From 0948df52b850b908e7a74cb589d19fa29eb30368 Mon Sep 17 00:00:00 2001
From: Alexey <AlexeyAB@users.noreply.github.com>
Date: Tue, 08 May 2018 14:27:45 +0000
Subject: [PATCH] Merge pull request #741 from IlyaOvodov/Fix_detector_output
---
src/convolutional_layer.c | 748 +++++++++++++++++++++++++++++++++++++++++++++++++--------
1 files changed, 643 insertions(+), 105 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index d4aff73..b8065fd 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,150 +1,688 @@
#include "convolutional_layer.h"
+#include "utils.h"
+#include "batchnorm_layer.h"
+#include "im2col.h"
+#include "col2im.h"
+#include "blas.h"
+#include "gemm.h"
#include <stdio.h>
+#include <time.h>
-image get_convolutional_image(convolutional_layer layer)
+#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)
{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.n;
- return double_to_image(h,w,c,layer.output);
+ 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
}
-image get_convolutional_delta(convolutional_layer layer)
+void binarize_weights(float *weights, int n, int size, float *binary)
{
- int h = (layer.h-1)/layer.stride + 1;
- int w = (layer.w-1)/layer.stride + 1;
- int c = layer.n;
- return double_to_image(h,w,c,layer.delta);
+ 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;
+ }
+ }
}
-convolutional_layer *make_convolutional_layer(int h, int w, int c, int n, int size, int stride, ACTIVATION activator)
+void binarize_cpu(float *input, int n, float *binary)
{
- printf("Convolutional Layer: %d x %d x %d image, %d filters\n", h,w,c,n);
int i;
- convolutional_layer *layer = calloc(1, sizeof(convolutional_layer));
- layer->h = h;
- layer->w = w;
- layer->c = c;
- layer->n = n;
- layer->stride = stride;
- layer->kernels = calloc(n, sizeof(image));
- layer->kernel_updates = calloc(n, sizeof(image));
- layer->biases = calloc(n, sizeof(double));
- layer->bias_updates = calloc(n, sizeof(double));
for(i = 0; i < n; ++i){
- layer->biases[i] = .005;
- layer->kernels[i] = make_random_kernel(size, c);
- layer->kernel_updates[i] = make_random_kernel(size, c);
+ binary[i] = (input[i] > 0) ? 1 : -1;
}
- layer->output = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * n, sizeof(double));
- layer->delta = calloc(((h-1)/stride+1) * ((w-1)/stride+1) * n, sizeof(double));
- layer->upsampled = make_image(h,w,n);
-
- 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;
- }
- return layer;
}
-void forward_convolutional_layer(const convolutional_layer layer, double *in)
+void binarize_input(float *input, int n, int size, float *binary)
{
- image input = double_to_image(layer.h, layer.w, layer.c, in);
- image output = get_convolutional_image(layer);
- int i,j;
- for(i = 0; i < layer.n; ++i){
- convolve(input, layer.kernels[i], layer.stride, i, output);
- }
- for(i = 0; i < output.c; ++i){
- for(j = 0; j < output.h*output.w; ++j){
- int index = i*output.h*output.w + j;
- output.data[index] += layer.biases[i];
- output.data[index] = layer.activation(output.data[index]);
+ 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;
}
}
}
-void backward_convolutional_layer(convolutional_layer layer, double *input, double *delta)
+int convolutional_out_height(convolutional_layer l)
+{
+ return (l.h + 2*l.pad - l.size) / l.stride + 1;
+}
+
+int convolutional_out_width(convolutional_layer l)
+{
+ return (l.w + 2*l.pad - l.size) / l.stride + 1;
+}
+
+image get_convolutional_image(convolutional_layer l)
+{
+ int h,w,c;
+ h = convolutional_out_height(l);
+ w = convolutional_out_width(l);
+ c = l.n;
+ return float_to_image(w,h,c,l.output);
+}
+
+image get_convolutional_delta(convolutional_layer l)
+{
+ int h,w,c;
+ h = convolutional_out_height(l);
+ w = convolutional_out_width(l);
+ c = l.n;
+ return float_to_image(w,h,c,l.delta);
+}
+
+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);
+}
+
+#ifdef GPU
+#ifdef CUDNN
+void cudnn_convolutional_setup(layer *l, int cudnn_preference)
+{
+
+#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
+
+#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
+
+ // 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};
+ l.type = CONVOLUTIONAL;
- image in_image = double_to_image(layer.h, layer.w, layer.c, input);
- image in_delta = double_to_image(layer.h, layer.w, layer.c, delta);
- image out_delta = get_convolutional_delta(layer);
- zero_image(in_delta);
+ l.h = h;
+ l.w = w;
+ l.c = c;
+ l.n = n;
+ l.binary = binary;
+ l.xnor = xnor;
+ l.batch = batch;
+ l.stride = stride;
+ l.size = size;
+ l.pad = padding;
+ l.batch_normalize = batch_normalize;
- for(i = 0; i < layer.n; ++i){
- back_convolve(in_delta, layer.kernels[i], layer.stride, i, out_delta);
+ 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.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;
+ l.out_w = out_w;
+ l.out_c = n;
+ l.outputs = l.out_h * l.out_w * l.out_c;
+ l.inputs = l.w * l.h * l.c;
+
+ 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_weights = calloc(c*n*size*size, sizeof(float));
+ l.cweights = calloc(c*n*size*size, sizeof(char));
+ l.scales = calloc(n, sizeof(float));
}
- for(i = 0; i < layer.h*layer.w*layer.c; ++i){
- in_delta.data[i] *= layer.gradient(in_image.data[i]);
+ 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){
+ l.scales = calloc(n, sizeof(float));
+ l.scale_updates = calloc(n, sizeof(float));
+ for(i = 0; i < n; ++i){
+ l.scales[i] = 1;
+ }
+
+ 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.forward_gpu = forward_convolutional_layer_gpu;
+ l.backward_gpu = backward_convolutional_layer_gpu;
+ l.update_gpu = update_convolutional_layer_gpu;
+
+ 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.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.biases_gpu = cuda_make_array(l.biases, n);
+ l.bias_updates_gpu = cuda_make_array(l.bias_updates, n);
+
+ 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(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);
+ }
+
+ if(batch_normalize){
+ l.mean_gpu = cuda_make_array(l.mean, n);
+ l.variance_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.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, "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;
+}
+
+void denormalize_convolutional_layer(convolutional_layer l)
+{
+ int i, j;
+ 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.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 backpropagate_convolutional_layer_convolve(image input, convolutional_layer layer)
+void test_convolutional_layer()
{
- int i,j;
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
+ 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,
+ 1,1,1,1,1,
+ 1,1,1,1,1,
+ 1,1,1,1,1,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 2,2,2,2,2,
+ 3,3,3,3,3,
+ 3,3,3,3,3,
+ 3,3,3,3,3,
+ 3,3,3,3,3,
+ 3,3,3,3,3};
+ network_state state = {0};
+ state.input = data;
+ forward_convolutional_layer(l, state);
+}
+
+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);
+ int out_h = convolutional_out_height(*l);
+
+ l->out_w = out_w;
+ l->out_h = out_h;
+
+ l->outputs = l->out_h * l->out_w * l->out_c;
+ l->inputs = l->w * l->h * l->c;
+
+ 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));
}
- zero_image(input);
- upsample_image(layer.output, layer.stride, layer.upsampled);
- for(j = 0; j < input.c; ++j){
- for(i = 0; i < layer.n; ++i){
- two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, input, j);
+#ifdef GPU
+ if (old_w < w || old_h < h) {
+ cuda_free(l->delta_gpu);
+ cuda_free(l->output_gpu);
+
+ 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! Need memory: %d, available: %d\n", l->workspace_size, (free_byte < total_byte/2) ? free_byte : total_byte/2);
+ cudnn_convolutional_setup(l, cudnn_smallest);
+ l->workspace_size = get_workspace_size(*l);
+ }
+#endif
+}
+
+void add_bias(float *output, float *biases, int batch, int n, int size)
+{
+ int i,j,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ for(j = 0; j < size; ++j){
+ output[(b*n + i)*size + j] += biases[i];
+ }
}
}
-
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
-}
-*/
-
-void learn_convolutional_layer(convolutional_layer layer, double *input)
-{
- int i;
- image in_image = double_to_image(layer.h, layer.w, layer.c, input);
- image out_delta = get_convolutional_delta(layer);
- for(i = 0; i < layer.n; ++i){
- kernel_update(in_image, layer.kernel_updates[i], layer.stride, i, out_delta);
- layer.bias_updates[i] += avg_image_layer(out_delta, i);
- }
}
-void update_convolutional_layer(convolutional_layer layer, double step)
+void scale_bias(float *output, float *scales, int batch, int n, int size)
{
- return;
- int i,j;
- for(i = 0; i < layer.n; ++i){
- layer.biases[i] += step*layer.bias_updates[i];
- layer.bias_updates[i] = 0;
- int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c;
- for(j = 0; j < pixels; ++j){
- layer.kernels[i].data[j] += step*layer.kernel_updates[i].data[j];
+ int i,j,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ for(j = 0; j < size; ++j){
+ output[(b*n + i)*size + j] *= scales[i];
+ }
}
- zero_image(layer.kernel_updates[i]);
}
}
-void visualize_convolutional_layer(convolutional_layer layer)
+void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
+{
+ int i,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ bias_updates[i] += sum_array(delta+size*(i+b*n), size);
+ }
+ }
+}
+
+void forward_convolutional_layer(convolutional_layer l, network_state state)
+{
+ int out_h = convolutional_out_height(l);
+ int out_w = convolutional_out_width(l);
+ int i;
+
+ 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.weights;
+ float *b = state.workspace;
+ float *c = l.output;
+
+ for(i = 0; i < l.batch; ++i){
+ im2col_cpu(state.input, l.c, l.h, l.w,
+ l.size, l.stride, l.pad, b);
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ c += n*m;
+ state.input += l.c*l.h*l.w;
+ }
+
+ if(l.batch_normalize){
+ 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)
{
int i;
+ int m = l.n;
+ int n = l.size*l.size*l.c;
+ int k = convolutional_out_height(l)*
+ convolutional_out_width(l);
+
+ 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 = state.workspace;
+ float *c = l.weight_updates;
+
+ float *im = state.input+i*l.c*l.h*l.w;
+
+ im2col_cpu(im, l.c, l.h, l.w,
+ l.size, l.stride, l.pad, b);
+ gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+
+ if(state.delta){
+ a = l.weights;
+ b = l.delta + i*m*k;
+ c = state.workspace;
+
+ gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
+
+ 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);
+ }
+ }
+}
+
+void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay)
+{
+ int size = l.size*l.size*l.c*l.n;
+ axpy_cpu(l.n, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
+ scal_cpu(l.n, momentum, l.bias_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_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.weights+i*h*w*c);
+}
+
+void rgbgr_weights(convolutional_layer l)
+{
+ int i;
+ for(i = 0; i < l.n; ++i){
+ image im = get_convolutional_weight(l, i);
+ if (im.c == 3) {
+ rgbgr_image(im);
+ }
+ }
+}
+
+void rescale_weights(convolutional_layer l, float scale, float trans)
+{
+ int i;
+ for(i = 0; i < l.n; ++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);
+ l.biases[i] += sum*trans;
+ }
+ }
+}
+
+image *get_weights(convolutional_layer l)
+{
+ image *weights = calloc(l.n, sizeof(image));
+ int i;
+ for(i = 0; i < l.n; ++i){
+ weights[i] = copy_image(get_convolutional_weight(l, i));
+ //normalize_image(weights[i]);
+ }
+ return weights;
+}
+
+image *visualize_convolutional_layer(convolutional_layer l, char *window, image *prev_weights)
+{
+ 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);
char buff[256];
- //image vis = make_image(layer.n*layer.size, layer.size*layer.kernels[0].c, 3);
- for(i = 0; i < layer.n; ++i){
- image k = layer.kernels[i];
- sprintf(buff, "Kernel %d", i);
- if(k.c <= 3) show_image(k, buff);
- else show_image_layers(k, buff);
- }
+ sprintf(buff, "%s: Output", window);
+ //show_image(dc, buff);
+ //save_image(dc, buff);
+ free_image(dc);
+ return single_weights;
}
--
Gitblit v1.10.0