Text Files: savetxt, loadtxt, genfromtxt

Most real data arrives as CSV — a scorecard exported from Excel, a stats file from a cricket website. CSV is plain text: humans can read it and every tool understands it. NumPy can write and read such files directly.

Definition

np.savetxt writes a 1D or 2D array to a text file, one row per line, values separated by a delimiter and formatted with a format string.

np.loadtxt reads a text file of clean, complete numeric data into an array. It fails if a value is missing or not a number.

np.genfromtxt is the forgiving reader: it handles missing values (filling them with nan), header rows, column names and mixed types.

• Delimiter — the character between values: "," for CSV, "\t" for TSV, whitespace by default.
• Format string (fmt) — printf-style: "%d" integer, "%.2f" 2 decimals, "%s" string. Default is "%.18e" (scientific notation, 18 digits).
• nan — "Not a Number", NumPy's float marker for missing data. Use np.isnan, np.nanmean, np.nansum to work with it.

Syntax

np.savetxt(fname, X, fmt="%.18e", delimiter=" ", header="", comments="# ")
np.loadtxt(fname, dtype=float, delimiter=None, skiprows=0, usecols=None)
np.genfromtxt(fname, dtype=float, delimiter=None, skip_header=0, usecols=None,
              names=None, filling_values=None, encoding=None)

The playground has no disk, so the examples use io.StringIO — an in-memory text file. On your computer just pass a file name like "scores.csv".

Example — savetxt: write a CSV

import numpy as np
import io
scores = np.array([[72, 45, 88], [56, 0, 74]])
buf = io.StringIO()
np.savetxt(buf, scores, fmt="%d", delimiter=",", header="m1,m2,m3", comments="")
print(buf.getvalue())

m1,m2,m3
72,45,88
56,0,74

How it works: fmt="%d" writes whole numbers, delimiter="," makes it a CSV, and comments="" stops NumPy from writing # in front of the header (Excel would otherwise show # m1). Without fmt, the default is ugly scientific notation:

buf = io.StringIO()
np.savetxt(buf, np.array([[0.5, 1.25], [2.0, 3.333]]))
print(buf.getvalue())
buf2 = io.StringIO()
np.savetxt(buf2, np.array([[0.5, 1.25], [2.0, 3.333]]), fmt="%.2f", delimiter=",")
print(buf2.getvalue())

5.000000000000000000e-01 1.250000000000000000e+00
2.000000000000000000e+00 3.333000000000000185e+00

0.50,1.25
2.00,3.33

Example — loadtxt: read clean numeric data

csv = io.StringIO("""m1,m2,m3
72,45,88
56,0,74
45,32,0""")
data = np.loadtxt(csv, delimiter=",", skiprows=1, dtype=int)
print(data)
print("per match totals:", data.sum(axis=0))

[[72 45 88]
 [56  0 74]
 [45 32  0]]
per match totals: [173  77 162]

skiprows=1 skips the header text (loadtxt can't turn "m1" into a number); dtype=int gives integers instead of the default floats.

Example — genfromtxt: missing values

csv = io.StringIO("""player,runs,balls
Virat,72,48
Rohit,,30
Hardik,45,NA""")
data = np.genfromtxt(csv, delimiter=",", skip_header=1, usecols=(1, 2))
print(data)
print("missing values:", np.isnan(data).sum())
print("mean runs ignoring NaN:", np.nanmean(data[:, 0]))

[[72. 48.]
 [nan 30.]
 [45. nan]]
missing values: 2
mean runs ignoring NaN: 58.5

Rohit's empty runs cell and Hardik's NA balls both became nan. np.nanmean skips them: (72 + 45) / 2 = 58.5. The same file with loadtxt crashes:

ValueError: could not convert string '' to float64 at row 1, column 2.

Example — text columns and named columns

csv = io.StringIO("""player,runs,balls
Virat,72,48
Rohit,56,30
Hardik,45,25""")
names = np.genfromtxt(csv, delimiter=",", skip_header=1, usecols=0, dtype=str)
print(names)
csv.seek(0)
nums = np.genfromtxt(csv, delimiter=",", names=True, dtype=None, encoding="utf-8")
print(nums["runs"], nums.dtype.names)

['Virat' 'Rohit' 'Hardik']
[72 56 45] ('player', 'runs', 'balls')

dtype=str reads a text column. names=True + dtype=None creates a structured array whose columns you access by name — at this point, though, Pandas (pd.read_csv) is usually the better tool, and it is the next course.

Comparison — loadtxt vs genfromtxt

Common mistakes

• ❌ loadtxt on a file with a header row → ValueError: could not convert string 'm1' to float64 at row 0, column 1. ✅ skiprows=1.
• ❌ Forgetting delimiter="," for a CSV — the default splits on whitespace, so a whole line becomes one "value". ✅ Always set the delimiter.
• ❌ np.mean on data with nan → result is nan. ✅ np.nanmean, np.nansum, np.nanmax.
• ❌ savetxt on a 3D array → error. ✅ Reshape to 2D first or use np.save.

Interview tip

Q: When would you use np.genfromtxt instead of np.loadtxt? A: When the file has missing or non-numeric entries, a header you want as column names, or mixed column types. loadtxt is faster but only handles clean, complete numeric data.