NumPy · dtypes

Every element of an array shares one dtype, and the dtype decides how many bytes each element takes.

dtype	bytes	typical use
int64	8	default integers on most platforms
int32	4	compact integers
float64	8	default floats
float32	4	deep learning (half the memory)
bool	1	masks

• astype converts and returns a copy. Going from float to int truncates toward zero.
• Unlike Python ints, NumPy ints are fixed width, so they wrap around on overflow. That’s the C++ lesson again.
• float32 stores fewer significant digits than float64.

import numpy as np

x = np.array([1, 2, 3])
y = np.array([1.5, 2.7, -3.9])
print(x.dtype, y.dtype, x.itemsize, y.itemsize)

print(y.astype(np.int64))                     # truncates toward zero
small = np.array([1.0, 2.0, 3.0], dtype=np.float32)
print(small.dtype, small.nbytes, "bytes vs", small.astype(np.float64).nbytes)

i8 = np.array([120, 127], dtype=np.int8)
print(i8 + np.int8(10))                       # wraps around!
print(np.iinfo(np.int8), sep="")
print(np.float32(0.1) == np.float64(0.1), np.finfo(np.float32).eps)

Output:

int64 float64 8 8
[ 1  2 -3]
float32 12 bytes vs 24
[-126 -119]
Machine parameters for int8
---------------------------------------------------------------
min = -128
max = 127
---------------------------------------------------------------

False 1.1920929e-07