w,c; int n; @@ -48,10 +52,10 @@ typedef struct { void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in); #endif -convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation); +convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, float learning_rate, float momentum, float decay); void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c); void forward_convolutional_layer(const convolutional_layer layer, float *in); -void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay); +void update_convolutional_layer(convolutional_layer layer); image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters); void backward_convolutional_layer(convolutional_layer layer, float *delta); diff --git a/src/convolutional_layer_gpu.c b/src/convolutional_layer_gpu.c new file mode 100644 index 00000000..e69de29b diff --git a/src/data.c b/src/data.c index 30ee9ecb..846b950a 100644 --- a/src/data.c +++ b/src/data.c @@ -131,6 +131,7 @@ data load_cifar10_data(char *filename) d.y = y; FILE *fp = fopen(filename, "rb"); + if(!fp) file_error(filename); for(i = 0; i < 10000; ++i){ unsigned char bytes[3073]; fread(bytes, 1, 3073, fp); @@ -140,10 +141,46 @@ data load_cifar10_data(char *filename) X.vals[i][j] = (double)bytes[j+1]; } } + translate_data_rows(d, -144); + scale_data_rows(d, 1./128); + //normalize_data_rows(d); fclose(fp); return d; } +data load_all_cifar10() +{ + data d; + d.shallow = 0; + int i,j,b; + matrix X = make_matrix(50000, 3072); + matrix y = make_matrix(50000, 10); + d.X = X; + d.y = y; + + + for(b = 0; b < 5; ++b){ + char buff[256]; + sprintf(buff, "data/cifar10/data_batch_%d.bin", b+1); + FILE *fp = fopen(buff, "rb"); + if(!fp) file_error(buff); + for(i = 0; i < 10000; ++i){ + unsigned char bytes[3073]; + fread(bytes, 1, 3073, fp); + int class = bytes[0]; + y.vals[i+b*10000][class] = 1; + for(j = 0; j < X.cols; ++j){ + X.vals[i+b*10000][j] = (double)bytes[j+1]; + } + } + fclose(fp); + } + //normalize_data_rows(d); + translate_data_rows(d, -144); + scale_data_rows(d, 1./128); + return d; +} + void randomize_data(data d) { int i; diff --git a/src/data.h b/src/data.h index c639d5fa..0a1830e6 100644 --- a/src/data.h +++ b/src/data.h @@ -18,6 +18,7 @@ data load_data_image_pathfile_part(char *filename, int part, int total, data load_data_image_pathfile_random(char *filename, int n, char **labels, int k, int h, int w); data load_cifar10_data(char *filename); +data load_all_cifar10(); list *get_paths(char *filename); data load_categorical_data_csv(char *filename, int target, int k); void normalize_data_rows(data d); diff --git a/src/dropout_layer.c b/src/dropout_layer.c new file mode 100644 index 00000000..fcad7b9e --- /dev/null +++ b/src/dropout_layer.c @@ -0,0 +1,26 @@ +#include "dropout_layer.h" +#include "stdlib.h" +#include "stdio.h" + +dropout_layer *make_dropout_layer(int batch, int inputs, float probability) +{ + fprintf(stderr, "Dropout Layer: %d inputs, %f probability\n", inputs, probability); + dropout_layer *layer = calloc(1, sizeof(dropout_layer)); + layer->probability = probability; + layer->inputs = inputs; + layer->batch = batch; + return layer; +} + +void forward_dropout_layer(dropout_layer layer, float *input) +{ + int i; + for(i = 0; i < layer.batch * layer.inputs; ++i){ + if((float)rand()/RAND_MAX < layer.probability) input[i] = 0; + else input[i] /= (1-layer.probability); + } +} +void backward_dropout_layer(dropout_layer layer, float *input, float *delta) +{ + // Don't do shit LULZ +} diff --git a/src/dropout_layer.h b/src/dropout_layer.h new file mode 100644 index 00000000..b164a921 --- /dev/null +++ b/src/dropout_layer.h @@ -0,0 +1,15 @@ +#ifndef DROPOUT_LAYER_H +#define DROPOUT_LAYER_H + +typedef struct{ + int batch; + int inputs; + float probability; +} dropout_layer; + +dropout_layer *make_dropout_layer(int batch, int inputs, float probability); + +void forward_dropout_layer(dropout_layer layer, float *input); +void backward_dropout_layer(dropout_layer layer, float *input, float *delta); + +#endif diff --git a/src/im2col.c b/src/im2col.c index 89748c90..6ed9d891 100644 --- a/src/im2col.c +++ b/src/im2col.c @@ -51,11 +51,11 @@ void im2col_cpu_batch(float* data_im, //From Berkeley Vision's Caffe! //https://github.com/BVLC/caffe/blob/master/LICENSE -void im2col_cpu(float* data_im, +void im2col_cpu(float* data_im, const int batch, const int channels, const int height, const int width, const int ksize, const int stride, int pad, float* data_col) { - int c,h,w; + int c,h,w,b; int height_col = (height - ksize) / stride + 1; int width_col = (width - ksize) / stride + 1; if (pad){ @@ -64,19 +64,25 @@ void im2col_cpu(float* data_im, pad = ksize/2; } int channels_col = channels * ksize * ksize; - for (c = 0; c < channels_col; ++c) { - int w_offset = c % ksize; - int h_offset = (c / ksize) % ksize; - int c_im = c / ksize / ksize; - for (h = 0; h < height_col; ++h) { - for (w = 0; w < width_col; ++w) { - int im_row = h_offset + h * stride; - int im_col = w_offset + w * stride; - int col_index = (c * height_col + h) * width_col + w; - data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, - im_row, im_col, c_im, pad); + int im_size = height*width*channels; + int col_size = height_col*width_col*channels_col; + for (b = 0; b < batch; ++b) { + for (c = 0; c < channels_col; ++c) { + int w_offset = c % ksize; + int h_offset = (c / ksize) % ksize; + int c_im = c / ksize / ksize; + for (h = 0; h < height_col; ++h) { + for (w = 0; w < width_col; ++w) { + int im_row = h_offset + h * stride; + int im_col = w_offset + w * stride; + int col_index = (c * height_col + h) * width_col + w; + data_col[col_index] = im2col_get_pixel(data_im, height, width, channels, + im_row, im_col, c_im, pad); + } } } + data_im += im_size; + data_col += col_size; } } diff --git a/src/im2col.cl b/src/im2col.cl index 0226d282..765a92df 100644 --- a/src/im2col.cl +++ b/src/im2col.cl @@ -1,7 +1,7 @@ -__kernel void im2col(__global float *data_im, - const int batch, const int channels, const int height, const int width, - const int ksize, const int stride, __global float *data_col) +__kernel void im2col(__global float *data_im, const int im_offset, + const int channels, const int height, const int width, + const int ksize, const int stride, __global float *data_col, const int col_offset) { int b = get_global_id(0); int c = get_global_id(1); diff --git a/src/image.c b/src/image.c index e2c451b7..b25bf05b 100644 --- a/src/image.c +++ b/src/image.c @@ -138,7 +138,7 @@ void show_image(image p, char *name) } free_image(copy); if(disp->height < 500 || disp->width < 500 || disp->height > 1000){ - int w = 1500; + int w = 500; int h = w*p.h/p.w; if(h > 1000){ h = 1000; @@ -720,7 +720,7 @@ image collapse_images_horz(image *ims, int n) void show_images(image *ims, int n, char *window) { image m = collapse_images_vert(ims, n); - save_image(m, window); + //save_image(m, window); show_image(m, window); free_image(m); } diff --git a/src/maxpool_layer.c b/src/maxpool_layer.c index 54a734a8..08c9f2f2 100644 --- a/src/maxpool_layer.c +++ b/src/maxpool_layer.c @@ -17,14 +17,15 @@ image get_maxpool_delta(maxpool_layer layer) return float_to_image(h,w,c,layer.delta); } -maxpool_layer *make_maxpool_layer(int batch, int h, int w, int c, int stride) +maxpool_layer *make_maxpool_layer(int batch, int h, int w, int c, int size, int stride) { - fprintf(stderr, "Maxpool Layer: %d x %d x %d image, %d stride\n", h,w,c,stride); + fprintf(stderr, "Maxpool Layer: %d x %d x %d image, %d size, %d stride\n", h,w,c,size,stride); maxpool_layer *layer = calloc(1, sizeof(maxpool_layer)); layer->batch = batch; layer->h = h; layer->w = w; layer->c = c; + layer->size = size; layer->stride = stride; layer->output = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * c*batch, sizeof(float)); layer->delta = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * c*batch, sizeof(float)); @@ -40,6 +41,32 @@ void resize_maxpool_layer(maxpool_layer *layer, int h, int w, int c) layer->delta = realloc(layer->delta, ((h-1)/layer->stride+1) * ((w-1)/layer->stride+1) * c * layer->batch*sizeof(float)); } +float get_max_region(image im, int h, int w, int c, int size) +{ + int i,j; + int lower = (-size-1)/2 + 1; + int upper = size/2 + 1; + + int lh = (h-lower < 0) ? 0 : h-lower; + int uh = (h+upper > im.h) ? im.h : h+upper; + + int lw = (w-lower < 0) ? 0 : w-lower; + int uw = (w+upper > im.w) ? im.w : w+upper; + + //printf("%d\n", -3/2); + //printf("%d %d\n", lower, upper); + //printf("%d %d %d %d\n", lh, uh, lw, uw); + + float max = -FLT_MAX; + for(i = lh; i < uh; ++i){ + for(j = lw; j < uw; ++j){ + float val = get_pixel(im, i, j, c); + if (val > max) max = val; + } + } + return max; +} + void forward_maxpool_layer(const maxpool_layer layer, float *in) { int b; @@ -52,19 +79,40 @@ void forward_maxpool_layer(const maxpool_layer layer, float *in) image output = float_to_image(h,w,c,layer.output+b*h*w*c); int i,j,k; - for(i = 0; i < output.h*output.w*output.c; ++i) output.data[i] = -DBL_MAX; for(k = 0; k < input.c; ++k){ - for(i = 0; i < input.h; ++i){ - for(j = 0; j < input.w; ++j){ - float val = get_pixel(input, i, j, k); - float cur = get_pixel(output, i/layer.stride, j/layer.stride, k); - if(val > cur) set_pixel(output, i/layer.stride, j/layer.stride, k, val); + for(i = 0; i < input.h; i += layer.stride){ + for(j = 0; j < input.w; j += layer.stride){ + float max = get_max_region(input, i, j, k, layer.size); + set_pixel(output, i/layer.stride, j/layer.stride, k, max); } } } } } +float set_max_region_delta(image im, image delta, int h, int w, int c, int size, float max, float error) +{ + int i,j; + int lower = (-size-1)/2 + 1; + int upper = size/2 + 1; + + int lh = (h-lower < 0) ? 0 : h-lower; + int uh = (h+upper > im.h) ? im.h : h+upper; + + int lw = (w-lower < 0) ? 0 : w-lower; + int uw = (w+upper > im.w) ? im.w : w+upper; + + for(i = lh; i < uh; ++i){ + for(j = lw; j < uw; ++j){ + float val = get_pixel(im, i, j, c); + if (val == max){ + add_pixel(delta, i, j, c, error); + } + } + } + return max; +} + void backward_maxpool_layer(const maxpool_layer layer, float *in, float *delta) { int b; @@ -76,18 +124,15 @@ void backward_maxpool_layer(const maxpool_layer layer, float *in, float *delta) int c = layer.c; image output = float_to_image(h,w,c,layer.output+b*h*w*c); image output_delta = float_to_image(h,w,c,layer.delta+b*h*w*c); + zero_image(input_delta); int i,j,k; for(k = 0; k < input.c; ++k){ - for(i = 0; i < input.h; ++i){ - for(j = 0; j < input.w; ++j){ - float val = get_pixel(input, i, j, k); - float cur = get_pixel(output, i/layer.stride, j/layer.stride, k); - float d = get_pixel(output_delta, i/layer.stride, j/layer.stride, k); - if(val == cur) { - set_pixel(input_delta, i, j, k, d); - } - else set_pixel(input_delta, i, j, k, 0); + for(i = 0; i < input.h; i += layer.stride){ + for(j = 0; j < input.w; j += layer.stride){ + float max = get_pixel(output, i/layer.stride, j/layer.stride, k); + float error = get_pixel(output_delta, i/layer.stride, j/layer.stride, k); + set_max_region_delta(input, input_delta, i, j, k, layer.size, max, error); } } } diff --git a/src/maxpool_layer.h b/src/maxpool_layer.h index 92d41e66..cde84458 100644 --- a/src/maxpool_layer.h +++ b/src/maxpool_layer.h @@ -7,12 +7,13 @@ typedef struct { int batch; int h,w,c; int stride; + int size; float *delta; float *output; } maxpool_layer; image get_maxpool_image(maxpool_layer layer); -maxpool_layer *make_maxpool_layer(int batch, int h, int w, int c, int stride); +maxpool_layer *make_maxpool_layer(int batch, int h, int w, int c, int size, int stride); void resize_maxpool_layer(maxpool_layer *layer, int h, int w, int c); void forward_maxpool_layer(const maxpool_layer layer, float *in); void backward_maxpool_layer(const maxpool_layer layer, float *in, float *delta); diff --git a/src/mini_blas.h b/src/mini_blas.h index 95e924bf..c80e6ad5 100644 --- a/src/mini_blas.h +++ b/src/mini_blas.h @@ -25,7 +25,7 @@ void gemm_ongpu(int TA, int TB, int M, int N, int K, float ALPHA, cl_mem C_gpu, int ldc); #endif -void im2col_cpu(float* data_im, +void im2col_cpu(float* data_im, const int batch, const int channels, const int height, const int width, const int ksize, const int stride, int pad, float* data_col); diff --git a/src/network.c b/src/network.c index 70883989..ed927a8c 100644 --- a/src/network.c +++ b/src/network.c @@ -9,6 +9,7 @@ #include "maxpool_layer.h" #include "normalization_layer.h" #include "softmax_layer.h" +#include "dropout_layer.h" network make_network(int n, int batch) { @@ -25,94 +26,6 @@ network make_network(int n, int batch) return net; } -void print_convolutional_cfg(FILE *fp, convolutional_layer *l, int first) -{ - int i; - fprintf(fp, "[convolutional]\n"); - if(first) fprintf(fp, "batch=%d\n" - "height=%d\n" - "width=%d\n" - "channels=%d\n", - l->batch,l->h, l->w, l->c); - fprintf(fp, "filters=%d\n" - "size=%d\n" - "stride=%d\n" - "activation=%s\n", - l->n, l->size, l->stride, - get_activation_string(l->activation)); - fprintf(fp, "data="); - for(i = 0; i < l->n; ++i) fprintf(fp, "%g,", l->biases[i]); - for(i = 0; i < l->n*l->c*l->size*l->size; ++i) fprintf(fp, "%g,", l->filters[i]); - fprintf(fp, "\n\n"); -} -void print_connected_cfg(FILE *fp, connected_layer *l, int first) -{ - int i; - fprintf(fp, "[connected]\n"); - if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs); - fprintf(fp, "output=%d\n" - "activation=%s\n", - l->outputs, - get_activation_string(l->activation)); - fprintf(fp, "data="); - for(i = 0; i < l->outputs; ++i) fprintf(fp, "%g,", l->biases[i]); - for(i = 0; i < l->inputs*l->outputs; ++i) fprintf(fp, "%g,", l->weights[i]); - fprintf(fp, "\n\n"); -} - -void print_maxpool_cfg(FILE *fp, maxpool_layer *l, int first) -{ - fprintf(fp, "[maxpool]\n"); - if(first) fprintf(fp, "batch=%d\n" - "height=%d\n" - "width=%d\n" - "channels=%d\n", - l->batch,l->h, l->w, l->c); - fprintf(fp, "stride=%d\n\n", l->stride); -} - -void print_normalization_cfg(FILE *fp, normalization_layer *l, int first) -{ - fprintf(fp, "[localresponsenormalization]\n"); - if(first) fprintf(fp, "batch=%d\n" - "height=%d\n" - "width=%d\n" - "channels=%d\n", - l->batch,l->h, l->w, l->c); - fprintf(fp, "size=%d\n" - "alpha=%g\n" - "beta=%g\n" - "kappa=%g\n\n", l->size, l->alpha, l->beta, l->kappa); -} - -void print_softmax_cfg(FILE *fp, softmax_layer *l, int first) -{ - fprintf(fp, "[softmax]\n"); - if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs); - fprintf(fp, "\n"); -} - -void save_network(network net, char *filename) -{ - FILE *fp = fopen(filename, "w"); - if(!fp) file_error(filename); - int i; - for(i = 0; i < net.n; ++i) - { - if(net.types[i] == CONVOLUTIONAL) - print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i], i==0); - else if(net.types[i] == CONNECTED) - print_connected_cfg(fp, (connected_layer *)net.layers[i], i==0); - else if(net.types[i] == MAXPOOL) - print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i], i==0); - else if(net.types[i] == NORMALIZATION) - print_normalization_cfg(fp, (normalization_layer *)net.layers[i], i==0); - else if(net.types[i] == SOFTMAX) - print_softmax_cfg(fp, (softmax_layer *)net.layers[i], i==0); - } - fclose(fp); -} - #ifdef GPU void forward_network(network net, float *input, int train) { @@ -169,7 +82,7 @@ void forward_network(network net, float *input, int train) } else if(net.types[i] == CONNECTED){ connected_layer layer = *(connected_layer *)net.layers[i]; - forward_connected_layer(layer, input, train); + forward_connected_layer(layer, input); input = layer.output; } else if(net.types[i] == SOFTMAX){ @@ -187,17 +100,22 @@ void forward_network(network net, float *input, int train) forward_normalization_layer(layer, input); input = layer.output; } + else if(net.types[i] == DROPOUT){ + if(!train) continue; + dropout_layer layer = *(dropout_layer *)net.layers[i]; + forward_dropout_layer(layer, input); + } } } #endif -void update_network(network net, float step, float momentum, float decay) +void update_network(network net) { int i; for(i = 0; i < net.n; ++i){ if(net.types[i] == CONVOLUTIONAL){ convolutional_layer layer = *(convolutional_layer *)net.layers[i]; - update_convolutional_layer(layer, step, momentum, decay); + update_convolutional_layer(layer); } else if(net.types[i] == MAXPOOL){ //maxpool_layer layer = *(maxpool_layer *)net.layers[i]; @@ -210,7 +128,7 @@ void update_network(network net, float step, float momentum, float decay) } else if(net.types[i] == CONNECTED){ connected_layer layer = *(connected_layer *)net.layers[i]; - update_connected_layer(layer, step, momentum, decay); + update_connected_layer(layer); } } } @@ -226,6 +144,8 @@ float *get_network_output_layer(network net, int i) } else if(net.types[i] == SOFTMAX){ softmax_layer layer = *(softmax_layer *)net.layers[i]; return layer.output; + } else if(net.types[i] == DROPOUT){ + return get_network_output_layer(net, i-1); } else if(net.types[i] == CONNECTED){ connected_layer layer = *(connected_layer *)net.layers[i]; return layer.output; @@ -251,6 +171,8 @@ float *get_network_delta_layer(network net, int i) } else if(net.types[i] == SOFTMAX){ softmax_layer layer = *(softmax_layer *)net.layers[i]; return layer.delta; + } else if(net.types[i] == DROPOUT){ + return get_network_delta_layer(net, i-1); } else if(net.types[i] == CONNECTED){ connected_layer layer = *(connected_layer *)net.layers[i]; return layer.delta; @@ -326,17 +248,17 @@ float backward_network(network net, float *input, float *truth) return error; } -float train_network_datum(network net, float *x, float *y, float step, float momentum, float decay) +float train_network_datum(network net, float *x, float *y) { forward_network(net, x, 1); //int class = get_predicted_class_network(net); float error = backward_network(net, x, y); - update_network(net, step, momentum, decay); + update_network(net); //return (y[class]?1:0); return error; } -float train_network_sgd(network net, data d, int n, float step, float momentum,float decay) +float train_network_sgd(network net, data d, int n) { int batch = net.batch; float *X = calloc(batch*d.X.cols, sizeof(float)); @@ -350,9 +272,9 @@ float train_network_sgd(network net, data d, int n, float step, float momentum,f memcpy(X+j*d.X.cols, d.X.vals[index], d.X.cols*sizeof(float)); memcpy(y+j*d.y.cols, d.y.vals[index], d.y.cols*sizeof(float)); } - float err = train_network_datum(net, X, y, step, momentum, decay); + float err = train_network_datum(net, X, y); sum += err; - //train_network_datum(net, X, y, step, momentum, decay); + //train_network_datum(net, X, y); /* float *y = d.y.vals[index]; int class = get_predicted_class_network(net); @@ -382,7 +304,7 @@ float train_network_sgd(network net, data d, int n, float step, float momentum,f free(y); return (float)sum/(n*batch); } -float train_network_batch(network net, data d, int n, float step, float momentum,float decay) +float train_network_batch(network net, data d, int n) { int i,j; float sum = 0; @@ -395,18 +317,18 @@ float train_network_batch(network net, data d, int n, float step, float momentum forward_network(net, x, 1); sum += backward_network(net, x, y); } - update_network(net, step, momentum, decay); + update_network(net); } return (float)sum/(n*batch); } -void train_network(network net, data d, float step, float momentum, float decay) +void train_network(network net, data d) { int i; int correct = 0; for(i = 0; i < d.X.rows; ++i){ - correct += train_network_datum(net, d.X.vals[i], d.y.vals[i], step, momentum, decay); + correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]); if(i%100 == 0){ visualize_network(net); cvWaitKey(10); @@ -430,6 +352,9 @@ int get_network_input_size_layer(network net, int i) else if(net.types[i] == CONNECTED){ connected_layer layer = *(connected_layer *)net.layers[i]; return layer.inputs; + } else if(net.types[i] == DROPOUT){ + dropout_layer layer = *(dropout_layer *) net.layers[i]; + return layer.inputs; } else if(net.types[i] == SOFTMAX){ softmax_layer layer = *(softmax_layer *)net.layers[i]; @@ -453,6 +378,9 @@ int get_network_output_size_layer(network net, int i) else if(net.types[i] == CONNECTED){ connected_layer layer = *(connected_layer *)net.layers[i]; return layer.outputs; + } else if(net.types[i] == DROPOUT){ + dropout_layer layer = *(dropout_layer *) net.layers[i]; + return layer.inputs; } else if(net.types[i] == SOFTMAX){ softmax_layer layer = *(softmax_layer *)net.layers[i]; diff --git a/src/network.h b/src/network.h index 35a58ca9..a9a6797d 100644 --- a/src/network.h +++ b/src/network.h @@ -11,12 +11,16 @@ typedef enum { CONNECTED, MAXPOOL, SOFTMAX, - NORMALIZATION + NORMALIZATION, + DROPOUT } LAYER_TYPE; typedef struct { int n; int batch; + float learning_rate; + float momentum; + float decay; void **layers; LAYER_TYPE *types; int outputs; @@ -31,10 +35,10 @@ typedef struct { network make_network(int n, int batch); void forward_network(network net, float *input, int train); float backward_network(network net, float *input, float *truth); -void update_network(network net, float step, float momentum, float decay); -float train_network_sgd(network net, data d, int n, float step, float momentum,float decay); -float train_network_batch(network net, data d, int n, float step, float momentum,float decay); -void train_network(network net, data d, float step, float momentum, float decay); +void update_network(network net); +float train_network_sgd(network net, data d, int n); +float train_network_batch(network net, data d, int n); +void train_network(network net, data d); matrix network_predict_data(network net, data test); float network_accuracy(network net, data d); float *get_network_output(network net); @@ -48,7 +52,6 @@ image get_network_image_layer(network net, int i); int get_predicted_class_network(network net); void print_network(network net); void visualize_network(network net); -void save_network(network net, char *filename); int resize_network(network net, int h, int w, int c); int get_network_input_size(network net); diff --git a/src/normalization_layer.c b/src/normalization_layer.c index 2d844e0e..67d873c9 100644 --- a/src/normalization_layer.c +++ b/src/normalization_layer.c @@ -72,7 +72,7 @@ void forward_normalization_layer(const normalization_layer layer, float *in) int next = k+layer.size/2; int prev = k-layer.size/2-1; if(next < layer.c) add_square_array(in+next*imsize, layer.sums, imsize); - if(prev > 0) sub_square_array(in+prev*imsize, layer.sums, imsize); + if(prev > 0) sub_square_array(in+prev*imsize, layer.sums, imsize); for(i = 0; i < imsize; ++i){ layer.output[k*imsize + i] = in[k*imsize+i] / pow(layer.kappa + layer.alpha * layer.sums[i], layer.beta); } diff --git a/src/opencl.c b/src/opencl.c index d78537b4..8f9edd3c 100644 --- a/src/opencl.c +++ b/src/opencl.c @@ -110,6 +110,15 @@ void cl_copy_array(cl_mem src, cl_mem dst, int n) check_error(cl); } +cl_mem cl_sub_array(cl_mem src, int offset, int size) +{ + cl_buffer_region r; + r.origin = offset*sizeof(float); + r.size = size*sizeof(float); + cl_mem sub = clCreateSubBuffer(src, CL_MEM_USE_HOST_PTR, CL_BUFFER_CREATE_TYPE_REGION, &r, 0); + return sub; +} + cl_mem cl_make_array(float *x, int n) { cl_setup(); diff --git a/src/opencl.h b/src/opencl.h index a7ee0bdb..9cf3acd4 100644 --- a/src/opencl.h +++ b/src/opencl.h @@ -25,5 +25,6 @@ void cl_read_array(cl_mem mem, float *x, int n); void cl_write_array(cl_mem mem, float *x, int n); cl_mem cl_make_array(float *x, int n); void cl_copy_array(cl_mem src, cl_mem dst, int n); +cl_mem cl_sub_array(cl_mem src, int offset, int size); #endif #endif diff --git a/src/option_list.c b/src/option_list.c index bb8b7101..76e10166 100644 --- a/src/option_list.c +++ b/src/option_list.c @@ -53,6 +53,13 @@ int option_find_int(list *l, char *key, int def) return def; } +float option_find_float_quiet(list *l, char *key, float def) +{ + char *v = option_find(l, key); + if(v) return atof(v); + return def; +} + float option_find_float(list *l, char *key, float def) { char *v = option_find(l, key); diff --git a/src/option_list.h b/src/option_list.h index 26cd36fc..fa795f3e 100644 --- a/src/option_list.h +++ b/src/option_list.h @@ -14,6 +14,7 @@ char *option_find(list *l, char *key); char *option_find_str(list *l, char *key, char *def); int option_find_int(list *l, char *key, int def); float option_find_float(list *l, char *key, float def); +float option_find_float_quiet(list *l, char *key, float def); void option_unused(list *l); #endif diff --git a/src/parser.c b/src/parser.c index b008882d..16563465 100644 --- a/src/parser.c +++ b/src/parser.c @@ -9,6 +9,7 @@ #include "maxpool_layer.h" #include "normalization_layer.h" #include "softmax_layer.h" +#include "dropout_layer.h" #include "list.h" #include "option_list.h" #include "utils.h" @@ -21,6 +22,7 @@ typedef struct{ int is_convolutional(section *s); int is_connected(section *s); int is_maxpool(section *s); +int is_dropout(section *s); int is_softmax(section *s); int is_normalization(section *s); list *read_cfg(char *filename); @@ -41,10 +43,11 @@ void free_section(section *s) free(s); } -convolutional_layer *parse_convolutional(list *options, network net, int count) +convolutional_layer *parse_convolutional(list *options, network *net, int count) { int i; int h,w,c; + float learning_rate, momentum, decay; int n = option_find_int(options, "filters",1); int size = option_find_int(options, "size",1); int stride = option_find_int(options, "stride",1); @@ -52,18 +55,27 @@ convolutional_layer *parse_convolutional(list *options, network net, int count) char *activation_s = option_find_str(options, "activation", "sigmoid"); ACTIVATION activation = get_activation(activation_s); if(count == 0){ + learning_rate = option_find_float(options, "learning_rate", .001); + momentum = option_find_float(options, "momentum", .9); + decay = option_find_float(options, "decay", .0001); h = option_find_int(options, "height",1); w = option_find_int(options, "width",1); c = option_find_int(options, "channels",1); - net.batch = option_find_int(options, "batch",1); + net->batch = option_find_int(options, "batch",1); + net->learning_rate = learning_rate; + net->momentum = momentum; + net->decay = decay; }else{ - image m = get_network_image_layer(net, count-1); + learning_rate = option_find_float_quiet(options, "learning_rate", net->learning_rate); + momentum = option_find_float_quiet(options, "momentum", net->momentum); + decay = option_find_float_quiet(options, "decay", net->decay); + image m = get_network_image_layer(*net, count-1); h = m.h; w = m.w; c = m.c; if(h == 0) error("Layer before convolutional layer must output image."); } - convolutional_layer *layer = make_convolutional_layer(net.batch,h,w,c,n,size,stride,pad,activation); + convolutional_layer *layer = make_convolutional_layer(net->batch,h,w,c,n,size,stride,pad,activation,learning_rate,momentum,decay); char *data = option_find_str(options, "data", 0); if(data){ char *curr = data; @@ -81,25 +93,60 @@ convolutional_layer *parse_convolutional(list *options, network net, int count) curr = next+1; } } + char *weights = option_find_str(options, "weights", 0); + char *biases = option_find_str(options, "biases", 0); + if(biases){ + char *curr = biases; + char *next = biases; + int done = 0; + for(i = 0; i < n && !done; ++i){ + while(*++next !='\0' && *next != ','); + if(*next == '\0') done = 1; + *next = '\0'; + sscanf(curr, "%g", &layer->biases[i]); + curr = next+1; + } + } + if(weights){ + char *curr = weights; + char *next = weights; + int done = 0; + for(i = 0; i < c*n*size*size && !done; ++i){ + while(*++next !='\0' && *next != ','); + if(*next == '\0') done = 1; + *next = '\0'; + sscanf(curr, "%g", &layer->filters[i]); + curr = next+1; + } + } option_unused(options); return layer; } -connected_layer *parse_connected(list *options, network net, int count) +connected_layer *parse_connected(list *options, network *net, int count) { int i; int input; + float learning_rate, momentum, decay; int output = option_find_int(options, "output",1); - float dropout = option_find_float(options, "dropout", 0.); char *activation_s = option_find_str(options, "activation", "sigmoid"); ACTIVATION activation = get_activation(activation_s); if(count == 0){ input = option_find_int(options, "input",1); - net.batch = option_find_int(options, "batch",1); + net->batch = option_find_int(options, "batch",1); + learning_rate = option_find_float(options, "learning_rate", .001); + momentum = option_find_float(options, "momentum", .9); + decay = option_find_float(options, "decay", .0001); + net->learning_rate = learning_rate; + net->momentum = momentum; + net->decay = decay; }else{ - input = get_network_output_size_layer(net, count-1); + learning_rate = option_find_float_quiet(options, "learning_rate", net->learning_rate); + momentum = option_find_float_quiet(options, "momentum", net->momentum); + decay = option_find_float_quiet(options, "decay", net->decay); + input = get_network_output_size_layer(*net, count-1); } - connected_layer *layer = make_connected_layer(net.batch, input, output, dropout, activation); + connected_layer *layer = make_connected_layer(net->batch, input, output, activation,learning_rate,momentum,decay); char *data = option_find_str(options, "data", 0); if(data){ char *curr = data; @@ -121,42 +168,58 @@ connected_layer *parse_connected(list *options, network net, int count) return layer; } -softmax_layer *parse_softmax(list *options, network net, int count) +softmax_layer *parse_softmax(list *options, network *net, int count) { int input; if(count == 0){ input = option_find_int(options, "input",1); - net.batch = option_find_int(options, "batch",1); + net->batch = option_find_int(options, "batch",1); }else{ - input = get_network_output_size_layer(net, count-1); + input = get_network_output_size_layer(*net, count-1); } - softmax_layer *layer = make_softmax_layer(net.batch, input); + softmax_layer *layer = make_softmax_layer(net->batch, input); option_unused(options); return layer; } -maxpool_layer *parse_maxpool(list *options, network net, int count) +maxpool_layer *parse_maxpool(list *options, network *net, int count) { int h,w,c; int stride = option_find_int(options, "stride",1); + int size = option_find_int(options, "size",stride); if(count == 0){ h = option_find_int(options, "height",1); w = option_find_int(options, "width",1); c = option_find_int(options, "channels",1); - net.batch = option_find_int(options, "batch",1); + net->batch = option_find_int(options, "batch",1); }else{ - image m = get_network_image_layer(net, count-1); + image m = get_network_image_layer(*net, count-1); h = m.h; w = m.w; c = m.c; if(h == 0) error("Layer before convolutional layer must output image."); } - maxpool_layer *layer = make_maxpool_layer(net.batch,h,w,c,stride); + maxpool_layer *layer = make_maxpool_layer(net->batch,h,w,c,size,stride); option_unused(options); return layer; } -normalization_layer *parse_normalization(list *options, network net, int count) +dropout_layer *parse_dropout(list *options, network *net, int count) +{ + int input; + float probability = option_find_float(options, "probability", .5); + if(count == 0){ + net->batch = option_find_int(options, "batch",1); + input = option_find_int(options, "input",1); + }else{ + input = get_network_output_size_layer(*net, count-1); + } + dropout_layer *layer = make_dropout_layer(net->batch,input,probability); + option_unused(options); + return layer; +} + +normalization_layer *parse_normalization(list *options, network *net, int count) { int h,w,c; int size = option_find_int(options, "size",1); @@ -167,15 +230,15 @@ normalization_layer *parse_normalization(list *options, network net, int count) h = option_find_int(options, "height",1); w = option_find_int(options, "width",1); c = option_find_int(options, "channels",1); - net.batch = option_find_int(options, "batch",1); + net->batch = option_find_int(options, "batch",1); }else{ - image m = get_network_image_layer(net, count-1); + image m = get_network_image_layer(*net, count-1); h = m.h; w = m.w; c = m.c; if(h == 0) error("Layer before convolutional layer must output image."); } - normalization_layer *layer = make_normalization_layer(net.batch,h,w,c,size, alpha, beta, kappa); + normalization_layer *layer = make_normalization_layer(net->batch,h,w,c,size, alpha, beta, kappa); option_unused(options); return layer; } @@ -191,30 +254,29 @@ network parse_network_cfg(char *filename) section *s = (section *)n->val; list *options = s->options; if(is_convolutional(s)){ - convolutional_layer *layer = parse_convolutional(options, net, count); + convolutional_layer *layer = parse_convolutional(options, &net, count); net.types[count] = CONVOLUTIONAL; net.layers[count] = layer; - net.batch = layer->batch; }else if(is_connected(s)){ - connected_layer *layer = parse_connected(options, net, count); + connected_layer *layer = parse_connected(options, &net, count); net.types[count] = CONNECTED; net.layers[count] = layer; - net.batch = layer->batch; }else if(is_softmax(s)){ - softmax_layer *layer = parse_softmax(options, net, count); + softmax_layer *layer = parse_softmax(options, &net, count); net.types[count] = SOFTMAX; net.layers[count] = layer; - net.batch = layer->batch; }else if(is_maxpool(s)){ - maxpool_layer *layer = parse_maxpool(options, net, count); + maxpool_layer *layer = parse_maxpool(options, &net, count); net.types[count] = MAXPOOL; net.layers[count] = layer; - net.batch = layer->batch; }else if(is_normalization(s)){ - normalization_layer *layer = parse_normalization(options, net, count); + normalization_layer *layer = parse_normalization(options, &net, count); net.types[count] = NORMALIZATION; net.layers[count] = layer; - net.batch = layer->batch; + }else if(is_dropout(s)){ + dropout_layer *layer = parse_dropout(options, &net, count); + net.types[count] = DROPOUT; + net.layers[count] = layer; }else{ fprintf(stderr, "Type not recognized: %s\n", s->type); } @@ -243,6 +305,10 @@ int is_maxpool(section *s) return (strcmp(s->type, "[max]")==0 || strcmp(s->type, "[maxpool]")==0); } +int is_dropout(section *s) +{ + return (strcmp(s->type, "[dropout]")==0); +} int is_softmax(section *s) { @@ -308,3 +374,120 @@ list *read_cfg(char *filename) return sections; } +void print_convolutional_cfg(FILE *fp, convolutional_layer *l, network net, int count) +{ + int i; + fprintf(fp, "[convolutional]\n"); + if(count == 0) { + fprintf(fp, "batch=%d\n" + "height=%d\n" + "width=%d\n" + "channels=%d\n" + "learning_rate=%g\n" + "momentum=%g\n" + "decay=%g\n", + l->batch,l->h, l->w, l->c, l->learning_rate, l->momentum, l->decay); + } else { + if(l->learning_rate != net.learning_rate) + fprintf(fp, "learning_rate=%g\n", l->learning_rate); + if(l->momentum != net.momentum) + fprintf(fp, "momentum=%g\n", l->momentum); + if(l->decay != net.decay) + fprintf(fp, "decay=%g\n", l->decay); + } + fprintf(fp, "filters=%d\n" + "size=%d\n" + "stride=%d\n" + "pad=%d\n" + "activation=%s\n", + l->n, l->size, l->stride, l->pad, + get_activation_string(l->activation)); + fprintf(fp, "biases="); + for(i = 0; i < l->n; ++i) fprintf(fp, "%g,", l->biases[i]); + fprintf(fp, "\n"); + fprintf(fp, "weights="); + for(i = 0; i < l->n*l->c*l->size*l->size; ++i) fprintf(fp, "%g,", l->filters[i]); + fprintf(fp, "\n\n"); +} +void print_connected_cfg(FILE *fp, connected_layer *l, network net, int count) +{ + int i; + fprintf(fp, "[connected]\n"); + if(count == 0){ + fprintf(fp, "batch=%d\n" + "input=%d\n" + "learning_rate=%g\n" + "momentum=%g\n" + "decay=%g\n", + l->batch, l->inputs, l->learning_rate, l->momentum, l->decay); + } else { + if(l->learning_rate != net.learning_rate) + fprintf(fp, "learning_rate=%g\n", l->learning_rate); + if(l->momentum != net.momentum) + fprintf(fp, "momentum=%g\n", l->momentum); + if(l->decay != net.decay) + fprintf(fp, "decay=%g\n", l->decay); + } + fprintf(fp, "output=%d\n" + "activation=%s\n", + l->outputs, + get_activation_string(l->activation)); + fprintf(fp, "data="); + for(i = 0; i < l->outputs; ++i) fprintf(fp, "%g,", l->biases[i]); + for(i = 0; i < l->inputs*l->outputs; ++i) fprintf(fp, "%g,", l->weights[i]); + fprintf(fp, "\n\n"); +} + +void print_maxpool_cfg(FILE *fp, maxpool_layer *l, network net, int count) +{ + fprintf(fp, "[maxpool]\n"); + if(count == 0) fprintf(fp, "batch=%d\n" + "height=%d\n" + "width=%d\n" + "channels=%d\n", + l->batch,l->h, l->w, l->c); + fprintf(fp, "size=%d\nstride=%d\n\n", l->size, l->stride); +} + +void print_normalization_cfg(FILE *fp, normalization_layer *l, network net, int count) +{ + fprintf(fp, "[localresponsenormalization]\n"); + if(count == 0) fprintf(fp, "batch=%d\n" + "height=%d\n" + "width=%d\n" + "channels=%d\n", + l->batch,l->h, l->w, l->c); + fprintf(fp, "size=%d\n" + "alpha=%g\n" + "beta=%g\n" + "kappa=%g\n\n", l->size, l->alpha, l->beta, l->kappa); +} + +void print_softmax_cfg(FILE *fp, softmax_layer *l, network net, int count) +{ + fprintf(fp, "[softmax]\n"); + if(count == 0) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs); + fprintf(fp, "\n"); +} + +void save_network(network net, char *filename) +{ + FILE *fp = fopen(filename, "w"); + if(!fp) file_error(filename); + int i; + for(i = 0; i < net.n; ++i) + { + if(net.types[i] == CONVOLUTIONAL) + print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i], net, i); + else if(net.types[i] == CONNECTED) + print_connected_cfg(fp, (connected_layer *)net.layers[i], net, i); + else if(net.types[i] == MAXPOOL) + print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i], net, i); + else if(net.types[i] == NORMALIZATION) + print_normalization_cfg(fp, (normalization_layer *)net.layers[i], net, i); + else if(net.types[i] == SOFTMAX) + print_softmax_cfg(fp, (softmax_layer *)net.layers[i], net, i); + } + fclose(fp); +} + diff --git a/src/parser.h b/src/parser.h index 878baa35..891e658b 100644 --- a/src/parser.h +++ b/src/parser.h @@ -3,5 +3,6 @@ #include "network.h" network parse_network_cfg(char *filename); +void save_network(network net, char *filename); #endif diff --git a/src/softmax_layer.c b/src/softmax_layer.c index 12684238..b6e9fe9e 100644 --- a/src/softmax_layer.c +++ b/src/softmax_layer.c @@ -1,4 +1,5 @@ #include "softmax_layer.h" +#include "mini_blas.h" #include #include #include @@ -11,6 +12,7 @@ softmax_layer *make_softmax_layer(int batch, int inputs) layer->inputs = inputs; layer->output = calloc(inputs*batch, sizeof(float)); layer->delta = calloc(inputs*batch, sizeof(float)); + layer->jacobian = calloc(inputs*inputs*batch, sizeof(float)); return layer; } @@ -51,6 +53,28 @@ void forward_softmax_layer(const softmax_layer layer, float *input) void backward_softmax_layer(const softmax_layer layer, float *input, float *delta) { +/* + int i,j,b; + for(b = 0; b < layer.batch; ++b){ + for(i = 0; i < layer.inputs; ++i){ + for(j = 0; j < layer.inputs; ++j){ + int d = (i==j); + layer.jacobian[b*layer.inputs*layer.inputs + i*layer.inputs + j] = + layer.output[b*layer.inputs + i] * (d - layer.output[b*layer.inputs + j]); + } + } + } + for(b = 0; b < layer.batch; ++b){ + int M = layer.inputs; + int N = 1; + int K = layer.inputs; + float *A = layer.jacobian + b*layer.inputs*layer.inputs; + float *B = layer.delta + b*layer.inputs; + float *C = delta + b*layer.inputs; + gemm(0,0,M,N,K,1,A,K,B,N,0,C,N); + } + */ + int i; for(i = 0; i < layer.inputs*layer.batch; ++i){ delta[i] = layer.delta[i]; diff --git a/src/softmax_layer.h b/src/softmax_layer.h index 414030c6..22752508 100644 --- a/src/softmax_layer.h +++ b/src/softmax_layer.h @@ -6,6 +6,7 @@ typedef struct { int batch; float *delta; float *output; + float *jacobian; } softmax_layer; softmax_layer *make_softmax_layer(int batch, int inputs);