From e6c97a53a7b5ac4014d30d236ea2bf5adb4bb521 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Tue, 07 Aug 2018 20:19:50 +0000
Subject: [PATCH] Maxpool fixes
---
src/cost_layer.c | 86 ++++++++++++++++++++++++++++++++++---------
1 files changed, 68 insertions(+), 18 deletions(-)
diff --git a/src/cost_layer.c b/src/cost_layer.c
index 24f6ffa..39d2398 100644
--- a/src/cost_layer.c
+++ b/src/cost_layer.c
@@ -11,7 +11,8 @@
{
if (strcmp(s, "sse")==0) return SSE;
if (strcmp(s, "masked")==0) return MASKED;
- fprintf(stderr, "Couldn't find activation function %s, going with SSE\n", s);
+ if (strcmp(s, "smooth")==0) return SMOOTH;
+ fprintf(stderr, "Couldn't find cost type %s, going with SSE\n", s);
return SSE;
}
@@ -22,46 +23,73 @@
return "sse";
case MASKED:
return "masked";
+ case SMOOTH:
+ return "smooth";
}
return "sse";
}
-cost_layer make_cost_layer(int batch, int inputs, COST_TYPE cost_type)
+cost_layer make_cost_layer(int batch, int inputs, COST_TYPE cost_type, float scale)
{
- fprintf(stderr, "Cost Layer: %d inputs\n", inputs);
+ 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(1, 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
- l.delta_gpu = cuda_make_array(l.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 l;
}
+void resize_cost_layer(cost_layer *l, int inputs)
+{
+ 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 < l.batch*l.inputs; ++i){
- if(state.truth[i] == 0) state.input[i] = 0;
+ if(state.truth[i] == SECRET_NUM) state.input[i] = SECRET_NUM;
}
}
- 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);
+ 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 l, network_state state)
{
- copy_cpu(l.batch*l.inputs, l.delta, 1, state.delta, 1);
+ axpy_cpu(l.batch*l.inputs, l.scale, l.delta, 1, state.delta, 1);
}
#ifdef GPU
@@ -76,23 +104,45 @@
cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs);
}
+int float_abs_compare (const void * a, const void * b)
+{
+ 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);
+}
+
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, state.truth);
+ mask_ongpu(l.batch*l.inputs, state.input, SECRET_NUM, state.truth);
}
-
- 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);
+ 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 l, network_state state)
{
- copy_ongpu(l.batch*l.inputs, l.delta_gpu, 1, state.delta, 1);
+ axpy_ongpu(l.batch*l.inputs, l.scale, l.delta_gpu, 1, state.delta, 1);
}
#endif
--
Gitblit v1.10.0