提问人:Matthew J. Oldach 提问时间:3/19/2020 最后编辑:WaldiMatthew J. Oldach 更新时间:2/2/2021 访问量:293
从 JPEG 创建 Spark 对象并在非转换函数上使用 spark_apply()
Creating Spark Objects from JPEG and using spark_apply() on a non-translated function
问:
通常,当人们想要在自定义函数(即**非转换函数)上使用时,他们会将它们放在 spark_apply()
中。但是,我只遇到过一些示例,其中单个本地数据帧是 or 到远程数据源,然后在其上使用。举个例子,仅供说明之用:sparklyr
copy_to()
spark_read_csv()
spark_apply()
library(sparklyr)
sc <- spark_connect(master = "local")
n_sim = 100
iris_samps <- iris %>% dplyr::filter(Species == "virginica") %>%
sapply(rep.int, times=n_sim) %>% cbind(replicate = rep(1:n_sim, each = 50)) %>%
data.frame() %>%
dplyr::group_by(replicate) %>%
dplyr::sample_n(50, replace = TRUE)
iris_samps_tbl <- copy_to(sc, iris_samps)
iris_samps_tbl %>%
spark_apply(function(x) {mean(x$Petal_Length)},
group_by = "replicate") %>%
ggplot(aes(x = result)) + geom_histogram(bins = 20) + ggtitle("Histogram of 100 Bootstrapped Means using sparklyr")
因此,只要数据驻留在 Spark 对象中,就可以将其用于来自或包的任何范围的非转换函数。CRAN
Bioconductor
当我读到 SparkR
可以通过 Java 库中的 ImageIO
将压缩图像(.jpeg
、.png
等)加载到原始图像表示中时,我提出了一个针对图像的特定问题 - 似乎也可以做到这一点。.jpeg
sparklyr
RsimMosaic::composeMosaicFromImageRandom(inputImage, outputImage, pathToTilesLibrary)
函数获取输入图像和用于创建照片马赛克的图块路径,并输出图像(此处为示例)。
如果这个函数只拍摄一张图像,并且我知道如何将其转换为火花对象,我可能会想象该命令将如下所示:.但是,此函数正在获取 30,000 张图像的路径。composeMosaicFromImageRandom(inputImage, outputImage, spark_obj)
如何从这些图块 () 的路径创建 30,000 个 Spark 对象,然后使用此功能?.jpegs
如果底层代码确实需要修改,我曾经提供源代码:jimhester/lookup
function (originalImageFileName, outputImageFileName, imagesToUseInMosaic,
useGradients = FALSE, removeTiles = TRUE, fracLibSizeThreshold = 0.7,
repFracSize = 0.25, verbose = TRUE)
{
if (verbose) {
cat(paste("\n ------------------------------------------------ \n"))
cat(paste(" R Simple Mosaic composer - random version \n"))
cat(paste(" ------------------------------------------------ \n\n"))
}
if (verbose) {
cat(paste(" Creating the library... \n"))
}
libForMosaicFull <- createLibraryIndexDataFrame(imagesToUseInMosaic,
saveLibraryIndex = F, useGradients = useGradients)
libForMosaic <- libForMosaicFull
filenameArray <- list.files(imagesToUseInMosaic, full.names = TRUE)
originalImage <- jpeg::readJPEG(filenameArray[1])
xTileSize <- dim(originalImage[, , 1])[1]
yTileSize <- dim(originalImage[, , 1])[2]
if (verbose) {
cat(paste(" -- Tiles in the Library : ", length(libForMosaic[,
1]), "\n"))
cat(paste(" -- Tile dimensions : ", xTileSize, " x ",
yTileSize, "\n"))
}
if (verbose) {
cat(paste("\n"))
cat(paste(" Reading the original image... \n"))
}
originalImage <- jpeg::readJPEG(originalImageFileName)
xOrigImgSize <- dim(originalImage[, , 1])[1]
yOrigImgSize <- dim(originalImage[, , 1])[2]
if (verbose) {
cat(paste(" -- Original image dimensions : ", xOrigImgSize,
" x ", yOrigImgSize, "\n"))
cat(paste(" -- Output image dimensions : ", ((xOrigImgSize -
2) * xTileSize), " x ", ((yOrigImgSize - 2) * yTileSize),
"\n"))
}
if (verbose) {
cat(paste("\n"))
cat(paste(" Computing the mosaic... \n"))
}
outputImage <- array(dim = c(((xOrigImgSize - 2) * xTileSize),
((yOrigImgSize - 2) * yTileSize), 3))
removedList <- c()
l <- 1
pCoord <- matrix(nrow = ((xOrigImgSize - 2) * (yOrigImgSize -
2)), ncol = 2)
for (i in 2:(xOrigImgSize - 1)) {
for (j in 2:(yOrigImgSize - 1)) {
pCoord[l, 1] <- i
pCoord[l, 2] <- j
l <- l + 1
}
}
npixels <- length(pCoord[, 1])
for (i in 1:npixels) {
idx <- round(runif(1, 1, length(pCoord[, 1])))
pixelRGBandNeigArray <- computeStatisticalQuantitiesPixel(pCoord[idx,
1], pCoord[idx, 2], originalImage, useGradients)
tileFilename <- getCloseMatch(pixelRGBandNeigArray,
libForMosaic)
startI <- (pCoord[idx, 1] - 2) * xTileSize + 1
startJ <- (pCoord[idx, 2] - 2) * yTileSize + 1
outputImage[startI:(startI + xTileSize - 1), startJ:(startJ +
yTileSize - 1), ] <- jpeg::readJPEG(tileFilename)
if (removeTiles) {
libForMosaic <- removeTile(tileFilename, libForMosaic)
removedList <- c(removedList, tileFilename)
if (length(libForMosaic[, 1]) < (fracLibSizeThreshold *
length(libForMosaicFull[, 1]))) {
idxs <- runif(round(0.25 * length(libForMosaicFull[,
1])), 1, length(removedList))
for (ii in 1:length(idxs)) {
libForMosaic <- addBackTile(removedList[idxs[ii]],
libForMosaic, libForMosaicFull)
}
removedList <- removedList[-idxs]
}
}
if (length(pCoord[, 1]) > 2) {
pCoord <- pCoord[-idx, ]
}
}
if (verbose) {
cat(paste("\n"))
cat(paste(" Done!\n\n"))
}
jpeg::writeJPEG(outputImage, outputImageFileName)
}
请注意:我第一次尝试加速这段代码是 1) use 查找瓶颈(即 for 循环)2) 在 for 循环上使用包。这导致代码速度变慢,这表明我的并行化水平太低。据我所知,更多的是关于分布式计算,而不是并行化它,所以也许这可以工作。profvis
foreach
sparklyr
答: 暂无答案
评论