From 2fafd8dfcef34889bf2d315abbdcb8b00df5b249 Mon Sep 17 00:00:00 2001
From: Hashim Sharif <hsharif3@tyler.cs.illinois.edu>
Date: Sun, 17 Mar 2019 18:22:28 -0500
Subject: [PATCH] Adding Top-5% acc compuation to utils and VGG-16_CIFAR100
---
.../dnn_sources/include/utils.h | 72 +++++++-
.../dnn_sources/src/vgg16_cifar100_5.cc | 167 ++++++++++++++++++
2 files changed, 234 insertions(+), 5 deletions(-)
create mode 100644 llvm/projects/hpvm-tensor-rt/dnn_sources/src/vgg16_cifar100_5.cc
diff --git a/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h b/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h
index 799b6f45bc..f8c75f7102 100644
--- a/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h
+++ b/llvm/projects/hpvm-tensor-rt/dnn_sources/include/utils.h
@@ -6,6 +6,7 @@
#include <sstream>
#include <vector>
+#include <bits/stdc++.h>
#include "../../tensor_runtime/include/tensor.h"
#include "types.h"
@@ -331,10 +332,6 @@ uint8_t* readLabelsBatch(const char* labels_file, int start, int end){
size_t bytes_read = fread(labels, 1, sizeof(uint8_t) * num_labels, file);
- /*for(unsigned int i = 0 ; i < 20; i++){
- printf("labels[%d] = %u \n", i, labels[i]);
- }
- */
fclose(file);
@@ -398,7 +395,6 @@ float computeAccuracy2(uint8_t* labels, int num_labels, void* result_ptr, unsign
printf("batch_dim = %lu, channels = %lu \n", batch_dim, channels);
- //for(int i = 0; i < batch_dim; i++){
for(int i = 0; i < num_labels; i++){
int chosen = 0;
@@ -435,6 +431,72 @@ float computeAccuracy2(uint8_t* labels, int num_labels, void* result_ptr, unsign
}
+struct ClassProb{
+ float prob;
+ int index;
+};
+
+
+bool descendFloatComp(ClassProb obj1, ClassProb obj2){
+ return obj1.prob > obj2.prob;
+}
+
+
+float computeTop5Accuracy(uint8_t* labels, int num_labels, void* result_ptr, unsigned num_classes = 10){
+
+ struct Tensor* result = (struct Tensor*) result_ptr;
+
+ size_t batch_dim = result->dims.dim_sizes[0];
+ size_t channels = result->dims.dim_sizes[1];
+ float* data = (float*) result->host_data;
+ int num_errors = 0;
+
+ printf("batch_dim = %lu, channels = %lu \n", batch_dim, channels);
+
+ for(int i = 0; i < num_labels; i++){
+
+ std::vector<ClassProb> elem_probs;
+ for (int id = 0; id < num_classes; ++id){
+ ClassProb cProb;
+ cProb.prob = data[i * channels + id];
+ cProb.index = id;
+ elem_probs.push_back(cProb);
+ }
+
+ std:sort(elem_probs.begin(), elem_probs.end(), descendFloatComp);
+ // Check if any of top-5 predictions matches
+ bool matched = false;
+ for(int j = 0; j < 5; j++){
+ ClassProb cProb = elem_probs[j];
+ if(cProb.index == labels[i])
+ matched = true;
+ }
+
+ if(!matched)
+ num_errors +=1;
+ }
+
+ float accuracy = ((batch_dim - num_errors) * 1.0 / batch_dim * 1.0) * 100.0;
+ printf("****** Accuracy = %f \n\n", accuracy);
+
+ FILE* fp = fopen("final_accuracy", "w+");
+ if(fp != NULL){
+
+ std::ostringstream ss;
+ ss << std::fixed << accuracy;
+ std::string print_str = ss.str();
+
+ fwrite(print_str.c_str(), 1, print_str.length(), fp);
+ }
+
+ fclose(fp);
+
+ return accuracy;
+}
+
+
+
+
void dumpFinalAccuracy(float accuracy){
printf("\n\n **** Final Accuracy = %f \n", accuracy);
diff --git a/llvm/projects/hpvm-tensor-rt/dnn_sources/src/vgg16_cifar100_5.cc b/llvm/projects/hpvm-tensor-rt/dnn_sources/src/vgg16_cifar100_5.cc
new file mode 100644
index 0000000000..3ee273d70a
--- /dev/null
+++ b/llvm/projects/hpvm-tensor-rt/dnn_sources/src/vgg16_cifar100_5.cc
@@ -0,0 +1,167 @@
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <string.h>
+#include "../../tensor_runtime/include/tensor_runtime.h"
+#include "../include/utils.h"
+
+int main(){
+
+ llvm_hpvm_initTensorRt(0);
+
+ std::string dir_prefix = std::string("../model_params/vgg16_cifar100_front/");
+ //std::string input_path = dir_prefix + std::string("vgg16_cifar100_calib.bin");
+ //std::string labels_path = dir_prefix + std::string("vgg16_cifar100_train_labels.bin");
+
+ std::string input_path = dir_prefix + std::string("input.bin");
+ std::string labels_path = dir_prefix + std::string("labels.bin");
+
+ std::string conv2d_1_w_path = dir_prefix + std::string("conv2d_1_w.bin");
+ void* conv2d_1_w = readTrainedWeights(conv2d_1_w_path.c_str(), 0,64,3,3,3);
+ std::string conv2d_1_b_path = dir_prefix + std::string("conv2d_1_b.bin");
+ void* conv2d_1_b = readTrainedWeights(conv2d_1_b_path.c_str(), 0,1,64,1,1);
+ std::string conv2d_2_w_path = dir_prefix + std::string("conv2d_2_w.bin");
+ void* conv2d_2_w = readTrainedWeights(conv2d_2_w_path.c_str(), 0,64,64,3,3);
+ std::string conv2d_2_b_path = dir_prefix + std::string("conv2d_2_b.bin");
+ void* conv2d_2_b = readTrainedWeights(conv2d_2_b_path.c_str(), 0,1,64,1,1);
+ std::string conv2d_3_w_path = dir_prefix + std::string("conv2d_3_w.bin");
+ void* conv2d_3_w = readTrainedWeights(conv2d_3_w_path.c_str(), 0,128,64,3,3);
+ std::string conv2d_3_b_path = dir_prefix + std::string("conv2d_3_b.bin");
+ void* conv2d_3_b = readTrainedWeights(conv2d_3_b_path.c_str(), 0,1,128,1,1);
+ std::string conv2d_4_w_path = dir_prefix + std::string("conv2d_4_w.bin");
+ void* conv2d_4_w = readTrainedWeights(conv2d_4_w_path.c_str(), 0,128,128,3,3);
+ std::string conv2d_4_b_path = dir_prefix + std::string("conv2d_4_b.bin");
+ void* conv2d_4_b = readTrainedWeights(conv2d_4_b_path.c_str(), 0,1,128,1,1);
+ std::string conv2d_5_w_path = dir_prefix + std::string("conv2d_5_w.bin");
+ void* conv2d_5_w = readTrainedWeights(conv2d_5_w_path.c_str(), 0,256,128,3,3);
+ std::string conv2d_5_b_path = dir_prefix + std::string("conv2d_5_b.bin");
+ void* conv2d_5_b = readTrainedWeights(conv2d_5_b_path.c_str(), 0,1,256,1,1);
+ std::string conv2d_6_w_path = dir_prefix + std::string("conv2d_6_w.bin");
+ void* conv2d_6_w = readTrainedWeights(conv2d_6_w_path.c_str(), 0,256,256,3,3);
+ std::string conv2d_6_b_path = dir_prefix + std::string("conv2d_6_b.bin");
+ void* conv2d_6_b = readTrainedWeights(conv2d_6_b_path.c_str(), 0,1,256,1,1);
+ std::string conv2d_7_w_path = dir_prefix + std::string("conv2d_7_w.bin");
+ void* conv2d_7_w = readTrainedWeights(conv2d_7_w_path.c_str(), 0,256,256,3,3);
+ std::string conv2d_7_b_path = dir_prefix + std::string("conv2d_7_b.bin");
+ void* conv2d_7_b = readTrainedWeights(conv2d_7_b_path.c_str(), 0,1,256,1,1);
+ std::string conv2d_8_w_path = dir_prefix + std::string("conv2d_8_w.bin");
+ void* conv2d_8_w = readTrainedWeights(conv2d_8_w_path.c_str(), 0,512,256,3,3);
+ std::string conv2d_8_b_path = dir_prefix + std::string("conv2d_8_b.bin");
+ void* conv2d_8_b = readTrainedWeights(conv2d_8_b_path.c_str(), 0,1,512,1,1);
+ std::string conv2d_9_w_path = dir_prefix + std::string("conv2d_9_w.bin");
+ void* conv2d_9_w = readTrainedWeights(conv2d_9_w_path.c_str(), 0,512,512,3,3);
+ std::string conv2d_9_b_path = dir_prefix + std::string("conv2d_9_b.bin");
+ void* conv2d_9_b = readTrainedWeights(conv2d_9_b_path.c_str(), 0,1,512,1,1);
+ std::string conv2d_10_w_path = dir_prefix + std::string("conv2d_10_w.bin");
+ void* conv2d_10_w = readTrainedWeights(conv2d_10_w_path.c_str(), 0,512,512,3,3);
+ std::string conv2d_10_b_path = dir_prefix + std::string("conv2d_10_b.bin");
+ void* conv2d_10_b = readTrainedWeights(conv2d_10_b_path.c_str(), 0,1,512,1,1);
+ std::string conv2d_11_w_path = dir_prefix + std::string("conv2d_11_w.bin");
+ void* conv2d_11_w = readTrainedWeights(conv2d_11_w_path.c_str(), 0,512,512,3,3);
+ std::string conv2d_11_b_path = dir_prefix + std::string("conv2d_11_b.bin");
+ void* conv2d_11_b = readTrainedWeights(conv2d_11_b_path.c_str(), 0,1,512,1,1);
+ std::string conv2d_12_w_path = dir_prefix + std::string("conv2d_12_w.bin");
+ void* conv2d_12_w = readTrainedWeights(conv2d_12_w_path.c_str(), 0,512,512,3,3);
+ std::string conv2d_12_b_path = dir_prefix + std::string("conv2d_12_b.bin");
+ void* conv2d_12_b = readTrainedWeights(conv2d_12_b_path.c_str(), 0,1,512,1,1);
+ std::string conv2d_13_w_path = dir_prefix + std::string("conv2d_13_w.bin");
+ void* conv2d_13_w = readTrainedWeights(conv2d_13_w_path.c_str(), 0,512,512,3,3);
+ std::string conv2d_13_b_path = dir_prefix + std::string("conv2d_13_b.bin");
+ void* conv2d_13_b = readTrainedWeights(conv2d_13_b_path.c_str(), 0,1,512,1,1);
+ std::string dense_1_w_path = dir_prefix + std::string("dense_1_w.bin");
+ void* dense_1_w = readTrainedWeights(dense_1_w_path.c_str(), 0,1,1,512,512);
+ std::string dense_1_b_path = dir_prefix + std::string("dense_1_b.bin");
+ void* dense_1_b = readTrainedWeights(dense_1_b_path.c_str(), 0,1,512,1,1);
+ std::string dense_2_w_path = dir_prefix + std::string("dense_2_w.bin");
+ void* dense_2_w = readTrainedWeights(dense_2_w_path.c_str(), 0,1,1,512,100);
+ std::string dense_2_b_path = dir_prefix + std::string("dense_2_b.bin");
+ void* dense_2_b = readTrainedWeights(dense_2_b_path.c_str(), 0,1,100,1,1);
+
+
+ startMemTracking();
+
+ int test_input_size = 5000;
+ int batch_size = 2500;
+ int offset = 5000;
+
+ int batch_count = test_input_size / batch_size;
+ float final_accuracy = 0.0;
+
+ for(int i = 0; i < batch_count; i++){
+
+ int start = i * batch_size + offset;
+ int end = (i + 1) * batch_size + offset;
+
+ void* input = readInputBatch(input_path.c_str(),0,start,end,3,32,32);
+
+ void* var_0 = tensorConvolution(input, conv2d_1_w, 1, 1, 1, 1, 1, 0);
+ void* var_1 = tensorAdd(var_0, conv2d_1_b);
+ void* var_2 = tensorRelu(var_1);
+ void* var_4 = tensorConvolution(var_2, conv2d_2_w, 1, 1, 1, 1, 1, 0);
+ void* var_5 = tensorAdd(var_4, conv2d_2_b);
+ void* var_6 = tensorRelu(var_5);
+ void* var_7 = tensorPooling(var_6,0,2,2,0,0,2,2);
+ void* var_8 = tensorConvolution(var_7, conv2d_3_w, 1, 1, 1, 1, 1, 0);
+ void* var_9 = tensorAdd(var_8, conv2d_3_b);
+ void* var_10 = tensorRelu(var_9);
+ void* var_12 = tensorConvolution(var_10, conv2d_4_w, 1, 1, 1, 1, 1, 0);
+ void* var_13 = tensorAdd(var_12, conv2d_4_b);
+ void* var_14 = tensorRelu(var_13);
+ void* var_15 = tensorPooling(var_14,0,2,2,0,0,2,2);
+ void* var_16 = tensorConvolution(var_15, conv2d_5_w, 1, 1, 1, 1, 1, 0);
+ void* var_17 = tensorAdd(var_16, conv2d_5_b);
+ void* var_18 = tensorRelu(var_17);
+ void* var_20 = tensorConvolution(var_18, conv2d_6_w, 1, 1, 1, 1, 1, 0);
+ void* var_21 = tensorAdd(var_20, conv2d_6_b);
+ void* var_22 = tensorRelu(var_21);
+ void* var_24 = tensorConvolution(var_22, conv2d_7_w, 1, 1, 1, 1, 1, 0);
+ void* var_25 = tensorAdd(var_24, conv2d_7_b);
+ void* var_26 = tensorRelu(var_25);
+ void* var_27 = tensorPooling(var_26,0,2,2,0,0,2,2);
+ void* var_28 = tensorConvolution(var_27, conv2d_8_w, 1, 1, 1, 1, 1, 0);
+ void* var_29 = tensorAdd(var_28, conv2d_8_b);
+ void* var_30 = tensorRelu(var_29);
+ void* var_32 = tensorConvolution(var_30, conv2d_9_w, 1, 1, 1, 1, 1, 0);
+ void* var_33 = tensorAdd(var_32, conv2d_9_b);
+ void* var_34 = tensorRelu(var_33);
+ void* var_36 = tensorConvolution(var_34, conv2d_10_w, 1, 1, 1, 1, 1, 0);
+ void* var_37 = tensorAdd(var_36, conv2d_10_b);
+ void* var_38 = tensorRelu(var_37);
+ void* var_39 = tensorPooling(var_38,0,2,2,0,0,2,2);
+ void* var_40 = tensorConvolution(var_39, conv2d_11_w, 1, 1, 1, 1, 1, 0);
+ void* var_41 = tensorAdd(var_40, conv2d_11_b);
+ void* var_42 = tensorRelu(var_41);
+ void* var_44 = tensorConvolution(var_42, conv2d_12_w, 1, 1, 1, 1, 1, 0);
+ void* var_45 = tensorAdd(var_44, conv2d_12_b);
+ void* var_46 = tensorRelu(var_45);
+ void* var_48 = tensorConvolution(var_46, conv2d_13_w, 1, 1, 1, 1, 1, 0);
+ void* var_49 = tensorAdd(var_48, conv2d_13_b);
+ void* var_50 = tensorRelu(var_49);
+ void* var_51 = tensorPooling(var_50,0,2,2,0,0,2,2);
+ void* var_54 = tensorGemmGPU(var_51, dense_1_w);
+ void* var_55 = tensorAdd(var_54, dense_1_b);
+ void* var_56 = tensorRelu(var_55);
+ void* var_58 = tensorGemmGPU(var_56, dense_2_w);
+ void* var_59 = tensorAdd(var_58, dense_2_b);
+ void* var_60 = tensorSoftmax(var_59);
+
+ uint8_t* labels = readLabelsBatch(labels_path.c_str(),start,end);
+
+ //float accuracy = computeAccuracy2(labels, batch_size, var_60, 100);
+ float accuracy = computeTop5Accuracy(labels, batch_size, var_60, 100);
+ final_accuracy += accuracy;
+ freeBatchMemory();
+
+ }
+
+ final_accuracy = final_accuracy / batch_count;
+ dumpFinalAccuracy(final_accuracy);
+
+ llvm_hpvm_cleanupTensorRt();
+
+ return 0;
+}
--
GitLab