| | |
| | | #include "convolutional_layer.h" |
| | | #include "utils.h" |
| | | #include "im2col.h" |
| | | #include "col2im.h" |
| | | #include "blas.h" |
| | | #include "gemm.h" |
| | | #include <stdio.h> |
| | | #include <time.h> |
| | | |
| | | image get_convolutional_image(convolutional_layer layer) |
| | | int convolutional_out_height(convolutional_layer l) |
| | | { |
| | | int h = (layer.h-1)/layer.stride + 1; |
| | | int w = (layer.w-1)/layer.stride + 1; |
| | | int c = layer.n; |
| | | return double_to_image(h,w,c,layer.output); |
| | | int h = l.h; |
| | | if (!l.pad) h -= l.size; |
| | | else h -= 1; |
| | | return h/l.stride + 1; |
| | | } |
| | | |
| | | image get_convolutional_delta(convolutional_layer layer) |
| | | int convolutional_out_width(convolutional_layer l) |
| | | { |
| | | int h = (layer.h-1)/layer.stride + 1; |
| | | int w = (layer.w-1)/layer.stride + 1; |
| | | int c = layer.n; |
| | | return double_to_image(h,w,c,layer.delta); |
| | | int w = l.w; |
| | | if (!l.pad) w -= l.size; |
| | | else w -= 1; |
| | | return w/l.stride + 1; |
| | | } |
| | | |
| | | convolutional_layer *make_convolutional_layer(int h, int w, int c, int n, int size, int stride, ACTIVATION activator) |
| | | image get_convolutional_image(convolutional_layer l) |
| | | { |
| | | printf("Convolutional Layer: %d x %d x %d image, %d filters\n", h,w,c,n); |
| | | int h,w,c; |
| | | h = convolutional_out_height(l); |
| | | w = convolutional_out_width(l); |
| | | c = l.n; |
| | | return float_to_image(w,h,c,l.output); |
| | | } |
| | | |
| | | image get_convolutional_delta(convolutional_layer l) |
| | | { |
| | | int h,w,c; |
| | | h = convolutional_out_height(l); |
| | | w = convolutional_out_width(l); |
| | | c = l.n; |
| | | return float_to_image(w,h,c,l.delta); |
| | | } |
| | | |
| | | convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation) |
| | | { |
| | | int i; |
| | | convolutional_layer *layer = calloc(1, sizeof(convolutional_layer)); |
| | | layer->h = h; |
| | | layer->w = w; |
| | | layer->c = c; |
| | | layer->n = n; |
| | | layer->stride = stride; |
| | | layer->kernels = calloc(n, sizeof(image)); |
| | | layer->kernel_updates = calloc(n, sizeof(image)); |
| | | layer->biases = calloc(n, sizeof(double)); |
| | | layer->bias_updates = calloc(n, sizeof(double)); |
| | | convolutional_layer l = {0}; |
| | | l.type = CONVOLUTIONAL; |
| | | |
| | | l.h = h; |
| | | l.w = w; |
| | | l.c = c; |
| | | l.n = n; |
| | | l.batch = batch; |
| | | l.stride = stride; |
| | | l.size = size; |
| | | l.pad = pad; |
| | | |
| | | l.filters = calloc(c*n*size*size, sizeof(float)); |
| | | l.filter_updates = calloc(c*n*size*size, sizeof(float)); |
| | | |
| | | l.biases = calloc(n, sizeof(float)); |
| | | l.bias_updates = calloc(n, sizeof(float)); |
| | | //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){ |
| | | layer->biases[i] = .005; |
| | | layer->kernels[i] = make_random_kernel(size, c); |
| | | layer->kernel_updates[i] = make_random_kernel(size, c); |
| | | l.biases[i] = scale; |
| | | } |
| | | layer->output = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * n, sizeof(double)); |
| | | layer->delta = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * n, sizeof(double)); |
| | | layer->upsampled = make_image(h,w,n); |
| | | int out_h = convolutional_out_height(l); |
| | | int out_w = convolutional_out_width(l); |
| | | l.out_h = out_h; |
| | | l.out_w = out_w; |
| | | l.out_c = n; |
| | | l.outputs = l.out_h * l.out_w * l.out_c; |
| | | l.inputs = l.w * l.h * l.c; |
| | | |
| | | if(activator == SIGMOID){ |
| | | layer->activation = sigmoid_activation; |
| | | layer->gradient = sigmoid_gradient; |
| | | }else if(activator == RELU){ |
| | | layer->activation = relu_activation; |
| | | layer->gradient = relu_gradient; |
| | | }else if(activator == IDENTITY){ |
| | | layer->activation = identity_activation; |
| | | layer->gradient = identity_gradient; |
| | | } |
| | | return layer; |
| | | l.col_image = calloc(out_h*out_w*size*size*c, sizeof(float)); |
| | | l.output = calloc(l.batch*out_h * out_w * n, sizeof(float)); |
| | | l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float)); |
| | | |
| | | #ifdef GPU |
| | | l.filters_gpu = cuda_make_array(l.filters, c*n*size*size); |
| | | l.filter_updates_gpu = cuda_make_array(l.filter_updates, c*n*size*size); |
| | | |
| | | l.biases_gpu = cuda_make_array(l.biases, n); |
| | | l.bias_updates_gpu = cuda_make_array(l.bias_updates, n); |
| | | |
| | | l.col_image_gpu = cuda_make_array(l.col_image, out_h*out_w*size*size*c); |
| | | l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n); |
| | | l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n); |
| | | #endif |
| | | l.activation = activation; |
| | | |
| | | fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n); |
| | | |
| | | return l; |
| | | } |
| | | |
| | | void forward_convolutional_layer(const convolutional_layer layer, double *in) |
| | | void resize_convolutional_layer(convolutional_layer *l, int w, int h) |
| | | { |
| | | image input = double_to_image(layer.h, layer.w, layer.c, in); |
| | | image output = get_convolutional_image(layer); |
| | | int i,j; |
| | | for(i = 0; i < layer.n; ++i){ |
| | | convolve(input, layer.kernels[i], layer.stride, i, output); |
| | | } |
| | | for(i = 0; i < output.c; ++i){ |
| | | for(j = 0; j < output.h*output.w; ++j){ |
| | | int index = i*output.h*output.w + j; |
| | | output.data[index] += layer.biases[i]; |
| | | output.data[index] = layer.activation(output.data[index]); |
| | | l->w = w; |
| | | l->h = h; |
| | | int out_w = convolutional_out_width(*l); |
| | | int out_h = convolutional_out_height(*l); |
| | | |
| | | l->out_w = out_w; |
| | | l->out_h = out_h; |
| | | |
| | | l->outputs = l->out_h * l->out_w * l->out_c; |
| | | l->inputs = l->w * l->h * l->c; |
| | | |
| | | l->col_image = realloc(l->col_image, |
| | | out_h*out_w*l->size*l->size*l->c*sizeof(float)); |
| | | l->output = realloc(l->output, |
| | | l->batch*out_h * out_w * l->n*sizeof(float)); |
| | | l->delta = realloc(l->delta, |
| | | l->batch*out_h * out_w * l->n*sizeof(float)); |
| | | |
| | | #ifdef GPU |
| | | cuda_free(l->col_image_gpu); |
| | | cuda_free(l->delta_gpu); |
| | | cuda_free(l->output_gpu); |
| | | |
| | | l->col_image_gpu = cuda_make_array(0, out_h*out_w*l->size*l->size*l->c); |
| | | l->delta_gpu = cuda_make_array(0, l->batch*out_h*out_w*l->n); |
| | | l->output_gpu = cuda_make_array(0, l->batch*out_h*out_w*l->n); |
| | | #endif |
| | | } |
| | | |
| | | void bias_output(float *output, float *biases, int batch, int n, int size) |
| | | { |
| | | int i,j,b; |
| | | for(b = 0; b < batch; ++b){ |
| | | for(i = 0; i < n; ++i){ |
| | | for(j = 0; j < size; ++j){ |
| | | output[(b*n + i)*size + j] = biases[i]; |
| | | } |
| | | } |
| | | } |
| | | } |
| | | |
| | | void backward_convolutional_layer(convolutional_layer layer, double *input, double *delta) |
| | | void backward_bias(float *bias_updates, float *delta, int batch, int n, int size) |
| | | { |
| | | int i; |
| | | |
| | | image in_image = double_to_image(layer.h, layer.w, layer.c, input); |
| | | image in_delta = double_to_image(layer.h, layer.w, layer.c, delta); |
| | | image out_delta = get_convolutional_delta(layer); |
| | | zero_image(in_delta); |
| | | |
| | | for(i = 0; i < layer.n; ++i){ |
| | | back_convolve(in_delta, layer.kernels[i], layer.stride, i, out_delta); |
| | | } |
| | | for(i = 0; i < layer.h*layer.w*layer.c; ++i){ |
| | | in_delta.data[i] *= layer.gradient(in_image.data[i]); |
| | | } |
| | | } |
| | | |
| | | /* |
| | | void backpropagate_convolutional_layer_convolve(image input, convolutional_layer layer) |
| | | { |
| | | int i,j; |
| | | for(i = 0; i < layer.n; ++i){ |
| | | rotate_image(layer.kernels[i]); |
| | | } |
| | | |
| | | zero_image(input); |
| | | upsample_image(layer.output, layer.stride, layer.upsampled); |
| | | for(j = 0; j < input.c; ++j){ |
| | | for(i = 0; i < layer.n; ++i){ |
| | | two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, input, j); |
| | | int i,b; |
| | | for(b = 0; b < batch; ++b){ |
| | | for(i = 0; i < n; ++i){ |
| | | bias_updates[i] += sum_array(delta+size*(i+b*n), size); |
| | | } |
| | | } |
| | | |
| | | for(i = 0; i < layer.n; ++i){ |
| | | rotate_image(layer.kernels[i]); |
| | | } |
| | | } |
| | | */ |
| | | |
| | | void learn_convolutional_layer(convolutional_layer layer, double *input) |
| | | |
| | | void forward_convolutional_layer(const convolutional_layer l, network_state state) |
| | | { |
| | | int out_h = convolutional_out_height(l); |
| | | int out_w = convolutional_out_width(l); |
| | | int i; |
| | | |
| | | bias_output(l.output, l.biases, l.batch, l.n, out_h*out_w); |
| | | |
| | | int m = l.n; |
| | | int k = l.size*l.size*l.c; |
| | | int n = out_h*out_w; |
| | | |
| | | float *a = l.filters; |
| | | float *b = l.col_image; |
| | | float *c = l.output; |
| | | |
| | | for(i = 0; i < l.batch; ++i){ |
| | | im2col_cpu(state.input, l.c, l.h, l.w, |
| | | l.size, l.stride, l.pad, b); |
| | | gemm(0,0,m,n,k,1,a,k,b,n,1,c,n); |
| | | c += n*m; |
| | | state.input += l.c*l.h*l.w; |
| | | } |
| | | activate_array(l.output, m*n*l.batch, l.activation); |
| | | } |
| | | |
| | | void backward_convolutional_layer(convolutional_layer l, network_state state) |
| | | { |
| | | int i; |
| | | image in_image = double_to_image(layer.h, layer.w, layer.c, input); |
| | | image out_delta = get_convolutional_delta(layer); |
| | | for(i = 0; i < layer.n; ++i){ |
| | | kernel_update(in_image, layer.kernel_updates[i], layer.stride, i, out_delta); |
| | | layer.bias_updates[i] += avg_image_layer(out_delta, i); |
| | | } |
| | | } |
| | | int m = l.n; |
| | | int n = l.size*l.size*l.c; |
| | | int k = convolutional_out_height(l)* |
| | | convolutional_out_width(l); |
| | | |
| | | void update_convolutional_layer(convolutional_layer layer, double step) |
| | | { |
| | | return; |
| | | int i,j; |
| | | for(i = 0; i < layer.n; ++i){ |
| | | layer.biases[i] += step*layer.bias_updates[i]; |
| | | layer.bias_updates[i] = 0; |
| | | int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c; |
| | | for(j = 0; j < pixels; ++j){ |
| | | layer.kernels[i].data[j] += step*layer.kernel_updates[i].data[j]; |
| | | gradient_array(l.output, m*k*l.batch, l.activation, l.delta); |
| | | backward_bias(l.bias_updates, l.delta, l.batch, l.n, k); |
| | | |
| | | if(state.delta) memset(state.delta, 0, l.batch*l.h*l.w*l.c*sizeof(float)); |
| | | |
| | | for(i = 0; i < l.batch; ++i){ |
| | | float *a = l.delta + i*m*k; |
| | | float *b = l.col_image; |
| | | float *c = l.filter_updates; |
| | | |
| | | float *im = state.input+i*l.c*l.h*l.w; |
| | | |
| | | im2col_cpu(im, l.c, l.h, l.w, |
| | | l.size, l.stride, l.pad, b); |
| | | gemm(0,1,m,n,k,1,a,k,b,k,1,c,n); |
| | | |
| | | if(state.delta){ |
| | | a = l.filters; |
| | | b = l.delta + i*m*k; |
| | | c = l.col_image; |
| | | |
| | | gemm(1,0,n,k,m,1,a,n,b,k,0,c,k); |
| | | |
| | | col2im_cpu(l.col_image, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w); |
| | | } |
| | | zero_image(layer.kernel_updates[i]); |
| | | } |
| | | } |
| | | |
| | | void visualize_convolutional_layer(convolutional_layer layer) |
| | | void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay) |
| | | { |
| | | int size = l.size*l.size*l.c*l.n; |
| | | axpy_cpu(l.n, learning_rate/batch, l.bias_updates, 1, l.biases, 1); |
| | | scal_cpu(l.n, momentum, l.bias_updates, 1); |
| | | |
| | | axpy_cpu(size, -decay*batch, l.filters, 1, l.filter_updates, 1); |
| | | axpy_cpu(size, learning_rate/batch, l.filter_updates, 1, l.filters, 1); |
| | | scal_cpu(size, momentum, l.filter_updates, 1); |
| | | } |
| | | |
| | | |
| | | image get_convolutional_filter(convolutional_layer l, int i) |
| | | { |
| | | int h = l.size; |
| | | int w = l.size; |
| | | int c = l.c; |
| | | return float_to_image(w,h,c,l.filters+i*h*w*c); |
| | | } |
| | | |
| | | void rgbgr_filters(convolutional_layer l) |
| | | { |
| | | int i; |
| | | for(i = 0; i < l.n; ++i){ |
| | | image im = get_convolutional_filter(l, i); |
| | | if (im.c == 3) { |
| | | rgbgr_image(im); |
| | | } |
| | | } |
| | | } |
| | | |
| | | image *get_filters(convolutional_layer l) |
| | | { |
| | | image *filters = calloc(l.n, sizeof(image)); |
| | | int i; |
| | | for(i = 0; i < l.n; ++i){ |
| | | filters[i] = copy_image(get_convolutional_filter(l, i)); |
| | | } |
| | | return filters; |
| | | } |
| | | |
| | | image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_filters) |
| | | { |
| | | image *single_filters = get_filters(l); |
| | | show_images(single_filters, l.n, window); |
| | | |
| | | image delta = get_convolutional_image(l); |
| | | image dc = collapse_image_layers(delta, 1); |
| | | char buff[256]; |
| | | //image vis = make_image(layer.n*layer.size, layer.size*layer.kernels[0].c, 3); |
| | | for(i = 0; i < layer.n; ++i){ |
| | | image k = layer.kernels[i]; |
| | | sprintf(buff, "Kernel %d", i); |
| | | if(k.c <= 3) show_image(k, buff); |
| | | else show_image_layers(k, buff); |
| | | } |
| | | sprintf(buff, "%s: Output", window); |
| | | //show_image(dc, buff); |
| | | //save_image(dc, buff); |
| | | free_image(dc); |
| | | return single_filters; |
| | | } |
| | | |