Data type discrepancies between numpy and VTK about vedo HOT 11 CLOSED

I see why that happens. Would this be an acceptable solution?

settings.force_single_precision_points = False

a = vedo.Mesh([np.array([[1.5146416416146161461, 2.464164164164161461254, 3.145315134513465161441]], dtype=np.float64), [], []])
print(a.vertices.dtype)
# float64

(just pushed a version on master)

from vedo.

Yep, that's perfect! Thanks!

from vedo.

Hi, actually i'm not completely sure that that is the case.. e.g.

from vedo import *
Sphere().vertices.dtype
dtype('float32')

from vedo.

It doesn't occur with the bunny.obj dummy example but it does occur for high fidelity meshes. For example: https://www.artec3d.com/3d-models/buggy.

import vedo
import numpy as np

path = "Buggy.ply"

mesh = vedo.Mesh(path)
a = mesh.center_of_mass()
b = mesh.vertices.copy()
b.mean(axis=0)
c = b.astype(np.float64)
c.mean(axis=0)

# PLY: (Same for STL)
#    A: array([-160.66834098,   23.57756032,  -87.22478219])
#    B: array([-160.66664 ,   23.578615,  -87.15967 ], dtype=float32)
#    C: array([-160.66834098,   23.57756032,  -87.22478219])
# Difference between B and A/C is floating point precision.

Note that vedo's definition of center of mass is actually just the mean for x, y and z (another topic).

As you can see from the above, computations are done in float64 in VTK but the data type outputted is float32. Therefore, any usage of numpy without explicitly setting the data type may result in this issue.

from vedo.

I don't fully understand though what is the issue here if I run

import vedo

mesh = vedo.Mesh("Buggy.ply")

a = mesh.center_of_mass()
print(a.dtype)

b = mesh.vertices#.copy()
print(b.dtype)

I get

float64
float32

but why do you say that this would cause problems?

from vedo.

There are several areas within the codebase where numpy is used within functions, with np.array(x) or np.array().float(), without specifying the data type. I've had to accommodate for this difference in a few areas. There are a few spots in my codebase where I round to 6 dp for comparisons and recently realised it's due to this issue.

For instance:

• vedo/vedo/utils.py

  Line 791 in db0c0bf

  def vtk2numpy(varr):

• Vertices being converted to float32 here:

  vedo/vedo/file_io.py

  Line 1435 in db0c0bf

  adict["points"] = obj.vertices.astype(float)

• Possibly here too

  vedo/vedo/backends.py

  Line 203 in db0c0bf

  kimage.astype(np.float32),

from vedo.

Hi, the vtk2numpy() function is not overriding the input type so it looks to me it's doing the right thing(?)

The other ones in file_io and backends are rather innocuous.

Note that you should not compare floats to floats and doubles to doubles directly without declaring a tolerance.

from vedo.

Fair. It isn't explicitly mentioned for users developing with vedo that VTK uses doubles by default. Could maybe throw something into the getting started section of the documents but regardless, will close this issue.

from vedo.

This is still causing me problems. Here's a useful example:

a = vedo.Mesh([np.array([[1.5146416416146161461, 2.464164164164161461254, 3145315134513465161441]], dtype=np.float64), [], []])
print(a.vertices)
# array([[1.5146416e+00, 2.4641643e+00, 3.1453151e+21]], dtype=float32)

from vedo.

Hi @marcomusy, currently running into a problem with this solution:

from vedo import Axes, Box, show
settings.force_single_precision_points = False
box = Box(pos=(1,2,3), length=8, width=9, height=7).alpha(0.1)
axs = Axes(box, c='k')

Running this gives: "TypeError: Could not find a suitable VTK type for float16". From what I can tell it looks like a problem with the Axes -- specifically it seems that the mesh for the fonts has its vertices in Float16 which now don't get converted into Float32 anymore.

from vedo.

Thanks for spotting the bug, This should now be fixed in master.

from vedo.