Caffe中计算图像均值的实现(cifar10)

在深度学习中，在进行test时经常会减去train数据集的图像均值，这样做的好处是：属于数据预处理中的数据归一化，降低数据间相似性，可以将数值调整到一个合理的范围。

以下code是用于计算cifar10中训练集的图像均值：

#include "funset.hpp"
#include <string>
#include <vector>
#include <map>
#include "common.hpp"

int campute_image_mean(const std::string& db_type, const std::string& db_path, std::vector<float>& image_mean)
{
#ifdef CPU_ONLY
	caffe::Caffe::set_mode(caffe::Caffe::CPU);
#else
	caffe::Caffe::set_mode(caffe::Caffe::GPU);
#endif

	// reference: caffe/tools/compute_image_mean.cpp
	boost::scoped_ptr<caffe::db::DB> db(caffe::db::GetDB(db_type));
	db->Open(db_path, caffe::db::READ);
	boost::scoped_ptr<caffe::db::Cursor> cursor(db->NewCursor());

	caffe::BlobProto sum_blob;
	int count = 0;
	// load first datum
	caffe::Datum datum;
	datum.ParseFromString(cursor->value());

	caffe::DecodeDatumNative(&datum);

	sum_blob.set_num(1);
	sum_blob.set_channels(datum.channels());
	sum_blob.set_height(datum.height());
	sum_blob.set_width(datum.width());
	const int data_size = datum.channels() * datum.height() * datum.width();
	int size_in_datum = std::max<int>(datum.data().size(), datum.float_data_size());
	for (int i = 0; i < size_in_datum; ++i) {
		sum_blob.add_data(0.);
	}

	// Starting Iteration
	while (cursor->valid()) {
		caffe::Datum datum2;
		datum2.ParseFromString(cursor->value());
		caffe::DecodeDatumNative(&datum2);

		const std::string& data = datum2.data();
		size_in_datum = std::max<int>(datum2.data().size(), datum2.float_data_size());
		if (size_in_datum != data_size) {
			fprintf(stderr, "incorrect data field size: size_in_datum: %d, data_size: %d\n",
				size_in_datum, data_size);
			return -1;
		}
		if (data.size() != 0) {
			if (data.size() != size_in_datum) {
				fprintf(stderr, "data.size() != size_in_datum: data.size: %d, size_in_datum: %d\n",
					data.size(), size_in_datum);
				return -1;
			}
			for (int i = 0; i < size_in_datum; ++i) {
				sum_blob.set_data(i, sum_blob.data(i) + (uint8_t)data[i]);
			}
		} else {
			if (datum2.float_data_size() != size_in_datum) {
				fprintf(stderr, "datum.float_data_size() != size_in_datum: datum.float_data_size: %d, size_in_datum: %d",
					datum2.float_data_size(), size_in_datum);
				return -1;
			}
			for (int i = 0; i < size_in_datum; ++i) {
				sum_blob.set_data(i, sum_blob.data(i) + static_cast<float>(datum2.float_data(i)));
			}
		}
		++count;
		if (count % 10000 == 0) {
			fprintf(stderr, "Processed: %d files\n", count);
		}
		cursor->Next();
	}

	if (count % 10000 != 0) {
		fprintf(stderr, "Processed: %d files\n", count);
	}
	for (int i = 0; i < sum_blob.data_size(); ++i) {
		sum_blob.set_data(i, sum_blob.data(i) / count);
	}

	const int channels = sum_blob.channels();
	const int dim = sum_blob.height() * sum_blob.width();
	image_mean.resize(channels, 0.0);
	fprintf(stderr, "Number of channels: %d", channels);
	for (int c = 0; c < channels; ++c) {
		for (int i = 0; i < dim; ++i) {
			image_mean[c] += sum_blob.data(dim * c + i);
		}
		image_mean[c] /= dim;
	}
	fprintf(stderr, "\n");

	return 0;
}

int cifar10_compute_image_mean()
{
	const std::string db_type{ "lmdb" };
	const std::string db_path{ "E:/GitCode/Caffe_Test/test_data/cifar10/cifar10_train_lmdb" };
	std::vector<float> image_mean;
	if (campute_image_mean(db_type, db_path, image_mean) != 0) {
		fprintf(stderr, "compute image mean fail\n");
		return -1;
	}

	fprintf(stderr, "image mean:");
	for (const auto& value : image_mean) {
		fprintf(stderr, "  %f  ", value);
	}
	fprintf(stderr, "\n");

	return 0;
}

执行结果如下：

GitHub：https://github.com/fengbingchun/Caffe_Test