From e9d287817ec2dfced1eb02641c696b2663b7e42d Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Wed, 28 Mar 2018 20:42:22 +0000
Subject: [PATCH] Minor fix

---
 src/cost_layer.c |  157 +++++++++++++++++++++++++++++++---------------------
 1 files changed, 94 insertions(+), 63 deletions(-)

diff --git a/src/cost_layer.c b/src/cost_layer.c
index 08d3bb5..39d2398 100644
--- a/src/cost_layer.c
+++ b/src/cost_layer.c
@@ -1,6 +1,7 @@
 #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>
@@ -9,8 +10,9 @@
 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;
 }
 
@@ -19,99 +21,128 @@
     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));
+    fprintf(stderr, "cost                                           %4d\n",  inputs);
+    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(inputs*batch, sizeof(float));
+    l.cost = calloc(1, sizeof(float));
+
+    l.forward = forward_cost_layer;
+    l.backward = backward_cost_layer;
     #ifdef GPU
-    layer->delta_cl = cl_make_array(layer->delta, inputs*batch);
+    l.forward_gpu = forward_cost_layer_gpu;
+    l.backward_gpu = backward_cost_layer_gpu;
+
+    l.delta_gpu = cuda_make_array(l.output, inputs*batch);
+    l.output_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));
+    l->output = realloc(l->output, inputs*l->batch*sizeof(float));
+#ifdef GPU
+    cuda_free(l->delta_gpu);
+    cuda_free(l->output_gpu);
+    l->delta_gpu = cuda_make_array(l->delta, inputs*l->batch);
+    l->output_gpu = cuda_make_array(l->output, 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);
-    //printf("cost: %f\n", *layer.output);
+    if(l.cost_type == SMOOTH){
+        smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
+    } else {
+        l2_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
+    }
+    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
 }
 
-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)
+int float_abs_compare (const void * a, const void * b)
 {
-    if (!truth) return;
+    float fa = *(const float*) a;
+    if(fa < 0) fa = -fa;
+    float fb = *(const float*) b;
+    if(fb < 0) fb = -fb;
+    return (fa > fb) - (fa < fb);
+}
 
-    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);
+void forward_cost_layer_gpu(cost_layer l, network_state state)
+{
+    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("cost: %f\n", *layer.output);
+    if(l.cost_type == SMOOTH){
+        smooth_l1_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
+    } else {
+        l2_gpu(l.batch*l.inputs, state.input, state.truth, l.delta_gpu, l.output_gpu);
+    }
+
+    if(l.ratio){
+        cuda_pull_array(l.delta_gpu, l.delta, l.batch*l.inputs);
+        qsort(l.delta, l.batch*l.inputs, sizeof(float), float_abs_compare);
+        int n = (1-l.ratio) * l.batch*l.inputs;
+        float thresh = l.delta[n];
+        thresh = 0;
+        printf("%f\n", thresh);
+        supp_ongpu(l.batch*l.inputs, thresh, l.delta_gpu, 1);
+    }
+
+    cuda_pull_array(l.output_gpu, l.output, l.batch*l.inputs);
+    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
 }
 
-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
 

--
Gitblit v1.10.0