| | |
| | | #include "cost_layer.h" |
| | | #include "utils.h" |
| | | #include "mini_blas.h" |
| | | #include "cuda.h" |
| | | #include "blas.h" |
| | | #include <math.h> |
| | | #include <string.h> |
| | | #include <stdlib.h> |
| | |
| | | COST_TYPE get_cost_type(char *s) |
| | | { |
| | | if (strcmp(s, "sse")==0) return SSE; |
| | | if (strcmp(s, "detection")==0) return DETECTION; |
| | | fprintf(stderr, "Couldn't find activation function %s, going with SSE\n", s); |
| | | if (strcmp(s, "masked")==0) return MASKED; |
| | | if (strcmp(s, "smooth")==0) return SMOOTH; |
| | | fprintf(stderr, "Couldn't find cost type %s, going with SSE\n", s); |
| | | return SSE; |
| | | } |
| | | |
| | |
| | | switch(a){ |
| | | case SSE: |
| | | return "sse"; |
| | | case DETECTION: |
| | | return "detection"; |
| | | case MASKED: |
| | | return "masked"; |
| | | case SMOOTH: |
| | | return "smooth"; |
| | | } |
| | | return "sse"; |
| | | } |
| | | |
| | | cost_layer *make_cost_layer(int batch, int inputs, COST_TYPE type) |
| | | cost_layer make_cost_layer(int batch, int inputs, COST_TYPE cost_type, float scale) |
| | | { |
| | | fprintf(stderr, "Cost Layer: %d inputs\n", inputs); |
| | | cost_layer *layer = calloc(1, sizeof(cost_layer)); |
| | | layer->batch = batch; |
| | | layer->inputs = inputs; |
| | | layer->type = type; |
| | | layer->delta = calloc(inputs*batch, sizeof(float)); |
| | | layer->output = calloc(1, sizeof(float)); |
| | | cost_layer l = {0}; |
| | | l.type = COST; |
| | | |
| | | l.scale = scale; |
| | | l.batch = batch; |
| | | l.inputs = inputs; |
| | | l.outputs = inputs; |
| | | l.cost_type = cost_type; |
| | | l.delta = calloc(inputs*batch, sizeof(float)); |
| | | l.output = calloc(1, sizeof(float)); |
| | | #ifdef GPU |
| | | layer->delta_cl = cl_make_array(layer->delta, inputs*batch); |
| | | l.delta_gpu = cuda_make_array(l.delta, inputs*batch); |
| | | #endif |
| | | return layer; |
| | | return l; |
| | | } |
| | | |
| | | void forward_cost_layer(cost_layer layer, float *input, float *truth) |
| | | void resize_cost_layer(cost_layer *l, int inputs) |
| | | { |
| | | if (!truth) return; |
| | | copy_cpu(layer.batch*layer.inputs, truth, 1, layer.delta, 1); |
| | | axpy_cpu(layer.batch*layer.inputs, -1, input, 1, layer.delta, 1); |
| | | if(layer.type == DETECTION){ |
| | | l->inputs = inputs; |
| | | l->outputs = inputs; |
| | | l->delta = realloc(l->delta, inputs*l->batch*sizeof(float)); |
| | | #ifdef GPU |
| | | cuda_free(l->delta_gpu); |
| | | l->delta_gpu = cuda_make_array(l->delta, inputs*l->batch); |
| | | #endif |
| | | } |
| | | |
| | | void forward_cost_layer(cost_layer l, network_state state) |
| | | { |
| | | if (!state.truth) return; |
| | | if(l.cost_type == MASKED){ |
| | | int i; |
| | | for(i = 0; i < layer.batch*layer.inputs; ++i){ |
| | | if((i%5) && !truth[(i/5)*5]) layer.delta[i] = 0; |
| | | for(i = 0; i < l.batch*l.inputs; ++i){ |
| | | if(state.truth[i] == SECRET_NUM) state.input[i] = SECRET_NUM; |
| | | } |
| | | } |
| | | *(layer.output) = dot_cpu(layer.batch*layer.inputs, layer.delta, 1, layer.delta, 1); |
| | | if(l.cost_type == SMOOTH){ |
| | | smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta); |
| | | } else { |
| | | copy_cpu(l.batch*l.inputs, state.truth, 1, l.delta, 1); |
| | | axpy_cpu(l.batch*l.inputs, -1, state.input, 1, l.delta, 1); |
| | | } |
| | | *(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1); |
| | | //printf("cost: %f\n", *l.output); |
| | | } |
| | | |
| | | void backward_cost_layer(const cost_layer layer, float *input, float *delta) |
| | | void backward_cost_layer(const cost_layer l, network_state state) |
| | | { |
| | | copy_cpu(layer.batch*layer.inputs, layer.delta, 1, delta, 1); |
| | | axpy_cpu(l.batch*l.inputs, l.scale, l.delta, 1, state.delta, 1); |
| | | } |
| | | |
| | | #ifdef GPU |
| | | |
| | | cl_kernel get_mask_kernel() |
| | | void pull_cost_layer(cost_layer l) |
| | | { |
| | | static int init = 0; |
| | | static cl_kernel kernel; |
| | | if(!init){ |
| | | kernel = get_kernel("src/axpy.cl", "mask", 0); |
| | | init = 1; |
| | | } |
| | | return kernel; |
| | | cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs); |
| | | } |
| | | |
| | | void mask_ongpu(int n, cl_mem x, cl_mem mask, int mod) |
| | | void push_cost_layer(cost_layer l) |
| | | { |
| | | cl_setup(); |
| | | cl_kernel kernel = get_mask_kernel(); |
| | | cl_command_queue queue = cl.queue; |
| | | |
| | | cl_uint i = 0; |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(n), (void*) &n); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(x), (void*) &x); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(mask), (void*) &mask); |
| | | cl.error = clSetKernelArg(kernel, i++, sizeof(mod), (void*) &mod); |
| | | check_error(cl); |
| | | |
| | | const size_t global_size[] = {n}; |
| | | |
| | | cl.error = clEnqueueNDRangeKernel(queue, kernel, 1, 0, global_size, 0, 0, 0, 0); |
| | | check_error(cl); |
| | | |
| | | cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs); |
| | | } |
| | | |
| | | void forward_cost_layer_gpu(cost_layer layer, cl_mem input, cl_mem truth) |
| | | void forward_cost_layer_gpu(cost_layer l, network_state state) |
| | | { |
| | | if (!truth) return; |
| | | |
| | | copy_ongpu(layer.batch*layer.inputs, truth, 1, layer.delta_cl, 1); |
| | | axpy_ongpu(layer.batch*layer.inputs, -1, input, 1, layer.delta_cl, 1); |
| | | |
| | | if(layer.type==DETECTION){ |
| | | mask_ongpu(layer.inputs*layer.batch, layer.delta_cl, truth, 5); |
| | | if (!state.truth) return; |
| | | if (l.cost_type == MASKED) { |
| | | mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth); |
| | | } |
| | | |
| | | cl_read_array(layer.delta_cl, layer.delta, layer.batch*layer.inputs); |
| | | *(layer.output) = dot_cpu(layer.batch*layer.inputs, layer.delta, 1, layer.delta, 1); |
| | | //printf("%f\n", *layer.output); |
| | | if(l.cost_type == SMOOTH){ |
| | | smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu); |
| | | } else { |
| | | copy_ongpu(l.batch*l.inputs, state.truth, 1, l.delta_gpu, 1); |
| | | axpy_ongpu(l.batch*l.inputs, -1, state.input, 1, l.delta_gpu, 1); |
| | | } |
| | | |
| | | cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs); |
| | | *(l.output) = dot_cpu(l.batch*l.inputs, l.delta, 1, l.delta, 1); |
| | | } |
| | | |
| | | void backward_cost_layer_gpu(const cost_layer layer, cl_mem input, cl_mem delta) |
| | | void backward_cost_layer_gpu(const cost_layer l, network_state state) |
| | | { |
| | | copy_ongpu(layer.batch*layer.inputs, layer.delta_cl, 1, delta, 1); |
| | | axpy_ongpu(l.batch*l.inputs, l.scale, l.delta_gpu, 1, state.delta, 1); |
| | | } |
| | | #endif |
| | | |