numpy 1.23 breaks scenepick
TadasBaltrusaitis opened this issue · 9 comments
Update to numpy==1.23.0 breaks scenepic 1.0.8 (on Windows and Ubuntu).
Reproduced by using a reduced jelly example:
import scenepic as sp
import numpy as np
scene = sp.Scene()
canvas = scene.create_canvas_3d(width=700, height=700)
jelly_mesh = sp.load_obj(r"C:\Users\tabaltru\Downloads\jelly.obj")
texture = scene.create_image("texture")
texture.load(r"C:\Users\tabaltru\Downloads\jelly.png")
base_mesh = scene.create_mesh("jelly_base")
base_mesh.texture_id = texture.image_id
base_mesh.use_texture_alpha = True
base_mesh.add_mesh(jelly_mesh)
positions = jelly_mesh.positions.copy()
mesh_update = scene.update_mesh_without_normals("jelly_base", positions)
Leads to the following error:
Exception has occurred: TypeError
update_mesh_without_normals(): incompatible function arguments. The following argument types are supported:
1. (self: scenepic._scenepic.Scene, base_mesh_id: str, positions: numpy.ndarray[numpy.float32[m, 3], flags.c_contiguous], mesh_id: str = '') -> scenepic._scenepic.MeshUpdate
Invoked with: <scenepic._scenepic.Scene object at 0x000001FB7FD8ECB0>, 'jelly_base', array([], shape=(0, 3), dtype=float32)
File "C:\code\mrai.moai\moai\sampling\scenepic_test.py", line 17, in <module>
mesh_update = scene.update_mesh_without_normals("jelly_base", positions)
Tried in a clean conda environment both with numpy==1.22.0 and numpy==1.23.0 with only the latter breaking.
I tried to reproduce, but couldn't do so with numpy 1.23.0 and scenepic 1.0.8 on Python 3.9.6 in a virtualenv. The following seems to complete without error (where test.py is the example above):
> virtualenv testnpsp
created virtual environment CPython3.9.6.final.0-64 in 8103ms
creator CPython3Windows(dest=D:\Repos\scenepic\testnpsp, clear=False, no_vcs_ignore=False, global=False)
seeder FromAppData(download=False, pip=bundle, setuptools=bundle, wheel=bundle, via=copy, app_data_dir=~\AppData\Local\pypa\virtualenv)
added seed packages: pip==22.1.2, setuptools==62.6.0, wheel==0.37.1
activators BashActivator,BatchActivator,FishActivator,NushellActivator,PowerShellActivator,PythonActivator
> cd testnpsp
testnpsp> Scripts\activate
(testnpsp) testnpsp> pip install numpy==1.23.0 scenepic==1.0.8
Collecting numpy==1.23.0
Using cached numpy-1.23.0-cp39-cp39-win_amd64.whl (14.7 MB)
Collecting scenepic==1.0.8
Using cached scenepic-1.0.8-cp39-cp39-win_amd64.whl (557 kB)
Collecting pillow>=5.2.0
Using cached Pillow-9.1.1-cp39-cp39-win_amd64.whl (3.3 MB)
Collecting scipy>=1.4.1
Using cached scipy-1.8.1-cp39-cp39-win_amd64.whl (36.9 MB)
Installing collected packages: pillow, numpy, scipy, scenepic
Successfully installed numpy-1.23.0 pillow-9.1.1 scenepic-1.0.8 scipy-1.8.1
(testnpsp) testnpsp> python test.py
scenepic[info]> Disabling VideoWriter due to ImportError
Maybe we need a bit more info on the configuration where the error occurs?
Interesting, here's my pip freeze on a Windows machine with Python 3.8.13 environment managed through Miniconda3:
numpy==1.23.0
Pillow==9.1.1
scenepic==1.0.8
scipy==1.8.1
wincertstore==0.2
Just tried with Python 3.9.12 on my machine and it still fails. Complete repro steps with miniconda:
conda create -n scenepic python=3.9
conda activate scenepic
pip install numpy==1.23
pip install scenepic
Still results in above error, if I do:
pip uninstall numpy
pip install numpy==1.22
The above error is fixed
Pip freeze after installing numpy==1.22, seems like all the other modules are the same, the only difference in numpy
numpy==1.22.0
Pillow==9.1.1
scenepic==1.0.8
scipy==1.8.1
wincertstore==0.2
unable to replicate with python 3.9 or 3.10 😒
Here is a minimal example that replicates the issue we are seeing - I think Tadas' above was not correct:
import scenepic as sp
import numpy as np
scene = sp.Scene()
mesh = scene.create_mesh("a_mesh", shared_color=sp.Colors.Blue)
mesh.add_mesh_without_normals(np.zeros((5, 3), dtype=float), np.zeros((5, 3), dtype=int))For me this works with numpy==1.22 and fails with numpy==1.23:
Traceback (most recent call last):
File "c:\Users\chewitt\Dev\tmp\scenepic_test.py", line 8, in <module>
mesh.add_mesh_without_normals(np.zeros((5, 3), dtype=np.float32), np.zeros((5, 3), dtype=np.uint8))
TypeError: add_mesh_without_normals(): incompatible function arguments. The following argument types are supported:
1. (self: scenepic._scenepic.Mesh, vertices: numpy.ndarray[numpy.float32[m, 3], flags.c_contiguous], triangles: numpy.ndarray[numpy.uint32[m, 3], flags.c_contiguous], colors: numpy.ndarray[numpy.float32[m, 3], flags.c_contiguous] = array([], shape=(0, 3), dtype=float32), uvs: numpy.ndarray[numpy.float32[m, 2], flags.c_contiguous] = array([], shape=(0, 2), dtype=float32), transform: numpy.ndarray[numpy.float32[4, 4]] = array([[1., 0., 0., 0.],
[0., 1., 0., 0.],
[0., 0., 1., 0.],
[0., 0., 0., 1.]], dtype=float32), reverse_triangle_order: bool = False, fill_triangles: bool = True, add_wireframe: bool = False) -> None
Invoked with: {
"CameraSpace": false,
"CommandType": "DefineMesh",
"Definition":
{
"Color": "eAEBDADz/4eGhj6GhQU/AACAPx33A+ADAAAAAQ==",
"IndexBufferType": "UInt16",
"LineBuffer": "eAEDAAAAAAEAAAAAAg==",
"PrimitiveType": "SingleColorMesh",
"TriangleBuffer": "eAEDAAAAAAEAAAAAAw==",
"VertexBuffer": "eAEDAAAAAAEAAAAABg=="
},
"DoubleSided": false,
"IsBillboard": false,
"IsLabel": false,
"LayerId": null,
"MeshId": "jelly_base"
}
, array([[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.],
[0., 0., 0.]], dtype=float32), array([[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
[0, 0, 0],
[0, 0, 0]], dtype=uint8)
@matajoh - any ideas on this, currently breaking a lot of our CI :(
I think this is the root issue: pybind/pybind11#4038 (and hence will be fixed by the next pybind11 release, or by pulling pybind11 from source at commit 479e9a5 or later).
I verified that Charlie's repro leads to an array with zero-size dimensions in pybind11, which leads to EigenConformable::stride_compatible returning false for earlier versions of pybind11.
Local test confirmed that @tjcashman's fix works so this issue can be closed 😄
Closing this issue as resolved, thanks for the help everyone!