From ae4ec1a06d5b4c9bcb5647cf555aafd4e5f800f7 Mon Sep 17 00:00:00 2001
From: Bartek Gąsiorzewski <g.bartek@gmail.com>
Date: Mon, 07 May 2018 20:11:57 +0000
Subject: [PATCH] Exit with nonzero status on error
---
src/convolutional_layer.c | 524 ++++++++++++++++++++++++++++++++++++++++++---------------
1 files changed, 386 insertions(+), 138 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 159951d..9a76bdf 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,5 +1,6 @@
#include "convolutional_layer.h"
#include "utils.h"
+#include "batchnorm_layer.h"
#include "im2col.h"
#include "col2im.h"
#include "blas.h"
@@ -7,20 +8,78 @@
#include <stdio.h>
#include <time.h>
+#ifdef CUDNN
+#pragma comment(lib, "cudnn.lib")
+#endif
+
+#ifdef AI2
+#include "xnor_layer.h"
+#endif
+
+#ifndef AI2
+#define AI2 0
+void forward_xnor_layer(layer l, network_state state);
+#endif
+
+void swap_binary(convolutional_layer *l)
+{
+ float *swap = l->weights;
+ l->weights = l->binary_weights;
+ l->binary_weights = swap;
+
+ #ifdef GPU
+ swap = l->weights_gpu;
+ l->weights_gpu = l->binary_weights_gpu;
+ l->binary_weights_gpu = swap;
+ #endif
+}
+
+void binarize_weights(float *weights, int n, int size, float *binary)
+{
+ int i, f;
+ for(f = 0; f < n; ++f){
+ float mean = 0;
+ for(i = 0; i < size; ++i){
+ mean += fabs(weights[f*size + i]);
+ }
+ mean = mean / size;
+ for(i = 0; i < size; ++i){
+ binary[f*size + i] = (weights[f*size + i] > 0) ? mean : -mean;
+ }
+ }
+}
+
+void binarize_cpu(float *input, int n, float *binary)
+{
+ int i;
+ for(i = 0; i < n; ++i){
+ binary[i] = (input[i] > 0) ? 1 : -1;
+ }
+}
+
+void binarize_input(float *input, int n, int size, float *binary)
+{
+ int i, s;
+ for(s = 0; s < size; ++s){
+ float mean = 0;
+ for(i = 0; i < n; ++i){
+ mean += fabs(input[i*size + s]);
+ }
+ mean = mean / n;
+ for(i = 0; i < n; ++i){
+ binary[i*size + s] = (input[i*size + s] > 0) ? mean : -mean;
+ }
+ }
+}
+
int convolutional_out_height(convolutional_layer l)
{
- int h = l.h;
- if (!l.pad) h -= l.size;
- else h -= 1;
- return h/l.stride + 1;
+ return (l.h + 2*l.pad - l.size) / l.stride + 1;
}
int convolutional_out_width(convolutional_layer l)
{
- int w = l.w;
- if (!l.pad) w -= l.size;
- else w -= 1;
- return w/l.stride + 1;
+ return (l.w + 2*l.pad - l.size) / l.stride + 1;
}
image get_convolutional_image(convolutional_layer l)
@@ -41,65 +100,163 @@
return float_to_image(w,h,c,l.delta);
}
-void backward_scale_cpu(float *x_norm, float *delta, int batch, int n, int size, float *scale_updates)
-{
- int i,b,f;
- for(f = 0; f < n; ++f){
- float sum = 0;
- for(b = 0; b < batch; ++b){
- for(i = 0; i < size; ++i){
- int index = i + size*(f + n*b);
- sum += delta[index] * x_norm[index];
- }
- }
- scale_updates[f] += sum;
+size_t get_workspace_size(layer l){
+#ifdef CUDNN
+ if(gpu_index >= 0){
+ size_t most = 0;
+ size_t s = 0;
+ cudnnGetConvolutionForwardWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.weightDesc,
+ l.convDesc,
+ l.dstTensorDesc,
+ l.fw_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardFilterWorkspaceSize(cudnn_handle(),
+ l.srcTensorDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dweightDesc,
+ l.bf_algo,
+ &s);
+ if (s > most) most = s;
+ cudnnGetConvolutionBackwardDataWorkspaceSize(cudnn_handle(),
+ l.weightDesc,
+ l.ddstTensorDesc,
+ l.convDesc,
+ l.dsrcTensorDesc,
+ l.bd_algo,
+ &s);
+ if (s > most) most = s;
+ return most;
}
+ #endif
+ return (size_t)l.out_h*l.out_w*l.size*l.size*l.c*sizeof(float);
}
-void mean_delta_cpu(float *delta, float *variance, int batch, int filters, int spatial, float *mean_delta)
+#ifdef GPU
+#ifdef CUDNN
+void cudnn_convolutional_setup(layer *l, int cudnn_preference)
{
- int i,j,k;
- for(i = 0; i < filters; ++i){
- mean_delta[i] = 0;
- for (j = 0; j < batch; ++j) {
- for (k = 0; k < spatial; ++k) {
- int index = j*filters*spatial + i*spatial + k;
- mean_delta[i] += delta[index];
- }
- }
- mean_delta[i] *= (-1./sqrt(variance[i] + .00001f));
- }
-}
-void variance_delta_cpu(float *x, float *delta, float *mean, float *variance, int batch, int filters, int spatial, float *variance_delta)
-{
+#ifdef CUDNN_HALF
+ // TRUE_HALF_CONFIG is only supported on architectures with true fp16 support (compute capability 5.3 and 6.0):
+ // Tegra X1, Jetson TX1, DRIVE CX, DRIVE PX, Quadro GP100, Tesla P100
+ // PSEUDO_HALF_CONFIG is required for Tensor Cores - our case!
+ const cudnnDataType_t data_type = CUDNN_DATA_HALF;
+#else
+ cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
+#endif
- int i,j,k;
- for(i = 0; i < filters; ++i){
- variance_delta[i] = 0;
- for(j = 0; j < batch; ++j){
- for(k = 0; k < spatial; ++k){
- int index = j*filters*spatial + i*spatial + k;
- variance_delta[i] += delta[index]*(x[index] - mean[i]);
- }
- }
- variance_delta[i] *= -.5 * pow(variance[i] + .00001f, (float)(-3./2.));
- }
-}
-void normalize_delta_cpu(float *x, float *mean, float *variance, float *mean_delta, float *variance_delta, int batch, int filters, int spatial, float *delta)
-{
- int f, j, k;
- for(j = 0; j < batch; ++j){
- for(f = 0; f < filters; ++f){
- for(k = 0; k < spatial; ++k){
- int index = j*filters*spatial + f*spatial + k;
- delta[index] = delta[index] * 1./(sqrt(variance[f]) + .00001f) + variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) + mean_delta[f]/(spatial*batch);
- }
- }
- }
-}
+#if(CUDNN_MAJOR >= 7)
+ // Tensor Core uses CUDNN_TENSOR_OP_MATH instead of CUDNN_DEFAULT_MATH
+ // For *_ALGO_WINOGRAD_NONFUSED can be used CUDNN_DATA_FLOAT
+ // otherwise Input, Filter and Output descriptors (xDesc, yDesc, wDesc, dxDesc, dyDesc and dwDesc as applicable) have dataType = CUDNN_DATA_HALF
+ // Three techniques for training using Mixed-precision: https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/
+ // 1. Accumulation into FP32
+ // 2. Loss Scaling - required only for: activation gradients. We do not use.
+ // 3. FP32 Master Copy of Weights
+ // More: http://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html#tensor_ops
+ cudnnSetConvolutionMathType(l->convDesc, CUDNN_TENSOR_OP_MATH);
+#endif
-convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int pad, ACTIVATION activation, int batch_normalize, int binary)
+ // INT8_CONFIG, INT8_EXT_CONFIG, INT8x4_CONFIG and INT8x4_EXT_CONFIG are only supported
+ // on architectures with DP4A support (compute capability 6.1 and later).
+ //cudnnDataType_t data_type = CUDNN_DATA_INT8;
+
+ // backward delta
+ cudnnSetTensor4dDescriptor(l->dsrcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->ddstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->dweightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ // forward
+ cudnnSetTensor4dDescriptor(l->srcTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->c, l->h, l->w);
+ cudnnSetTensor4dDescriptor(l->dstTensorDesc, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+ cudnnSetFilter4dDescriptor(l->weightDesc, data_type, CUDNN_TENSOR_NCHW, l->n, l->c, l->size, l->size);
+
+ // batch norm
+ cudnnSetTensor4dDescriptor(l->normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l->out_c, 1, 1);
+ cudnnSetTensor4dDescriptor(l->normDstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l->batch, l->out_c, l->out_h, l->out_w);
+
+ cudnnSetTensor4dDescriptor(l->normDstTensorDescF16, CUDNN_TENSOR_NCHW, data_type, l->batch, l->out_c, l->out_h, l->out_w);
+#if(CUDNN_MAJOR >= 6)
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION, CUDNN_DATA_FLOAT); // cudnn >= 6.0
+#else
+ cudnnSetConvolution2dDescriptor(l->convDesc, l->pad, l->pad, l->stride, l->stride, 1, 1, CUDNN_CROSS_CORRELATION); // cudnn 5.1
+#endif
+ int forward_algo = CUDNN_CONVOLUTION_FWD_PREFER_FASTEST;
+ int backward_algo = CUDNN_CONVOLUTION_BWD_DATA_PREFER_FASTEST;
+ int backward_filter = CUDNN_CONVOLUTION_BWD_FILTER_PREFER_FASTEST;
+ if (cudnn_preference == cudnn_smallest)
+ {
+ forward_algo = CUDNN_CONVOLUTION_FWD_NO_WORKSPACE;
+ backward_algo = CUDNN_CONVOLUTION_BWD_DATA_NO_WORKSPACE;
+ backward_filter = CUDNN_CONVOLUTION_BWD_FILTER_NO_WORKSPACE;
+ printf(" CUDNN-slow ");
+ }
+
+ cudnnGetConvolutionForwardAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->weightDesc,
+ l->convDesc,
+ l->dstTensorDesc,
+ forward_algo,
+ 0,
+ &l->fw_algo);
+ cudnnGetConvolutionBackwardDataAlgorithm(cudnn_handle(),
+ l->weightDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dsrcTensorDesc,
+ backward_algo,
+ 0,
+ &l->bd_algo);
+ cudnnGetConvolutionBackwardFilterAlgorithm(cudnn_handle(),
+ l->srcTensorDesc,
+ l->ddstTensorDesc,
+ l->convDesc,
+ l->dweightDesc,
+ backward_filter,
+ 0,
+ &l->bf_algo);
+
+ if (data_type == CUDNN_DATA_HALF)
+ {
+ // HALF-16 if(data_type == CUDNN_DATA_HALF)
+ l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
+ l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
+ l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
+
+ // FLOAT-32 if(data_type == CUDNN_DATA_FLOAT)
+ //l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED;
+ //l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED;
+ //l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED;
+
+ int fw = 0, bd = 0, bf = 0;
+ if (l->fw_algo == CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM) fw = 1;
+ //printf("Tensor Cores - Forward enabled: l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM \n");
+ if (l->fw_algo == CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED) fw = 2;
+ //printf("Tensor Cores - Forward enabled: l->fw_algo = CUDNN_CONVOLUTION_FWD_ALGO_WINOGRAD_NONFUSED \n");
+
+ if (l->bd_algo == CUDNN_CONVOLUTION_BWD_DATA_ALGO_1) bd = 1;
+ //printf("Tensor Cores - Backward-data enabled: l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1 \n");
+ if (l->bd_algo == CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED) bd = 2;
+ //printf("Tensor Cores - Backward-data enabled: l->bd_algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_WINOGRAD_NONFUSED \n");
+
+ if (l->bf_algo == CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1) bf = 1;
+ //printf("Tensor Cores - Backward-filter enabled: l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1 \n");
+ if (l->bf_algo == CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED) bf = 2;
+ //printf("Tensor Cores - Backward-filter enabled: l->bf_algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_WINOGRAD_NONFUSED \n");
+
+ if (fw == 2 && bd == 2 && bf == 2) printf("TF ");
+ else if (fw == 1 && bd == 1 && bf == 1) printf("TH ");
+ }
+}
+#endif
+#endif
+
+convolutional_layer make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, int padding, ACTIVATION activation, int batch_normalize, int binary, int xnor, int adam)
{
int i;
convolutional_layer l = {0};
@@ -110,21 +267,22 @@
l.c = c;
l.n = n;
l.binary = binary;
+ l.xnor = xnor;
l.batch = batch;
l.stride = stride;
l.size = size;
- l.pad = pad;
+ l.pad = padding;
l.batch_normalize = batch_normalize;
- l.filters = calloc(c*n*size*size, sizeof(float));
- l.filter_updates = calloc(c*n*size*size, sizeof(float));
+ l.weights = calloc(c*n*size*size, sizeof(float));
+ l.weight_updates = calloc(c*n*size*size, sizeof(float));
l.biases = calloc(n, sizeof(float));
l.bias_updates = calloc(n, sizeof(float));
// float scale = 1./sqrt(size*size*c);
float scale = sqrt(2./(size*size*c));
- for(i = 0; i < c*n*size*size; ++i) l.filters[i] = scale*rand_uniform(-1, 1);
+ for(i = 0; i < c*n*size*size; ++i) l.weights[i] = scale*rand_uniform(-1, 1);
int out_h = convolutional_out_height(l);
int out_w = convolutional_out_width(l);
l.out_h = out_h;
@@ -133,12 +291,20 @@
l.outputs = l.out_h * l.out_w * l.out_c;
l.inputs = l.w * l.h * l.c;
- l.col_image = calloc(out_h*out_w*size*size*c, sizeof(float));
- l.output = calloc(l.batch*out_h * out_w * n, sizeof(float));
- l.delta = calloc(l.batch*out_h * out_w * n, sizeof(float));
+ l.output = calloc(l.batch*l.outputs, sizeof(float));
+ l.delta = calloc(l.batch*l.outputs, sizeof(float));
+ l.forward = forward_convolutional_layer;
+ l.backward = backward_convolutional_layer;
+ l.update = update_convolutional_layer;
if(binary){
- l.binary_filters = calloc(c*n*size*size, sizeof(float));
+ l.binary_weights = calloc(c*n*size*size, sizeof(float));
+ l.cweights = calloc(c*n*size*size, sizeof(char));
+ l.scales = calloc(n, sizeof(float));
+ }
+ if(xnor){
+ l.binary_weights = calloc(c*n*size*size, sizeof(float));
+ l.binary_input = calloc(l.inputs*l.batch, sizeof(float));
}
if(batch_normalize){
@@ -151,45 +317,87 @@
l.mean = calloc(n, sizeof(float));
l.variance = calloc(n, sizeof(float));
+ l.mean_delta = calloc(n, sizeof(float));
+ l.variance_delta = calloc(n, sizeof(float));
+
l.rolling_mean = calloc(n, sizeof(float));
l.rolling_variance = calloc(n, sizeof(float));
+ l.x = calloc(l.batch*l.outputs, sizeof(float));
+ l.x_norm = calloc(l.batch*l.outputs, sizeof(float));
+ }
+ if(adam){
+ l.adam = 1;
+ l.m = calloc(c*n*size*size, sizeof(float));
+ l.v = calloc(c*n*size*size, sizeof(float));
}
#ifdef GPU
- l.filters_gpu = cuda_make_array(l.filters, c*n*size*size);
- l.filter_updates_gpu = cuda_make_array(l.filter_updates, c*n*size*size);
+ l.forward_gpu = forward_convolutional_layer_gpu;
+ l.backward_gpu = backward_convolutional_layer_gpu;
+ l.update_gpu = update_convolutional_layer_gpu;
- l.biases_gpu = cuda_make_array(l.biases, n);
- l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
+ if(gpu_index >= 0){
+ if (adam) {
+ l.m_gpu = cuda_make_array(l.m, c*n*size*size);
+ l.v_gpu = cuda_make_array(l.v, c*n*size*size);
+ }
- l.scales_gpu = cuda_make_array(l.scales, n);
- l.scale_updates_gpu = cuda_make_array(l.scale_updates, n);
+ l.weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+#ifdef CUDNN_HALF
+ l.weights_gpu16 = cuda_make_array(NULL, c*n*size*size / 2); //cuda_make_array(l.weights, c*n*size*size / 2);
+ l.weight_updates_gpu16 = cuda_make_array(NULL, c*n*size*size / 2); //cuda_make_array(l.weight_updates, c*n*size*size / 2);
+#endif
+ l.weight_updates_gpu = cuda_make_array(l.weight_updates, c*n*size*size);
- l.col_image_gpu = cuda_make_array(l.col_image, out_h*out_w*size*size*c);
- l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n);
- l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ l.biases_gpu = cuda_make_array(l.biases, n);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
- if(binary){
- l.binary_filters_gpu = cuda_make_array(l.filters, c*n*size*size);
- }
+ l.delta_gpu = cuda_make_array(l.delta, l.batch*out_h*out_w*n);
+ l.output_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
- if(batch_normalize){
- l.mean_gpu = cuda_make_array(l.mean, n);
- l.variance_gpu = cuda_make_array(l.variance, n);
+ if(binary){
+ l.binary_weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+ }
+ if(xnor){
+ l.binary_weights_gpu = cuda_make_array(l.weights, c*n*size*size);
+ l.binary_input_gpu = cuda_make_array(0, l.inputs*l.batch);
+ }
- l.rolling_mean_gpu = cuda_make_array(l.mean, n);
- l.rolling_variance_gpu = cuda_make_array(l.variance, n);
+ if(batch_normalize){
+ l.mean_gpu = cuda_make_array(l.mean, n);
+ l.variance_gpu = cuda_make_array(l.variance, n);
- l.mean_delta_gpu = cuda_make_array(l.mean, n);
- l.variance_delta_gpu = cuda_make_array(l.variance, n);
+ l.rolling_mean_gpu = cuda_make_array(l.mean, n);
+ l.rolling_variance_gpu = cuda_make_array(l.variance, n);
- l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
- l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ l.mean_delta_gpu = cuda_make_array(l.mean, n);
+ l.variance_delta_gpu = cuda_make_array(l.variance, n);
+
+ l.scales_gpu = cuda_make_array(l.scales, n);
+ l.scale_updates_gpu = cuda_make_array(l.scale_updates, n);
+
+ l.x_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ l.x_norm_gpu = cuda_make_array(l.output, l.batch*out_h*out_w*n);
+ }
+#ifdef CUDNN
+ cudnnCreateTensorDescriptor(&l.normDstTensorDesc);
+ cudnnCreateTensorDescriptor(&l.normDstTensorDescF16);
+ cudnnCreateTensorDescriptor(&l.normTensorDesc);
+ cudnnCreateTensorDescriptor(&l.srcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.dstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.weightDesc);
+ cudnnCreateTensorDescriptor(&l.dsrcTensorDesc);
+ cudnnCreateTensorDescriptor(&l.ddstTensorDesc);
+ cudnnCreateFilterDescriptor(&l.dweightDesc);
+ cudnnCreateConvolutionDescriptor(&l.convDesc);
+ cudnn_convolutional_setup(&l, cudnn_fastest);
+#endif
}
#endif
+ l.workspace_size = get_workspace_size(l);
l.activation = activation;
- fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n);
+ fprintf(stderr, "conv %5d %2d x%2d /%2d %4d x%4d x%4d -> %4d x%4d x%4d\n", n, size, size, stride, w, h, c, l.out_w, l.out_h, l.out_c);
return l;
}
@@ -200,15 +408,18 @@
for(i = 0; i < l.n; ++i){
float scale = l.scales[i]/sqrt(l.rolling_variance[i] + .00001);
for(j = 0; j < l.c*l.size*l.size; ++j){
- l.filters[i*l.c*l.size*l.size + j] *= scale;
+ l.weights[i*l.c*l.size*l.size + j] *= scale;
}
l.biases[i] -= l.rolling_mean[i] * scale;
+ l.scales[i] = 1;
+ l.rolling_mean[i] = 0;
+ l.rolling_variance[i] = 1;
}
}
void test_convolutional_layer()
{
- convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 0);
+ convolutional_layer l = make_convolutional_layer(1, 5, 5, 3, 2, 5, 2, 1, LEAKY, 1, 0, 0, 0);
l.batch_normalize = 1;
float data[] = {1,1,1,1,1,
1,1,1,1,1,
@@ -232,6 +443,8 @@
void resize_convolutional_layer(convolutional_layer *l, int w, int h)
{
+ int old_w = l->w;
+ int old_h = l->h;
l->w = w;
l->h = h;
int out_w = convolutional_out_width(*l);
@@ -243,21 +456,45 @@
l->outputs = l->out_h * l->out_w * l->out_c;
l->inputs = l->w * l->h * l->c;
- l->col_image = realloc(l->col_image,
- out_h*out_w*l->size*l->size*l->c*sizeof(float));
- l->output = realloc(l->output,
- l->batch*out_h * out_w * l->n*sizeof(float));
- l->delta = realloc(l->delta,
- l->batch*out_h * out_w * l->n*sizeof(float));
+ l->output = realloc(l->output, l->batch*l->outputs*sizeof(float));
+ l->delta = realloc(l->delta, l->batch*l->outputs*sizeof(float));
+ if(l->batch_normalize){
+ l->x = realloc(l->x, l->batch*l->outputs*sizeof(float));
+ l->x_norm = realloc(l->x_norm, l->batch*l->outputs*sizeof(float));
+ }
#ifdef GPU
- cuda_free(l->col_image_gpu);
- cuda_free(l->delta_gpu);
- cuda_free(l->output_gpu);
+ if (old_w < w || old_h < h) {
+ cuda_free(l->delta_gpu);
+ cuda_free(l->output_gpu);
- l->col_image_gpu = cuda_make_array(l->col_image, out_h*out_w*l->size*l->size*l->c);
- l->delta_gpu = cuda_make_array(l->delta, l->batch*out_h*out_w*l->n);
- l->output_gpu = cuda_make_array(l->output, l->batch*out_h*out_w*l->n);
+ l->delta_gpu = cuda_make_array(l->delta, l->batch*l->outputs);
+ l->output_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+
+ if (l->batch_normalize) {
+ cuda_free(l->x_gpu);
+ cuda_free(l->x_norm_gpu);
+
+ l->x_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ l->x_norm_gpu = cuda_make_array(l->output, l->batch*l->outputs);
+ }
+ }
+#ifdef CUDNN
+ cudnn_convolutional_setup(l, cudnn_fastest);
+#endif
+#endif
+ l->workspace_size = get_workspace_size(*l);
+
+#ifdef CUDNN
+ // check for excessive memory consumption
+ size_t free_byte;
+ size_t total_byte;
+ check_error(cudaMemGetInfo(&free_byte, &total_byte));
+ if (l->workspace_size > free_byte || l->workspace_size >= total_byte / 2) {
+ printf(" used slow CUDNN algo without Workspace! \n");
+ cudnn_convolutional_setup(l, cudnn_smallest);
+ l->workspace_size = get_workspace_size(*l);
+ }
#endif
}
@@ -295,7 +532,7 @@
}
}
-void forward_convolutional_layer(const convolutional_layer l, network_state state)
+void forward_convolutional_layer(convolutional_layer l, network_state state)
{
int out_h = convolutional_out_height(l);
int out_w = convolutional_out_width(l);
@@ -303,12 +540,20 @@
fill_cpu(l.outputs*l.batch, 0, l.output, 1);
+ if(l.xnor){
+ binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
+ swap_binary(&l);
+ binarize_cpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input);
+ state.input = l.binary_input;
+ }
+
int m = l.n;
int k = l.size*l.size*l.c;
int n = out_h*out_w;
- float *a = l.filters;
- float *b = l.col_image;
+
+ float *a = l.weights;
+ float *b = state.workspace;
float *c = l.output;
for(i = 0; i < l.batch; ++i){
@@ -320,18 +565,12 @@
}
if(l.batch_normalize){
- if(state.train){
- mean_cpu(l.output, l.batch, l.n, l.out_h*l.out_w, l.mean);
- variance_cpu(l.output, l.mean, l.batch, l.n, l.out_h*l.out_w, l.variance);
- normalize_cpu(l.output, l.mean, l.variance, l.batch, l.n, l.out_h*l.out_w);
- } else {
- normalize_cpu(l.output, l.rolling_mean, l.rolling_variance, l.batch, l.n, l.out_h*l.out_w);
- }
- scale_bias(l.output, l.scales, l.batch, l.n, out_h*out_w);
+ forward_batchnorm_layer(l, state);
}
add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);
activate_array(l.output, m*n*l.batch, l.activation);
+ if(l.binary || l.xnor) swap_binary(&l);
}
void backward_convolutional_layer(convolutional_layer l, network_state state)
@@ -345,10 +584,14 @@
gradient_array(l.output, m*k*l.batch, l.activation, l.delta);
backward_bias(l.bias_updates, l.delta, l.batch, l.n, k);
+ if(l.batch_normalize){
+ backward_batchnorm_layer(l, state);
+ }
+
for(i = 0; i < l.batch; ++i){
float *a = l.delta + i*m*k;
- float *b = l.col_image;
- float *c = l.filter_updates;
+ float *b = state.workspace;
+ float *c = l.weight_updates;
float *im = state.input+i*l.c*l.h*l.w;
@@ -357,13 +600,13 @@
gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
if(state.delta){
- a = l.filters;
+ a = l.weights;
b = l.delta + i*m*k;
- c = l.col_image;
+ c = state.workspace;
gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
- col2im_cpu(l.col_image, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w);
+ col2im_cpu(state.workspace, l.c, l.h, l.w, l.size, l.stride, l.pad, state.delta+i*l.c*l.h*l.w);
}
}
}
@@ -374,36 +617,41 @@
axpy_cpu(l.n, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
scal_cpu(l.n, momentum, l.bias_updates, 1);
- axpy_cpu(size, -decay*batch, l.filters, 1, l.filter_updates, 1);
- axpy_cpu(size, learning_rate/batch, l.filter_updates, 1, l.filters, 1);
- scal_cpu(size, momentum, l.filter_updates, 1);
+ if(l.scales){
+ axpy_cpu(l.n, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
+ scal_cpu(l.n, momentum, l.scale_updates, 1);
+ }
+
+ axpy_cpu(size, -decay*batch, l.weights, 1, l.weight_updates, 1);
+ axpy_cpu(size, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
+ scal_cpu(size, momentum, l.weight_updates, 1);
}
-image get_convolutional_filter(convolutional_layer l, int i)
+image get_convolutional_weight(convolutional_layer l, int i)
{
int h = l.size;
int w = l.size;
int c = l.c;
- return float_to_image(w,h,c,l.filters+i*h*w*c);
+ return float_to_image(w,h,c,l.weights+i*h*w*c);
}
-void rgbgr_filters(convolutional_layer l)
+void rgbgr_weights(convolutional_layer l)
{
int i;
for(i = 0; i < l.n; ++i){
- image im = get_convolutional_filter(l, i);
+ image im = get_convolutional_weight(l, i);
if (im.c == 3) {
rgbgr_image(im);
}
}
}
-void rescale_filters(convolutional_layer l, float scale, float trans)
+void rescale_weights(convolutional_layer l, float scale, float trans)
{
int i;
for(i = 0; i < l.n; ++i){
- image im = get_convolutional_filter(l, i);
+ image im = get_convolutional_weight(l, i);
if (im.c == 3) {
scale_image(im, scale);
float sum = sum_array(im.data, im.w*im.h*im.c);
@@ -412,21 +660,21 @@
}
}
-image *get_filters(convolutional_layer l)
+image *get_weights(convolutional_layer l)
{
- image *filters = calloc(l.n, sizeof(image));
+ image *weights = calloc(l.n, sizeof(image));
int i;
for(i = 0; i < l.n; ++i){
- filters[i] = copy_image(get_convolutional_filter(l, i));
- //normalize_image(filters[i]);
+ weights[i] = copy_image(get_convolutional_weight(l, i));
+ //normalize_image(weights[i]);
}
- return filters;
+ return weights;
}
-image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_filters)
+image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_weights)
{
- image *single_filters = get_filters(l);
- show_images(single_filters, l.n, window);
+ image *single_weights = get_weights(l);
+ show_images(single_weights, l.n, window);
image delta = get_convolutional_image(l);
image dc = collapse_image_layers(delta, 1);
@@ -435,6 +683,6 @@
//show_image(dc, buff);
//save_image(dc, buff);
free_image(dc);
- return single_filters;
+ return single_weights;
}
--
Gitblit v1.10.0