我正在尝试编写Tensorflow RecordWriter类的纯Java / Scala实现,以便将Spark DataFrame转换为TFRecords文件。根据文档,在TFRecords中,每条记录的格式如下:
uint64 length
uint32 masked_crc32_of_length
byte data[length]
uint32 masked_crc32_of_data
和CRC掩码
masked_crc = ((crc >> 15) | (crc << 17)) + 0xa282ead8ul
目前,我使用以下代码使用guava实现计算CRC:
import com.google.common.hash.Hashing
object CRC32 {
val kMaskDelta = 0xa282ead8
def hash(in: Array[Byte]): Int = {
val hashing = Hashing.crc32c()
hashing.hashBytes(in).asInt()
}
def mask(crc: Int): Int ={
((crc >> 15) | (crc << 17)) + kMaskDelta
}
}
我的其余代码是:
数据编码部分使用以下代码完成:
object LittleEndianEncoding {
def encodeLong(in: Long): Array[Byte] = {
val baos = new ByteArrayOutputStream()
val out = new LittleEndianDataOutputStream(baos)
out.writeLong(in)
baos.toByteArray
}
def encodeInt(in: Int): Array[Byte] = {
val baos = new ByteArrayOutputStream()
val out = new LittleEndianDataOutputStream(baos)
out.writeInt(in)
baos.toByteArray
}
}
使用协议缓冲区生成记录:
import com.google.protobuf.ByteString
import org.tensorflow.example._
import collection.JavaConversions._
import collection.mutable._
object TFRecord {
def int64Feature(in: Long): Feature = {
val valueBuilder = Int64List.newBuilder()
valueBuilder.addValue(in)
Feature.newBuilder()
.setInt64List(valueBuilder.build())
.build()
}
def floatFeature(in: Float): Feature = {
val valueBuilder = FloatList.newBuilder()
valueBuilder.addValue(in)
Feature.newBuilder()
.setFloatList(valueBuilder.build())
.build()
}
def floatVectorFeature(in: Array[Float]): Feature = {
val valueBuilder = FloatList.newBuilder()
in.foreach(valueBuilder.addValue)
Feature.newBuilder()
.setFloatList(valueBuilder.build())
.build()
}
def bytesFeature(in: Array[Byte]): Feature = {
val valueBuilder = BytesList.newBuilder()
valueBuilder.addValue(ByteString.copyFrom(in))
Feature.newBuilder()
.setBytesList(valueBuilder.build())
.build()
}
def makeFeatures(features: HashMap[String, Feature]): Features = {
Features.newBuilder().putAllFeature(features).build()
}
def makeExample(features: Features): Example = {
Example.newBuilder().setFeatures(features).build()
}
}
以下是我如何一起使用以生成TFRecords文件的示例:
val label = TFRecord.int64Feature(1)
val feature = TFRecord.floatVectorFeature(Array[Float](1, 2, 3, 4))
val features = TFRecord.makeFeatures(HashMap[String, Feature] ("feature"->feature, "label"-> label))
val ex = TFRecord.makeExample(features)
val exSerialized = ex.toByteArray()
val length = LittleEndianEncoding.encodeLong(exSerialized.length)
val crcLength = LittleEndianEncoding.encodeInt(CRC32.mask(CRC32.hash(length)))
val crcEx = LittleEndianEncoding.encodeInt(CRC32.mask(CRC32.hash(exSerialized)))
val out = new FileOutputStream(new File("test.tfrecords"))
out.write(length)
out.write(crcLength)
out.write(exSerialized)
out.write(crcEx)
out.close()
当我尝试使用TFRecordReader读取Tensorflow内部的文件时,出现以下错误:
W tensorflow/core/common_runtime/executor.cc:1076] 0x24cc430 Compute status: Data loss: corrupted record at 0
我怀疑CRC掩码计算不正确或字节序 java和c ++之间生成的文件不一样。
答案 0 :(得分:6)
我的实现的问题是CRC掩码的计算。以下是我找到的修复:
def mask(crc: Int): Int ={
((crc >>> 15) | (crc << 17)) + kMaskDelta
}
关键是使用无符号移位按位运算符>>>
而不是>>
答案 1 :(得分:3)
FWIW,Tensorflow团队提供了用于读/写TFRecords的实用程序代码,可以是found in the ecosystem repo