From 9c2c220c88aff4d0bcbdd5b03b10c6d1a7db56d3 Mon Sep 17 00:00:00 2001
From: Edmond Yoo <hj3yoo@uwaterloo.ca>
Date: Sun, 16 Sep 2018 03:14:14 +0000
Subject: [PATCH] Moving files from MTGCardDetector #2
---
src/connected_layer.c | 357 +++++++++++++++++++++++++++++++++++++++++++++++++----------
1 files changed, 295 insertions(+), 62 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index d77a10c..e6dc759 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,97 +1,330 @@
#include "connected_layer.h"
+#include "batchnorm_layer.h"
+#include "utils.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 inputs, int outputs, ACTIVATION activator)
+connected_layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize)
{
- printf("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;
+ connected_layer l = {0};
+ l.type = CONNECTED;
- layer->output = calloc(outputs, sizeof(double*));
- layer->delta = calloc(outputs, sizeof(double*));
+ 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;
- layer->weight_updates = calloc(inputs*outputs, sizeof(double));
- layer->weight_momentum = calloc(inputs*outputs, sizeof(double));
- layer->weights = calloc(inputs*outputs, sizeof(double));
- for(i = 0; i < inputs*outputs; ++i)
- layer->weights[i] = .01*(.5 - (double)rand()/RAND_MAX);
+ l.output = calloc(batch*outputs, sizeof(float));
+ l.delta = calloc(batch*outputs, sizeof(float));
- layer->bias_updates = calloc(outputs, sizeof(double));
- layer->bias_momentum = calloc(outputs, sizeof(double));
- layer->biases = calloc(outputs, sizeof(double));
- for(i = 0; i < outputs; ++i)
- layer->biases[i] = 1;
+ l.weight_updates = calloc(inputs*outputs, sizeof(float));
+ l.bias_updates = calloc(outputs, sizeof(float));
- if(activator == SIGMOID){
- layer->activation = sigmoid_activation;
- layer->gradient = sigmoid_gradient;
- }else if(activator == RELU){
- layer->activation = relu_activation;
- layer->gradient = relu_gradient;
- }else if(activator == IDENTITY){
- layer->activation = identity_activation;
- layer->gradient = identity_gradient;
+ l.weights = calloc(outputs*inputs, sizeof(float));
+ l.biases = calloc(outputs, sizeof(float));
+
+ l.forward = forward_connected_layer;
+ l.backward = backward_connected_layer;
+ l.update = update_connected_layer;
+
+ //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);
}
- return layer;
+ 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.forward_gpu = forward_connected_layer_gpu;
+ l.backward_gpu = backward_connected_layer_gpu;
+ l.update_gpu = update_connected_layer_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);
+#ifdef CUDNN
+ cudnnCreateTensorDescriptor(&l.normTensorDesc);
+ cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+ cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w);
+ cudnnSetTensor4dDescriptor(l.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l.out_c, 1, 1);
+#endif
+ }
+#endif
+ l.activation = activation;
+ fprintf(stderr, "connected %4d -> %4d\n", inputs, outputs);
+ return l;
}
-void forward_connected_layer(connected_layer layer, double *input)
+void update_connected_layer(connected_layer l, int batch, float learning_rate, float momentum, float decay)
+{
+ 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);
+ }
+
+ 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 l, network_state state)
+{
+ int i;
+ 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 < layer.outputs; ++i){
- layer.output[i] = layer.biases[i];
- for(j = 0; j < layer.inputs; ++j){
- layer.output[i] += input[j]*layer.weights[i*layer.inputs + j];
+ for(i = 0; i < l.outputs; ++i){
+ float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .000001);
+ for(j = 0; j < l.inputs; ++j){
+ l.weights[i*l.inputs + j] *= scale;
}
- layer.output[i] = layer.activation(layer.output[i]);
+ l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
}
}
-void learn_connected_layer(connected_layer layer, double *input)
+
+void statistics_connected_layer(layer l)
{
- int i, j;
- for(i = 0; i < layer.outputs; ++i){
- layer.bias_updates[i] += layer.delta[i];
- for(j = 0; j < layer.inputs; ++j){
- layer.weight_updates[i*layer.inputs + j] += layer.delta[i]*input[j];
- }
+ if(l.batch_normalize){
+ printf("Scales ");
+ print_statistics(l.scales, l.outputs);
+ /*
+ printf("Rolling Mean ");
+ print_statistics(l.rolling_mean, l.outputs);
+ printf("Rolling Variance ");
+ print_statistics(l.rolling_variance, l.outputs);
+ */
}
+ printf("Biases ");
+ print_statistics(l.biases, l.outputs);
+ printf("Weights ");
+ print_statistics(l.weights, l.outputs);
}
-void update_connected_layer(connected_layer layer, double step, double momentum, double decay)
+#ifdef GPU
+
+void pull_connected_layer(connected_layer l)
{
- int i,j;
- for(i = 0; i < layer.outputs; ++i){
- layer.bias_momentum[i] = step*(layer.bias_updates[i] - decay*layer.biases[i]) + momentum*layer.bias_momentum[i];
- layer.biases[i] += layer.bias_momentum[i];
- for(j = 0; j < layer.inputs; ++j){
- int index = i*layer.inputs+j;
- layer.weight_momentum[index] = step*(layer.weight_updates[index] - decay*layer.weights[index]) + momentum*layer.weight_momentum[index];
- layer.weights[index] += layer.weight_momentum[index];
- }
+ 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);
}
- memset(layer.bias_updates, 0, layer.outputs*sizeof(double));
- memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(double));
}
-void backward_connected_layer(connected_layer layer, double *input, double *delta)
+void push_connected_layer(connected_layer l)
{
- int i, j;
-
- for(j = 0; j < layer.inputs; ++j){
- double grad = layer.gradient(input[j]);
- delta[j] = 0;
- for(i = 0; i < layer.outputs; ++i){
- delta[j] += layer.delta[i]*layer.weights[i*layer.inputs + j];
- }
- delta[j] *= grad;
+ 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;
+ 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);
+ }
+ else {
+ add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.outputs, 1);
+ }
+ //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_gpu(connected_layer l, network_state state)
+{
+ int i;
+ constrain_ongpu(l.outputs*l.batch, 1, 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);
+ }
+
+ if(l.batch_normalize){
+ backward_batchnorm_layer_gpu(l, state);
+ }
+
+ 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