写在前面
百万张图片的 imagenet 数据原始大小约为 148G，整理成 TFRecord 格式文件后约为 144G，因此至少要准备 300G 大小。

参考：https://github.com/tensorflow/models/tree/master/research/inception#getting-started

1 在 iimagenet 网站 http://image-net.org 注册，记下用户名与密码。
2 下载代码。是tensorflow 的model zoom，https://github.com/tensorflow/models， 代码在models/research/inception/inception/data，
3 修改 download_and_preprocess_imagenet.sh 脚本，如我改的结果如下：
主要是 WORK_DIR，BUILD_SCRIPT，OUTPUT_DIRECTORY 三个，另外，将 BUILD_SCRIPT 设为可执行， 在 build_imagenet_data.py 文件头添加了 #!/usr/bin/env python， 目的是 BUILD_SCRIPT 默认是shell可执行文件。

# usage:
#  ./download_and_preprocess_imagenet.sh [data-dir]
set -e

if [ -z "$1" ]; then
  echo "Usage: download_and_preprocess_imagenet.sh [data dir]"
  exit
fi

# Create the output and temporary directories.
DATA_DIR="${1%/}"
SCRATCH_DIR="${DATA_DIR}/raw-data/"
#mkdir -p "${DATA_DIR}"
#mkdir -p "${SCRATCH_DIR}"
#WORK_DIR="$0.runfiles/inception/inception"
WORK_DIR="/home/linlf/project/linlf/mlperf/reference/image_classification/dataset/inception"
# Download the ImageNet data.
LABELS_FILE="${WORK_DIR}/data/imagenet_lsvrc_2015_synsets.txt"
DOWNLOAD_SCRIPT="${WORK_DIR}/data/download_imagenet.sh"
"${DOWNLOAD_SCRIPT}" "${SCRATCH_DIR}" "${LABELS_FILE}"

# Note the locations of the train and validation data.
TRAIN_DIRECTORY="${SCRATCH_DIR}train/"
VALIDATION_DIRECTORY="${SCRATCH_DIR}validation/"

# Preprocess the validation data by moving the images into the appropriate
# sub-directory based on the label (synset) of the image.
echo "Organizing the validation data into sub-directories."
PREPROCESS_VAL_SCRIPT="${WORK_DIR}/data/preprocess_imagenet_validation_data.py"
VAL_LABELS_FILE="${WORK_DIR}/data/imagenet_2012_validation_synset_labels.txt"

"${PREPROCESS_VAL_SCRIPT}" "${VALIDATION_DIRECTORY}" "${VAL_LABELS_FILE}"

# Convert the XML files for bounding box annotations into a single CSV.
echo "Extracting bounding box information from XML."
BOUNDING_BOX_SCRIPT="${WORK_DIR}/data/process_bounding_boxes.py"
BOUNDING_BOX_FILE="${SCRATCH_DIR}/imagenet_2012_bounding_boxes.csv"
BOUNDING_BOX_DIR="${SCRATCH_DIR}bounding_boxes/"

"${BOUNDING_BOX_SCRIPT}" "${BOUNDING_BOX_DIR}" "${LABELS_FILE}" \
 | sort > "${BOUNDING_BOX_FILE}"
echo "Finished downloading and preprocessing the ImageNet data."

# Build the TFRecords version of the ImageNet data.
#BUILD_SCRIPT="${WORK_DIR}/build_imagenet_data"
BUILD_SCRIPT="${WORK_DIR}/data/build_imagenet_data_newdataset.py"
#OUTPUT_DIRECTORY="${DATA_DIR}"
# mime -- output new dir
OUTPUT_DIRECTORY="/home/linlf/dataset"
IMAGENET_METADATA_FILE="${WORK_DIR}/data/imagenet_metadata.txt"

"${BUILD_SCRIPT}" \
  --train_directory="${TRAIN_DIRECTORY}" \
  --validation_directory="${VALIDATION_DIRECTORY}" \
  --output_directory="${OUTPUT_DIRECTORY}" \
  --imagenet_metadata_file="${IMAGENET_METADATA_FILE}" \
  --labels_file="${LABELS_FILE}" \
  --bounding_box_file="${BOUNDING_BOX_FILE}"

4 在 build_imagenet_data.py 可以自由调整多少个 TFRecord 训练集 和 TFRecord 测试集，在代码110行：
如下我调整成 512 个 TFRecord 训练集和 64 个 TFRecord 测试集。

 tf.app.flags.DEFINE_integer('train_shards', 512,
                            'Number of shards in training TFRecord files.')
 tf.app.flags.DEFINE_integer('validation_shards', 64,
                            'Number of shards in validation TFRecord files.')