Textscan on file with large number of lines

问题

I'm trying to analyze a very large file using textscan in MATLAB. The file in question is about 12 GB in size and contains about 250 million lines with seven (floating) numbers in each (delimited by a whitespace); because this obviously would not fit into the RAM of my desktop, I'm using the approach suggested in the MATLAB documentation (i.e. loading and analyzing a smaller block of the file at a time. According to the documentation this should allow for processing "arbitrarily large delimited text file[s]"). This only allows me to scan about 43% of the file, after which textscan starts returning empty cells (despite there still being data left to scan in the file).

To debug, I attempted to go to several positions in the file using the fseek function, for example like this:

fileInfo = dir(fileName);
fid = fileopen(fileName);
fseek(fid, floor(fileInfo.bytes/10), 'bof');
textscan(fid,'%f %f %f %f %f %f %f','Delimiter',' ');

I'm assuming that the way I'm using fseek here moves the position indicator to about 10% of my file. (I'm aware this doesn't necessarily mean the indicator is at the beginning of a line, but if I run textscan twice I get a satisfactory answer.) Now, if I substitute fileInfo.bytes/10 by fileInfo.bytes/2 (i.e. moving it to about 50% of the file) everything breaks down and textscan only returns an empty 1x7 cell.

I looked at the file using a text editor for large files, and this shows that the entire file looks fine, and that there should be no reason for textscan to be confused. The only possible explanation that I can think of is that something goes wrong on a much deeper level that I have little understanding of. Any suggestions would be greatly appreciated!

EDIT

The relevant part of my code used to look like this:

while ~feof(fid)
    data = textscan(fid, FormatString, nLines, 'Delimiter', ' '); %// Read nLines
        %// do some stuff
end

First I tried fixing it using ftell and fseek as suggested by Hoki below. This gave exactly the same error as I got before: MATLAB was unable to read in more than approximately 43% of the file. Then I tried using the HeaderLines solution (also suggested below), like this:

i = 0;
while ~feof(fid)
    frewind(fid)
    data = textscan(fid, FormatString, nLines, 'Delimiter',' ', 'HeaderLines', i*nLines);
        %// do some stuff
    i = i + 1;
end

This seems to read in the data without producing errors; it is, however, incredibly slow.

I'm not entirely sure I understand what HeaderLines does in this context, but it seems to make textscan completely ignore everything that comes before the specified line. This doesn't seem to happen when using textscan in the "appropriate" way (either with or without ftell and fseek): in both cases it tries to continue from its last position, but to no avail because of some reason I don't understand yet.

回答1:

fseek a pointer in a file is only good when you know precisely where (or by how many bytes) you want to move the cursor. It is very useful for binary files when you just want to skip some records of known length. But on a text file it is more dangerous and confusing than anything (unless you are absolutely sure that each line is the same size and each element on the line is at the same exact place/column, but that doesn't happen often).

There are several ways to read a text file block by block:

1) Use the HeaderLines option

To simply skip a block of lines on a text file, you can use the HeaderLines parameter of textscan, so for example:

readFormat = '%f %f %f %f %f %f %f' ;   %// read format specifier
nLines = 10000 ;                        %// number of line to read per block

fileInfo = dir(fileName);

%// read FIRST block
fid = fileopen(fileName);
M = textscan(fid, readFormat, nLines,'Delimiter',' '); %// read the first 10000 lines
fclose(fid)
    %// Now do something with your "M" data

Then when you want to read the second block:

%// later read the SECOND block:
fid = fileopen(fileName);
M = textscan(fid, readFormat, nLines,'Delimiter',' ','HeaderLines', nLines); %// read lines 10001 to 20000
fclose(fid)

And if you have many blocks, for the Nth block, just adapt:

%// and then for the Nth BLOCK block:
fid = fileopen(fileName);
M = textscan(fid, readFormat, nLines,'Delimiter',' ','HeaderLines', (N-1)*nLines);
fclose(fid)

If necessary (if you have many blocks), just code this last version in a loop.

Note that this is good if you close your file after each block reading (so the file pointer will start at the beginning of the file when you open it again). Closing the file after reading a block of data is safer if your processing might take a long time or may error out (you don't want to have files which remain open too long or loose the fid if you crash).

---

2) Read by block (without closing the file)

If the processing of the block is quick and safe enough so you're sure it won't bomb out, you could afford to not close the file. In this case, the textscan file pointer will stay where you stopped, so you could also :

• read a block (do not close the file): M = textscan(fid, readFormat, nLines)
• Process it then save your result (and release memory)
• read the next block with the same call: M = textscan(fid, readFormat, nLines)

In this case you wouldn't need the headerlines parameter because textscan will resume reading exactly where it stopped.

---

3) use ftell and fseek

Lastly, you could use fseek to start reading the file at the precise position you want, but in this case I recommend using it in conjunction with ftell.

ftell will return the current position in an open file, so use that to know at which position you stop reading last, then use fseek the next time to go straight at this position. Something like:

%// read FIRST block
fid = fileopen(fileName);
M = textscan(fid, readFormat, nLines,'Delimiter',' ');
lastPosition = ftell(fid) ;
fclose(fid)

%// do some stuff

%// then read another block:
fid = fileopen(fileName);
fseek( fid , 'bof' , lastPosition ) ;
M = textscan(fid, readFormat, nLines,'Delimiter',' ');
lastPosition = ftell(fid) ;
fclose(fid)
%// and so on ...

来源：https://stackoverflow.com/questions/32095944/textscan-on-file-with-large-number-of-lines