X-Git-Url: https://git.auder.net/?a=blobdiff_plain;f=epclust%2FR%2Fde_serialize.R;h=5a9dd1f236b055a69777e3ff5a3ab18f94007fc7;hb=d9bb53c5e1392018bf67f92140edb10137f3423c;hp=8dde2581545ea442d37a60217cb1dbb8bce38158;hpb=4bcfdbee4e2157f232427a5bfdf240f34760110d;p=epclust.git

diff --git a/epclust/R/de_serialize.R b/epclust/R/de_serialize.R
index 8dde258..5a9dd1f 100644
--- a/epclust/R/de_serialize.R
+++ b/epclust/R/de_serialize.R
@@ -1,22 +1,28 @@
 #' @name de_serialize
 #' @rdname de_serialize
-#' @aliases binarize getDataInFile
+#' @aliases binarize binarizeTransform getDataInFile
 #'
-#' @title (De)Serialization of a matrix
+#' @title (De)Serialization of a [big]matrix or data stream
 #'
 #' @description \code{binarize()} serializes a matrix or CSV file with minimal overhead,
 #'   into a binary file. \code{getDataInFile()} achieves the inverse task: she retrieves
-#'   (ASCII) data rows from indices in the binary file
+#'   (ASCII) data rows from indices in the binary file. Finally,
+#'   \code{binarizeTransform()} serialize transformations of all data chunks; to use it,
+#'   a data-retrieval function must be provided, thus \code{binarize} will most likely be
+#'   used first (and then a function defined to seek in generated binary file)
 #'
 #' @param data_ascii Either a matrix or CSV file, with items in rows
 #' @param indices Indices of the lines to retrieve
 #' @param data_bin_file Name of binary file on output (\code{binarize})
-#'   or intput (\code{getDataInFile})
-#' @param nb_per_chunk Number of lines to process in one batch
+#'   or input (\code{getDataInFile})
+#' @param nb_per_chunk Number of lines to process in one batch (big.matrix or connection)
 #' @inheritParams claws
+#' @param getData Function to retrieve data chunks
+#' @param transform Transformation function to apply on data chunks
 #'
 #' @return For \code{getDataInFile()}, the matrix with rows corresponding to the
-#'   requested indices
+#'   requested indices. \code{binarizeTransform} returns the number of processed lines.
+#'   \code{binarize} is designed to serialize in several calls, thus returns nothing.
 NULL
 
 #' @rdname de_serialize
@@ -24,77 +30,121 @@ NULL
 binarize = function(data_ascii, data_bin_file, nb_per_chunk,
 	sep=",", nbytes=4, endian=.Platform$endian)
 {
+	# data_ascii can be of two types: [big.]matrix, or connection
 	if (is.character(data_ascii))
 		data_ascii = file(data_ascii, open="r")
 	else if (methods::is(data_ascii,"connection") && !isOpen(data_ascii))
 		open(data_ascii)
+	is_matrix = !methods::is(data_ascii,"connection")
 
+	# At first call, the length of a stored row is written. So it's important to determine
+	# if the serialization process already started.
 	first_write = (!file.exists(data_bin_file) || file.info(data_bin_file)$size == 0)
+
+	# Open the binary file for writing (or 'append' if already exists)
 	data_bin = file(data_bin_file, open=ifelse(first_write,"wb","ab"))
 
-	#write data length on first call
 	if (first_write)
 	{
-		#number of items always on 8 bytes
+		# Write data length on first call: number of items always on 8 bytes
 		writeBin(0L, data_bin, size=8, endian=endian)
-		if (is.matrix(data_ascii))
-			data_length = ncol(data_ascii)
-		else #if (is(data, "connection"))
+		if (is_matrix)
+			data_length = nrow(data_ascii)
+		else #connection
 		{
+			# Read the first line to know data length, and write it then
 			data_line = scan(data_ascii, double(), sep=sep, nlines=1, quiet=TRUE)
 			writeBin(data_line, data_bin, size=nbytes, endian=endian)
 			data_length = length(data_line)
 		}
 	}
 
-	if (is.matrix(data_ascii))
+	if (is_matrix)
+	{
+		# Data is processed by chunks; although this may not be so useful for (normal) matrix
+		# input, it could for a file-backed big.matrix. It's easier to follow a unified pattern.
 		index = 1
+	}
 	repeat
 	{
-		if (is.matrix(data_ascii))
+		if (is_matrix)
 		{
-			range = index:min(nrow(data_ascii),index+nb_per_chunk)
 			data_chunk =
-				if (range[1] <= nrow(data_ascii))
-					as.double(t(data_ascii[range,]))
+				if (index <= ncol(data_ascii))
+					as.double(data_ascii[,index:min(ncol(data_ascii),index+nb_per_chunk-1)])
 				else
-					integer(0)
+					double(0)
 			index = index + nb_per_chunk
 		}
-		else
+		else #connection
 			data_chunk = scan(data_ascii, double(), sep=sep, nlines=nb_per_chunk, quiet=TRUE)
+
+		# Data size is unknown in the case of a connection
 		if (length(data_chunk)==0)
 			break
+
+		# Write this chunk of data to the binary file
 		writeBin(data_chunk, data_bin, size=nbytes, endian=endian)
 	}
 
 	if (first_write)
 	{
-		#ecrire file_size-1 / (nbytes*nbWritten) en 0 dans bin_data ! ignored == file_size
+		# Write data_length, = (file_size-1) / (nbytes*nbWritten) at offset 0 in data_bin
 		ignored = seek(data_bin, 0)
 		writeBin(data_length, data_bin, size=8, endian=endian)
 	}
 	close(data_bin)
 
-	if (methods::is(data_ascii,"connection"))
+	if ( ! is_matrix )
 		close(data_ascii)
 }
 
+#' @rdname de_serialize
+#' @export
+binarizeTransform = function(getData, transform, data_bin_file, nb_per_chunk,
+	nbytes=4, endian=.Platform$endian)
+{
+	nb_items = 0 #side-effect: store the number of transformed items
+	index = 1
+	repeat
+	{
+		# Retrieve a chunk of data in a binary file (generally obtained by binarize())
+		data_chunk = getData((index-1)+seq_len(nb_per_chunk))
+		if (is.null(data_chunk))
+			break
+
+		# Apply transformation on the current chunk (by columns)
+		transformed_chunk = transform(data_chunk)
+
+		# Save the result in binary format
+		binarize(transformed_chunk, data_bin_file, nb_per_chunk, ",", nbytes, endian)
+
+		index = index + nb_per_chunk
+		nb_items = nb_items + nrow(data_chunk)
+	}
+	nb_items #number of transformed items
+}
+
 #' @rdname de_serialize
 #' @export
 getDataInFile = function(indices, data_bin_file, nbytes=4, endian=.Platform$endian)
 {
-	data_bin = file(data_bin_file, "rb")
-	data_size = file.info(data_bin_file)$size
+	data_bin = file(data_bin_file, "rb") #source binary file
+
+	data_size = file.info(data_bin_file)$size #number of bytes in the file
+	# data_length: length of a vector in the binary file (first element, 8 bytes)
 	data_length = readBin(data_bin, "integer", n=1, size=8, endian=endian)
-	#Ou t(sapply(...)) (+ rapide ?)
-	data_ascii = do.call( rbind, lapply( indices, function(i) {
+
+	# Seek all 'indices' columns in the binary file, using data_length and nbytes
+	# to compute the offset ( index i at 8 + i*data_length*nbytes )
+	data_ascii = do.call( cbind, lapply( indices, function(i) {
 		offset = 8+(i-1)*data_length*nbytes
 		if (offset > data_size)
-			return (vector("double",0))
-		ignored = seek(data_bin, offset)
+			return (NULL)
+		ignored = seek(data_bin, offset) #position cursor at computed offset
 		readBin(data_bin, "double", n=data_length, size=nbytes, endian=endian)
 	} ) )
 	close(data_bin)
-	if (ncol(data_ascii)>0) data_ascii else NULL
+
+	data_ascii #retrieved data, in columns
 }