Open a MATLAB .mat file without MATLAB, and save its variables as CSV

What it shows

• The file's own header. A Level 5 file starts with 116 bytes of text such as MATLAB 5.0 MAT-file, Platform: GLNXA64, Created on: Tue Mar 12 09:14:02 2024: who wrote it, on which system and when. It is shown as written, with the byte order and how many variables are compressed.
• Every variable, like whos. Name, class (double, single, int8 to uint64, logical, char, cell, struct, sparse), dimensions such as 1000×64×3, whether it is complex, sparse or global, and its size in bytes counted the way MATLAB's whos counts it.
• Matrices as a table. Numeric and logical arrays open as a grid with row and column numbers counted from 1, as in MATLAB. The first 1,000 rows and 200 columns are drawn, with buttons for the next block and the true size stated above. A 3-D or larger array is shown one 2-D slice at a time: pick the index of each extra dimension, and the table shows X(:, :, k). Complex values show as 1.5-0.25i.
• Text. A char array is shown as text, one line per row, and as a table of characters. A 2×5 char matrix holding hello and world reads as those two lines.
• Structs and cell arrays as a tree. Tap a branch to open it: results.trial(3).rt or data{2}. Short values show in place; a matrix or text inside has an Open button that shows it as a table, ready to save as CSV.
• Sparse matrices as triplets. The non-zeros as (row, column, value), column by column as MATLAB stores them, with how full the matrix is. Save them as a three-column CSV, which sparse(r, c, v) in MATLAB or scipy.sparse.coo_matrix turns back into the matrix.

Saving as CSV and JSON

• Save as CSV writes the whole 2-D variable, or the slice you picked of an N-D one, not only the part drawn on screen: one line per row, values separated by commas. Numbers are written at full precision in the shortest form that reads back to the same double (0.1, not 0.10000000000000001), integers exactly (an int64 above 2⁵³ keeps every digit), NaN, Inf and -Inf as MATLAB writes them, logical values as 1 and 0.
• Save all as JSON writes every variable in one file: matrices as arrays of rows, N-D arrays as their size plus the values in MATLAB's column-major order, text as strings, cell arrays and structs nested as they are in the file. NaN and Inf become the strings "NaN" and "Inf", because JSON has no way to write them.

The formats, and what this reads in each

v7 (MATLAB's default)

What save writes unless told otherwise: a Level 5 file in which each variable is zlib-compressed on its own (the miCOMPRESSED data type). It is unpacked here with the bundled fflate library. Files from GNU Octave (save -v7) and SciPy (savemat(..., do_compression=True)) are the same format.

v6 and v5

The same Level 5 layout without compression: save -v6, MATLAB 5 to 6.5, and scipy.io.savemat by default. Little-endian (IM) and big-endian (MI, from old Sun and Mac machines) files both open.

Level 4

MATLAB 4's format, still written by save -v4, Octave, SciPy (format='4') and some instruments: a 20-byte header per matrix and its values, no compression, numeric and text matrices only (and sparse ones as a list of triplets). A Level 4 file has no signature, so outside this page viewhack only recognises it when its name ends in .mat.

v7.3

An HDF5 file behind a 512-byte MAT-file header, written by save -v7.3 and needed for variables over 2 GB. This page recognises it and says so, but cannot read it yet. Re-save it in MATLAB with S = load('data.mat'); save('data_v7.mat', '-struct', 'S', '-v7'); GNU Octave can usually load a v7.3 file of plain arrays and save it the same way.

What this cannot do

• Read v7.3 files. They are HDF5 inside, a different format altogether; see above for the two-line re-save.
• Show function handles. A saved @(x) x.^2 is listed by name and class only; what it points to is kept in MATLAB's own private encoding.
• Show MATLAB class objects. Objects of classdef classes, including string, table, timetable, datetime and categorical, are stored in an undocumented subsystem format and are listed by name and class only. Convert them in MATLAB before saving (table2array, char, datenum, double). Old-style @class objects show their fields.
• Edit or write .mat files. It reads and exports; it does not change values or save a .mat back.
• Run MATLAB code, or plot. It shows numbers, it does not compute with them or draw graphs. For that, the CSV opens in a spreadsheet, and GNU Octave reads the .mat itself for free.

Opening the CSV elsewhere

Excel or LibreOffice

Open the .csv directly. The values have no header row, just as the matrix has no column names; row 1 of the sheet is row 1 of the matrix.

Python

numpy.loadtxt('A.csv', delimiter=',') gives back the same matrix. NaN and Inf read correctly; complex values (written 1+2i) need MATLAB's i changed to Python's j first.

R

as.matrix(read.csv('A.csv', header = FALSE)).

MATLAB is a trademark of The MathWorks, Inc. viewhack is not affiliated with The MathWorks; this reader was written from the MAT-file format description MathWorks publishes.