| | |
| | | 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, float learning_rate, float momentum, float decay) |
| | | 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; |
| | | 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; |
| | |
| | | layer->biases = calloc(n, sizeof(float)); |
| | | layer->bias_updates = calloc(n, sizeof(float)); |
| | | float scale = 1./sqrt(size*size*c); |
| | | //scale = .05; |
| | | for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal(); |
| | | for(i = 0; i < n; ++i){ |
| | | //layer->biases[i] = rand_normal()*scale + scale; |
| | | layer->biases[i] = scale; |
| | | } |
| | | int out_h = convolutional_out_height(*layer); |
| | |
| | | return layer; |
| | | } |
| | | |
| | | void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c) |
| | | void resize_convolutional_layer(convolutional_layer *layer, int h, int w) |
| | | { |
| | | layer->h = h; |
| | | layer->w = w; |
| | | layer->c = c; |
| | | int out_h = convolutional_out_height(*layer); |
| | | int out_w = convolutional_out_width(*layer); |
| | | |
| | |
| | | layer->batch*out_h * out_w * layer->n*sizeof(float)); |
| | | layer->delta = realloc(layer->delta, |
| | | layer->batch*out_h * out_w * layer->n*sizeof(float)); |
| | | |
| | | #ifdef GPU |
| | | cuda_free(layer->col_image_gpu); |
| | | cuda_free(layer->delta_gpu); |
| | | cuda_free(layer->output_gpu); |
| | | |
| | | layer->col_image_gpu = cuda_make_array(layer->col_image, out_h*out_w*layer->size*layer->size*layer->c); |
| | | layer->delta_gpu = cuda_make_array(layer->delta, layer->batch*out_h*out_w*layer->n); |
| | | layer->output_gpu = cuda_make_array(layer->output, layer->batch*out_h*out_w*layer->n); |
| | | #endif |
| | | } |
| | | |
| | | void bias_output(const convolutional_layer layer) |
| | | void bias_output(float *output, float *biases, int batch, int n, int size) |
| | | { |
| | | int i,j,b; |
| | | int out_h = convolutional_out_height(layer); |
| | | int out_w = convolutional_out_width(layer); |
| | | for(b = 0; b < layer.batch; ++b){ |
| | | for(i = 0; i < layer.n; ++i){ |
| | | for(j = 0; j < out_h*out_w; ++j){ |
| | | layer.output[(b*layer.n + i)*out_h*out_w + j] = layer.biases[i]; |
| | | 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 forward_convolutional_layer(const convolutional_layer layer, float *in) |
| | | void backward_bias(float *bias_updates, float *delta, int batch, int n, int size) |
| | | { |
| | | float alpha = 1./batch; |
| | | int i,b; |
| | | for(b = 0; b < batch; ++b){ |
| | | for(i = 0; i < n; ++i){ |
| | | bias_updates[i] += alpha * sum_array(delta+size*(i+b*n), size); |
| | | } |
| | | } |
| | | } |
| | | |
| | | |
| | | void forward_convolutional_layer(const convolutional_layer layer, network_state state) |
| | | { |
| | | int out_h = convolutional_out_height(layer); |
| | | int out_w = convolutional_out_width(layer); |
| | | int i; |
| | | |
| | | bias_output(layer); |
| | | bias_output(layer.output, layer.biases, layer.batch, layer.n, out_h*out_w); |
| | | |
| | | int m = layer.n; |
| | | int k = layer.size*layer.size*layer.c; |
| | |
| | | float *c = layer.output; |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | im2col_cpu(in, layer.c, layer.h, layer.w, |
| | | im2col_cpu(state.input, 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.c*layer.h*layer.w; |
| | | state.input += layer.c*layer.h*layer.w; |
| | | } |
| | | activate_array(layer.output, m*n*layer.batch, layer.activation); |
| | | } |
| | | |
| | | void learn_bias_convolutional_layer(convolutional_layer layer) |
| | | void backward_convolutional_layer(convolutional_layer layer, network_state state) |
| | | { |
| | | int i,b; |
| | | int size = convolutional_out_height(layer) |
| | | *convolutional_out_width(layer); |
| | | for(b = 0; b < layer.batch; ++b){ |
| | | for(i = 0; i < layer.n; ++i){ |
| | | layer.bias_updates[i] += sum_array(layer.delta+size*(i+b*layer.n), size); |
| | | } |
| | | } |
| | | } |
| | | |
| | | void backward_convolutional_layer(convolutional_layer layer, float *in, float *delta) |
| | | { |
| | | float alpha = 1./layer.batch; |
| | | int i; |
| | | int m = layer.n; |
| | | int n = layer.size*layer.size*layer.c; |
| | |
| | | convolutional_out_width(layer); |
| | | |
| | | gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta); |
| | | backward_bias(layer.bias_updates, layer.delta, layer.batch, layer.n, k); |
| | | |
| | | learn_bias_convolutional_layer(layer); |
| | | |
| | | if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float)); |
| | | if(state.delta) memset(state.delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float)); |
| | | |
| | | for(i = 0; i < layer.batch; ++i){ |
| | | float *a = layer.delta + i*m*k; |
| | | float *b = layer.col_image; |
| | | float *c = layer.filter_updates; |
| | | |
| | | float *im = in+i*layer.c*layer.h*layer.w; |
| | | float *im = state.input+i*layer.c*layer.h*layer.w; |
| | | |
| | | im2col_cpu(im, layer.c, layer.h, layer.w, |
| | | layer.size, layer.stride, layer.pad, b); |
| | | gemm(0,1,m,n,k,1,a,k,b,k,1,c,n); |
| | | gemm(0,1,m,n,k,alpha,a,k,b,k,1,c,n); |
| | | |
| | | if(delta){ |
| | | if(state.delta){ |
| | | a = layer.filters; |
| | | b = layer.delta + i*m*k; |
| | | c = layer.col_image; |
| | | |
| | | gemm(1,0,n,k,m,1,a,n,b,k,0,c,k); |
| | | |
| | | col2im_cpu(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta+i*layer.c*layer.h*layer.w); |
| | | col2im_cpu(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, state.delta+i*layer.c*layer.h*layer.w); |
| | | } |
| | | } |
| | | } |
| | | |
| | | void update_convolutional_layer(convolutional_layer layer) |
| | | void update_convolutional_layer(convolutional_layer layer, float learning_rate, float momentum, float decay) |
| | | { |
| | | int size = layer.size*layer.size*layer.c*layer.n; |
| | | axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1); |
| | | scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1); |
| | | axpy_cpu(layer.n, learning_rate, layer.bias_updates, 1, layer.biases, 1); |
| | | scal_cpu(layer.n, momentum, layer.bias_updates, 1); |
| | | |
| | | axpy_cpu(size, -layer.decay, layer.filters, 1, layer.filter_updates, 1); |
| | | axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1); |
| | | scal_cpu(size, layer.momentum, layer.filter_updates, 1); |
| | | axpy_cpu(size, -decay, layer.filters, 1, layer.filter_updates, 1); |
| | | axpy_cpu(size, learning_rate, layer.filter_updates, 1, layer.filters, 1); |
| | | scal_cpu(size, momentum, layer.filter_updates, 1); |
| | | } |
| | | |
| | | |