From cf32e7e9b843560eb7ec3ed16e5b19f0f7156724 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@burninator.cs.washington.edu>
Date: Sat, 25 Jun 2016 23:12:00 +0000
Subject: [PATCH] colors

---
 src/connected_layer.c |  335 ++++++++++++++++++++++++++++++++++++++++++++-----------
 1 files changed, 264 insertions(+), 71 deletions(-)

diff --git a/src/connected_layer.c b/src/connected_layer.c
index 792f20b..f20aa93 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,101 +1,294 @@
 #include "connected_layer.h"
+#include "batchnorm_layer.h"
 #include "utils.h"
-#include "mini_blas.h"
+#include "cuda.h"
+#include "blas.h"
+#include "gemm.h"
 
 #include <math.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
 
-connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation)
+connected_layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize)
 {
+    int i;
+    connected_layer l = {0};
+    l.type = CONNECTED;
+
+    l.inputs = inputs;
+    l.outputs = outputs;
+    l.batch=batch;
+    l.batch_normalize = batch_normalize;
+    l.h = 1;
+    l.w = 1;
+    l.c = inputs;
+    l.out_h = 1;
+    l.out_w = 1;
+    l.out_c = outputs;
+
+    l.output = calloc(batch*outputs, sizeof(float));
+    l.delta = calloc(batch*outputs, sizeof(float));
+
+    l.weight_updates = calloc(inputs*outputs, sizeof(float));
+    l.bias_updates = calloc(outputs, sizeof(float));
+
+    l.weights = calloc(outputs*inputs, sizeof(float));
+    l.biases = calloc(outputs, sizeof(float));
+
+    //float scale = 1./sqrt(inputs);
+    float scale = sqrt(2./inputs);
+    for(i = 0; i < outputs*inputs; ++i){
+        l.weights[i] = scale*rand_uniform(-1, 1);
+    }
+
+    for(i = 0; i < outputs; ++i){
+        l.biases[i] = 0;
+    }
+
+    if(batch_normalize){
+        l.scales = calloc(outputs, sizeof(float));
+        l.scale_updates = calloc(outputs, sizeof(float));
+        for(i = 0; i < outputs; ++i){
+            l.scales[i] = 1;
+        }
+
+        l.mean = calloc(outputs, sizeof(float));
+        l.mean_delta = calloc(outputs, sizeof(float));
+        l.variance = calloc(outputs, sizeof(float));
+        l.variance_delta = calloc(outputs, sizeof(float));
+
+        l.rolling_mean = calloc(outputs, sizeof(float));
+        l.rolling_variance = calloc(outputs, sizeof(float));
+
+        l.x = calloc(batch*outputs, sizeof(float));
+        l.x_norm = calloc(batch*outputs, sizeof(float));
+    }
+
+#ifdef GPU
+    l.weights_gpu = cuda_make_array(l.weights, outputs*inputs);
+    l.biases_gpu = cuda_make_array(l.biases, outputs);
+
+    l.weight_updates_gpu = cuda_make_array(l.weight_updates, outputs*inputs);
+    l.bias_updates_gpu = cuda_make_array(l.bias_updates, outputs);
+
+    l.output_gpu = cuda_make_array(l.output, outputs*batch);
+    l.delta_gpu = cuda_make_array(l.delta, outputs*batch);
+    if(batch_normalize){
+        l.scales_gpu = cuda_make_array(l.scales, outputs);
+        l.scale_updates_gpu = cuda_make_array(l.scale_updates, outputs);
+
+        l.mean_gpu = cuda_make_array(l.mean, outputs);
+        l.variance_gpu = cuda_make_array(l.variance, outputs);
+
+        l.rolling_mean_gpu = cuda_make_array(l.mean, outputs);
+        l.rolling_variance_gpu = cuda_make_array(l.variance, outputs);
+
+        l.mean_delta_gpu = cuda_make_array(l.mean, outputs);
+        l.variance_delta_gpu = cuda_make_array(l.variance, outputs);
+
+        l.x_gpu = cuda_make_array(l.output, l.batch*outputs);
+        l.x_norm_gpu = cuda_make_array(l.output, l.batch*outputs);
+    }
+#endif
+    l.activation = activation;
     fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
-    int i;
-    connected_layer *layer = calloc(1, sizeof(connected_layer));
-    layer->inputs = inputs;
-    layer->outputs = outputs;
-    layer->batch=batch;
-
-    layer->output = calloc(batch*outputs, sizeof(float*));
-    layer->delta = calloc(batch*outputs, sizeof(float*));
-
-    layer->weight_updates = calloc(inputs*outputs, sizeof(float));
-    layer->weight_adapt = calloc(inputs*outputs, sizeof(float));
-    layer->weight_momentum = calloc(inputs*outputs, sizeof(float));
-    layer->weights = calloc(inputs*outputs, sizeof(float));
-    float scale = 1./inputs;
-    for(i = 0; i < inputs*outputs; ++i)
-        layer->weights[i] = scale*(rand_uniform());
-
-    layer->bias_updates = calloc(outputs, sizeof(float));
-    layer->bias_adapt = calloc(outputs, sizeof(float));
-    layer->bias_momentum = calloc(outputs, sizeof(float));
-    layer->biases = calloc(outputs, sizeof(float));
-    for(i = 0; i < outputs; ++i)
-        //layer->biases[i] = rand_normal()*scale + scale;
-        layer->biases[i] = 1;
-
-    layer->activation = activation;
-    return layer;
+    return l;
 }
 
-void update_connected_layer(connected_layer layer, float step, float momentum, float decay)
+void update_connected_layer(connected_layer l, int batch, float learning_rate, float momentum, float decay)
 {
-    int i;
-    for(i = 0; i < layer.outputs; ++i){
-        layer.bias_momentum[i] = step*(layer.bias_updates[i]) + momentum*layer.bias_momentum[i];
-        layer.biases[i] += layer.bias_momentum[i];
+    axpy_cpu(l.outputs, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
+    scal_cpu(l.outputs, momentum, l.bias_updates, 1);
+
+    if(l.batch_normalize){
+        axpy_cpu(l.outputs, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
+        scal_cpu(l.outputs, momentum, l.scale_updates, 1);
     }
-    for(i = 0; i < layer.outputs*layer.inputs; ++i){
-        layer.weight_momentum[i] = step*(layer.weight_updates[i] - decay*layer.weights[i]) + momentum*layer.weight_momentum[i];
-        layer.weights[i] += layer.weight_momentum[i];
-    }
-    memset(layer.bias_updates, 0, layer.outputs*sizeof(float));
-    memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(float));
+
+    axpy_cpu(l.inputs*l.outputs, -decay*batch, l.weights, 1, l.weight_updates, 1);
+    axpy_cpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
+    scal_cpu(l.inputs*l.outputs, momentum, l.weight_updates, 1);
 }
 
-void forward_connected_layer(connected_layer layer, float *input)
+void forward_connected_layer(connected_layer l, network_state state)
 {
     int i;
-    memcpy(layer.output, layer.biases, layer.outputs*sizeof(float));
-    int m = layer.batch;
-    int k = layer.inputs;
-    int n = layer.outputs;
-    float *a = input;
-    float *b = layer.weights;
-    float *c = layer.output;
-    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
-    for(i = 0; i < layer.outputs*layer.batch; ++i){
-        layer.output[i] = activate(layer.output[i], layer.activation);
+    fill_cpu(l.outputs*l.batch, 0, l.output, 1);
+    int m = l.batch;
+    int k = l.inputs;
+    int n = l.outputs;
+    float *a = state.input;
+    float *b = l.weights;
+    float *c = l.output;
+    gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+    if(l.batch_normalize){
+        if(state.train){
+            mean_cpu(l.output, l.batch, l.outputs, 1, l.mean);
+            variance_cpu(l.output, l.mean, l.batch, l.outputs, 1, l.variance);
+
+            scal_cpu(l.outputs, .95, l.rolling_mean, 1);
+            axpy_cpu(l.outputs, .05, l.mean, 1, l.rolling_mean, 1);
+            scal_cpu(l.outputs, .95, l.rolling_variance, 1);
+            axpy_cpu(l.outputs, .05, l.variance, 1, l.rolling_variance, 1);
+
+            copy_cpu(l.outputs*l.batch, l.output, 1, l.x, 1);
+            normalize_cpu(l.output, l.mean, l.variance, l.batch, l.outputs, 1);   
+            copy_cpu(l.outputs*l.batch, l.output, 1, l.x_norm, 1);
+        } else {
+            normalize_cpu(l.output, l.rolling_mean, l.rolling_variance, l.batch, l.outputs, 1);
+        }
+        scale_bias(l.output, l.scales, l.batch, l.outputs, 1);
+    }
+    for(i = 0; i < l.batch; ++i){
+        axpy_cpu(l.outputs, 1, l.biases, 1, l.output + i*l.outputs, 1);
+    }
+    activate_array(l.output, l.outputs*l.batch, l.activation);
+}
+
+void backward_connected_layer(connected_layer l, network_state state)
+{
+    int i;
+    gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);
+    for(i = 0; i < l.batch; ++i){
+        axpy_cpu(l.outputs, 1, l.delta + i*l.outputs, 1, l.bias_updates, 1);
+    }
+    if(l.batch_normalize){
+        backward_scale_cpu(l.x_norm, l.delta, l.batch, l.outputs, 1, l.scale_updates);
+
+        scale_bias(l.delta, l.scales, l.batch, l.outputs, 1);
+
+        mean_delta_cpu(l.delta, l.variance, l.batch, l.outputs, 1, l.mean_delta);
+        variance_delta_cpu(l.x, l.delta, l.mean, l.variance, l.batch, l.outputs, 1, l.variance_delta);
+        normalize_delta_cpu(l.x, l.mean, l.variance, l.mean_delta, l.variance_delta, l.batch, l.outputs, 1, l.delta);
+    }
+
+    int m = l.outputs;
+    int k = l.batch;
+    int n = l.inputs;
+    float *a = l.delta;
+    float *b = state.input;
+    float *c = l.weight_updates;
+    gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
+
+    m = l.batch;
+    k = l.outputs;
+    n = l.inputs;
+
+    a = l.delta;
+    b = l.weights;
+    c = state.delta;
+
+    if(c) gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+}
+
+
+void denormalize_connected_layer(layer l)
+{
+    int i, j;
+    for(i = 0; i < l.outputs; ++i){
+        float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .00001);
+        for(j = 0; j < l.inputs; ++j){
+            l.weights[i*l.inputs + j] *= scale;
+        }
+        l.biases[i] -= l.rolling_mean[i] * scale;
     }
 }
 
-void learn_connected_layer(connected_layer layer, float *input)
+#ifdef GPU
+
+void pull_connected_layer(connected_layer l)
+{
+    cuda_pull_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
+    cuda_pull_array(l.biases_gpu, l.biases, l.outputs);
+    cuda_pull_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
+    cuda_pull_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+    if (l.batch_normalize){
+        cuda_pull_array(l.scales_gpu, l.scales, l.outputs);
+        cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
+        cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
+    }
+}
+
+void push_connected_layer(connected_layer l)
+{
+    cuda_push_array(l.weights_gpu, l.weights, l.inputs*l.outputs);
+    cuda_push_array(l.biases_gpu, l.biases, l.outputs);
+    cuda_push_array(l.weight_updates_gpu, l.weight_updates, l.inputs*l.outputs);
+    cuda_push_array(l.bias_updates_gpu, l.bias_updates, l.outputs);
+    if (l.batch_normalize){
+        cuda_push_array(l.scales_gpu, l.scales, l.outputs);
+        cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.outputs);
+        cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.outputs);
+    }
+}
+
+void update_connected_layer_gpu(connected_layer l, int batch, float learning_rate, float momentum, float decay)
+{
+    axpy_ongpu(l.outputs, learning_rate/batch, l.bias_updates_gpu, 1, l.biases_gpu, 1);
+    scal_ongpu(l.outputs, momentum, l.bias_updates_gpu, 1);
+
+    if(l.batch_normalize){
+        axpy_ongpu(l.outputs, learning_rate/batch, l.scale_updates_gpu, 1, l.scales_gpu, 1);
+        scal_ongpu(l.outputs, momentum, l.scale_updates_gpu, 1);
+    }
+
+    axpy_ongpu(l.inputs*l.outputs, -decay*batch, l.weights_gpu, 1, l.weight_updates_gpu, 1);
+    axpy_ongpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates_gpu, 1, l.weights_gpu, 1);
+    scal_ongpu(l.inputs*l.outputs, momentum, l.weight_updates_gpu, 1);
+}
+
+void forward_connected_layer_gpu(connected_layer l, network_state state)
 {
     int i;
-    for(i = 0; i < layer.outputs*layer.batch; ++i){
-        layer.delta[i] *= gradient(layer.output[i], layer.activation);
-        layer.bias_updates[i%layer.batch] += layer.delta[i]/layer.batch;
+    fill_ongpu(l.outputs*l.batch, 0, l.output_gpu, 1);
+
+    int m = l.batch;
+    int k = l.inputs;
+    int n = l.outputs;
+    float * a = state.input;
+    float * b = l.weights_gpu;
+    float * c = l.output_gpu;
+    gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
+    if(l.batch_normalize){
+        forward_batchnorm_layer_gpu(l, state);
     }
-    int m = layer.inputs;
-    int k = layer.batch;
-    int n = layer.outputs;
-    float *a = input;
-    float *b = layer.delta;
-    float *c = layer.weight_updates;
-    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+    for(i = 0; i < l.batch; ++i){
+        axpy_ongpu(l.outputs, 1, l.biases_gpu, 1, l.output_gpu + i*l.outputs, 1);
+    }
+    activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);
+
 }
 
-void backward_connected_layer(connected_layer layer, float *input, float *delta)
+void backward_connected_layer_gpu(connected_layer l, network_state state)
 {
-    int m = layer.inputs;
-    int k = layer.outputs;
-    int n = layer.batch;
+    int i;
+    constrain_ongpu(l.outputs*l.batch, 5, l.delta_gpu, 1);
+    gradient_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
+    for(i = 0; i < l.batch; ++i){
+        axpy_ongpu(l.outputs, 1, l.delta_gpu + i*l.outputs, 1, l.bias_updates_gpu, 1);
+    }
 
-    float *a = layer.weights;
-    float *b = layer.delta;
-    float *c = delta;
+    if(l.batch_normalize){
+        backward_batchnorm_layer_gpu(l, state);
+    }
 
-    gemm(0,0,m,n,k,1,a,k,b,n,0,c,n);
+    int m = l.outputs;
+    int k = l.batch;
+    int n = l.inputs;
+    float * a = l.delta_gpu;
+    float * b = state.input;
+    float * c = l.weight_updates_gpu;
+    gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);
+
+    m = l.batch;
+    k = l.outputs;
+    n = l.inputs;
+
+    a = l.delta_gpu;
+    b = l.weights_gpu;
+    c = state.delta;
+
+    if(c) gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
 }
-
+#endif

--
Gitblit v1.10.0