Problem

[dclcs]

When I run the data_prep.py , I have a problem : found bundled python: /media/dcl/Data1/blender/2.79/python Read blend: /media/dcl/Data1/paper_code/GEOMetrics-master/scripts/manage.blend ( 0.0001 sec | 0.0001 sec) Importing OBJ 'data/objects/rifle/d442863599da17d951cfdb4c9f126c12.obj'... ( 0.0004 sec | 0.0003 sec) Parsing OBJ file... ( 0.0307 sec | 0.0304 sec) Done, loading materials and images... ProgressMaterial not found MTL: 'data/objects/rifle/model.mtl' ( 0.0313 sec | 0.0309 sec) Done, building geometries (verts:977 faces:4244 materials: 9 smoothgroups:1) ... ( 0.0667 sec | 0.0664 sec) Done. ( 0.0668 sec | 0.0667 sec) Finished importing: 'data/objects/rifle/d442863599da17d951cfdb4c9f126c12.obj' Progress: 100.00%

Writing: /tmp/manage.crash.txt 段错误 (核心已转储) Is the blender version problem? I am not familiar with blender.....

[EdwardSmith1884]

you figured it out?

[KnightOfTheMoonlight]

@EdwardSmith1884 Which version of blender did you use?

[EdwardSmith1884]

version 2.78 I think. I have not checked with the new 2.8 version, the scripts make not still work

[KnightOfTheMoonlight]

@EdwardSmith1884 Yes, 2.79b works for me too.

[EdwardSmith1884]

great!

[dysdsyd]

I am also facing the same problem with blender 2.79b. After tracking down the problem, I found that it was coming from:

bpy.ops.object.join(ctx) line 40 of blender_convert.py

After running the following blender command independently:

blender scripts/manage.blend -b -P scripts/blender_convert.py -- data/objects/cabinet/ba17ef05393844dfcb7105765410e2d6.obj data/mesh_info/cabinet/ba17ef05393844dfcb7105765410e2d6 data/managable_objects/cabinet/ba17ef05393844dfcb7105765410e2d6.obj

I am getting:

[EdwardSmith1884]

Does This occur for every object? Or for even more than a few? Perhaps you could just ignore these objects in the dataset, my data-loader should handle their absence. Alternatively you could just ignore the blender step in the datamaking script, this only effects the latent loss which is turned off by default.

[dysdsyd]

This is occurring for all the objects, I am hoping to get the downscaled meshes for an experiment that's why I can't skip over this step. I believe this is a problem with the blender code, I am using blender 2.79b as I was unable to run 2.78 on my server.

[EdwardSmith1884]

I am currently using blender 2.79 and this seems to work for me:

bpy.ops.import_scene.obj(filepath=obj_location)
obs = []
for ob in scene.objects:
    if ob.type == 'MESH' :
        obs.append(ob)

bpy.data.objects[ob.name].select_set(True)
bpy.context.view_layer.objects.active = ob
c = {}
c["object"] = c["active_object"] = bpy.context.object
c["selected_objects"] = c["selected_editable_objects"] = obs
bpy.ops.object.join(c)

can you let me know if this fixes your issue?

[dysdsyd]

Can you please share the updated blender_convert.py file? I am not very well versed with blender so, not exactly sure what lines to remove.

[EdwardSmith1884]

I won't post a new file tonight as I don't have access to Blender right now but if you replace:

obs = []
for ob in scene.objects:
    if ob.type == 'MESH':
        obs.append(ob)
ctx = bpy.context.copy()
ctx['active_object'] = obs[0]
ctx['selected_objects'] = obs
ctx['selected_editable_bases'] = [scene.object_bases[ob.name] for ob in obs]
bpy.ops.object.join(ctx)
o = bpy.context.selected_objects[0]

# removes split normal, helps with decimation
bpy.context.scene.objects.active = o 

with

obs = []
for ob in scene.objects:
    if ob.type == 'MESH' :
        obs.append(ob)

bpy.data.objects[ob.name].select_set(True)
bpy.context.view_layer.objects.active = ob
c = {}
c["object"] = c["active_object"] = bpy.context.object
c["selected_objects"] = c["selected_editable_objects"] = obs
bpy.ops.object.join(c)
bpy.context.scene.objects.active = ob 

it should work, if not let me known and I'll send a file some time this week.

[EdwardSmith1884]

This should work for newer versions of blender, at least it did for 2.81:

import os
import bpy
import bmesh
import scipy.io as sio
import sys
import numpy as np

# this whole file is to make managable obj files from overly large or samll ones
# this is done for the latent loss calcualtions only

def triangulate_edit_object(obj):
    me = obj.data
    bm = bmesh.from_edit_mesh(me)
    bmesh.ops.triangulate(bm, faces=bm.faces[:], quad_method='BEAUTY', ngon_method='BEAUTY')
    bmesh.update_edit_mesh(me, True)

# import arguements
model = sys.argv[-3]
location_info = sys.argv[-2]
location_obj = sys.argv[-1]

# import object
bpy.ops.import_scene.obj(filepath=model)
scene = bpy.context.scene

# join components of mesh
obs = []
for o in scene.objects:
    if o.type == 'MESH' :
        obs.append(o)

bpy.data.objects[o.name].select_set(True)
bpy.context.view_layer.objects.active = o
c = {}
c["object"] = c["active_object"] = bpy.context.object
c["selected_objects"] = c["selected_editable_objects"] = obs

bpy.ops.object.join(c)
bpy.context.view_layer.objects.active = o
bpy.ops.object.editmode_toggle()
bpy.ops.mesh.select_all(action='SELECT')
bpy.ops.mesh.customdata_custom_splitnormals_clear()
bpy.ops.mesh.remove_doubles()
bpy.ops.object.editmode_toggle()

# shrinking the mesh to be a uniform size
# no idea if this actually helps with training
# it does make objects smaller which makes loading them much quicker to load during traineing
# ideally all objects will be between 500 and 600 verts, but I allow between 400 and 700 verts
# the object is used during training regardless, but not for the latent loss
not_possible = False
num = float(len(o.data.vertices))
new_num = num
orig = num
full_ratio = .01
if num < 30:
    not_possible = True  # if its too small then subsampling doesnt work well

# if large then decimate to make the right size
# or at least try to
elif num > 550:
    for i in range(5):
        mod = o.modifiers.new(name='decimate', type='DECIMATE')
        mod.ratio = max(550. / num, full_ratio)
        full_ratio /= (mod.ratio)
        bpy.ops.object.modifier_apply(modifier=mod.name)
        o.modifiers.clear()
        if float(len(o.data.vertices)) < 550:
            new_num = float(len(o.data.vertices))
            break
        else:
            num = float(len(o.data.vertices))
        if i == 4 and float(len(o.data.vertices)) > 700:  # if it can't be made small enough then don't convert it
            not_possible = True
        new_num = float(len(o.data.vertices))
# if small then try to make it larger
elif num < 400:
    mod = o.modifiers.new(name="Remesh", type='REMESH')
    mod.octree_depth = 6
    mod.use_remove_disconnected = False
    bpy.ops.object.modifier_apply(modifier=mod.name)
    o.modifiers.clear()
    num = float(len(o.data.vertices))
    # and then shrink it again
    if num > 500:
        for i in range(5):
            mod = o.modifiers.new(name='decimate', type='DECIMATE')
            mod.ratio = 550. / num
            bpy.ops.object.modifier_apply(modifier=mod.name)
            o.modifiers.clear()
            if float(len(o.data.vertices)) < 600:
                break
            else:
                num = float(len(o.data.vertices))
    else:
        not_possible = True
print('-------------------------------------------------------------------------')
print(float(len(o.data.vertices)))
print('-------------------------------------------------------------------------')
if not_possible:
    exit()

# triangluate the object
bpy.ops.object.editmode_toggle()
bpy.ops.mesh.dissolve_limited()
triangulate_edit_object(o)
bpy.ops.object.editmode_toggle()

# now we record the object info

# get initial face info
me = o.data
adj_new = np.zeros((600, 600))
max_len = 0
faces = []
for poly in me.polygons:
    vs = []
    for loop_index in range(poly.loop_start, poly.loop_start + poly.loop_total):
        vs.append(me.loops[loop_index].vertex_index)
    faces.append(vs)

# get initial vertex info, and normal info if you want it (I dont)
bm = bmesh.new()
bm.from_mesh(me)
verts, normals = [0 for i in range(len(bm.verts))], [0 for i in range(len(bm.verts))]
for e, v in enumerate(bm.verts):
    verts[v.index] = v.co
    normals[v.index] = v.normal

# calculate adjacency matrix and final face, vertex and nromal infor
verts_map = {}
count = 0
for face in faces:
    v1, v2, v3 = face
    for v in face:
        if v not in verts_map:
            verts_map[v] = [count, verts[v], normals[v]]
            count += 1
adj = np.zeros((len(verts_map), len(verts_map)))
true_verts = np.zeros((len(verts_map), 3))
true_normals = np.zeros((len(verts_map), 3))
for e, face in enumerate(faces):
    v1, v2, v3 = face
    adj[verts_map[v1][0], verts_map[v1][0]] = 1
    adj[verts_map[v2][0], verts_map[v2][0]] = 1
    adj[verts_map[v3][0], verts_map[v3][0]] = 1
    adj[verts_map[v1][0], verts_map[v2][0]] = 1
    adj[verts_map[v2][0], verts_map[v1][0]] = 1
    adj[verts_map[v1][0], verts_map[v3][0]] = 1
    adj[verts_map[v3][0], verts_map[v1][0]] = 1
    adj[verts_map[v2][0], verts_map[v3][0]] = 1
    adj[verts_map[v3][0], verts_map[v2][0]] = 1
    faces[e] = [verts_map[v1][0], verts_map[v2][0], verts_map[v3][0]]

for _, info in verts_map.items():
    spot, position, normal = info
    true_verts[spot] = position
    true_normals[spot] = normal

for obj in bpy.data.objects:
    obj.select_set(False)
o.select_set(True)

# save updated object, and object info
bpy.ops.export_scene.obj(filepath=location_obj)
bpy.ops.export_scene.obj(filepath=location_obj)
sio.savemat(location_info, {'verts':np.array(true_verts), 
                'normals': np.array(true_normals), 
                'faces': np.array(faces),  
                'orig_adj': adj
                }
                )