-
Notifications
You must be signed in to change notification settings - Fork 26
/
sample.py
384 lines (327 loc) · 16.9 KB
/
sample.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
import os
import yaml
import torch
import argparse
import numpy as np
from tqdm import tqdm
from network import *
from diffusers import DDPMScheduler, PNDMScheduler
from OCC.Extend.DataExchange import write_stl_file, write_step_file
from utils import (
randn_tensor,
compute_bbox_center_and_size,
generate_random_string,
construct_brep,
detect_shared_vertex,
detect_shared_edge,
joint_optimize
)
text2int = {'uncond':0,
'bathtub':1,
'bed':2,
'bench':3,
'bookshelf':4,
'cabinet':5,
'chair':6,
'couch':7,
'lamp':8,
'sofa':9,
'table':10
}
def sample(eval_args):
# Inference configuration
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
batch_size = eval_args['batch_size']
z_threshold = eval_args['z_threshold']
bbox_threshold =eval_args['bbox_threshold']
save_folder = eval_args['save_folder']
num_surfaces = eval_args['num_surfaces']
num_edges = eval_args['num_edges']
if eval_args['use_cf']:
class_label = torch.LongTensor([text2int[eval_args['class_label']]]*batch_size + \
[text2int['uncond']]*batch_size).cuda().reshape(-1,1)
w = 0.6
else:
class_label = None
if not os.path.exists(save_folder):
os.makedirs(save_folder)
surfPos_model = SurfPosNet(eval_args['use_cf'])
surfPos_model.load_state_dict(torch.load(eval_args['surfpos_weight']))
surfPos_model = surfPos_model.to(device).eval()
surfZ_model = SurfZNet(eval_args['use_cf'])
surfZ_model.load_state_dict(torch.load(eval_args['surfz_weight']))
surfZ_model = surfZ_model.to(device).eval()
edgePos_model = EdgePosNet(eval_args['use_cf'])
edgePos_model.load_state_dict(torch.load(eval_args['edgepos_weight']))
edgePos_model = edgePos_model.to(device).eval()
edgeZ_model = EdgeZNet(eval_args['use_cf'])
edgeZ_model.load_state_dict(torch.load(eval_args['edgez_weight']))
edgeZ_model = edgeZ_model.to(device).eval()
surf_vae = AutoencoderKLFastDecode(in_channels=3,
out_channels=3,
down_block_types=['DownEncoderBlock2D', 'DownEncoderBlock2D', 'DownEncoderBlock2D', 'DownEncoderBlock2D'],
up_block_types= ['UpDecoderBlock2D', 'UpDecoderBlock2D', 'UpDecoderBlock2D', 'UpDecoderBlock2D'],
block_out_channels=[128, 256, 512, 512],
layers_per_block=2,
act_fn='silu',
latent_channels=3,
norm_num_groups=32,
sample_size=512,
)
surf_vae.load_state_dict(torch.load(eval_args['surfvae_weight']), strict=False)
surf_vae = surf_vae.to(device).eval()
edge_vae = AutoencoderKL1DFastDecode(
in_channels=3,
out_channels=3,
down_block_types=['DownBlock1D', 'DownBlock1D', 'DownBlock1D'],
up_block_types=['UpBlock1D', 'UpBlock1D', 'UpBlock1D'],
block_out_channels=[128, 256, 512],
layers_per_block=2,
act_fn='silu',
latent_channels=3,
norm_num_groups=32,
sample_size=512
)
edge_vae.load_state_dict(torch.load(eval_args['edgevae_weight']), strict=False)
edge_vae = edge_vae.to(device).eval()
pndm_scheduler = PNDMScheduler(
num_train_timesteps=1000,
beta_schedule='linear',
prediction_type='epsilon',
beta_start = 0.0001,
beta_end = 0.02,
)
ddpm_scheduler = DDPMScheduler(
num_train_timesteps=1000,
beta_schedule='linear',
prediction_type='epsilon',
beta_start = 0.0001,
beta_end = 0.02,
clip_sample = True,
clip_sample_range=3
)
with torch.no_grad():
with torch.cuda.amp.autocast():
###########################################
# STEP 1-1: generate the surface position #
###########################################
surfPos = randn_tensor((batch_size, num_surfaces, 6)).to(device)
pndm_scheduler.set_timesteps(200)
for t in tqdm(pndm_scheduler.timesteps[:158]):#
timesteps = t.reshape(-1).cuda()
if class_label is not None:
_surfPos_ = surfPos.repeat(2,1,1)
pred = surfPos_model(_surfPos_, timesteps, class_label)
pred = pred[:batch_size] * (1+w) - pred[batch_size:] * w
else:
pred = surfPos_model(surfPos, timesteps, class_label)
surfPos = pndm_scheduler.step(pred, t, surfPos).prev_sample
# Late increase for ABC/DeepCAD (slightly more efficient)
if not eval_args['use_cf']:
surfPos = surfPos.repeat(1,2,1)
num_surfaces *= 2
ddpm_scheduler.set_timesteps(1000)
for t in tqdm(ddpm_scheduler.timesteps[-250:]):
timesteps = t.reshape(-1).cuda()
if class_label is not None:
_surfPos_ = surfPos.repeat(2,1,1)
pred = surfPos_model(_surfPos_, timesteps, class_label)
pred = pred[:batch_size] * (1+w) - pred[batch_size:] * w
else:
pred = surfPos_model(surfPos, timesteps, class_label)
surfPos = ddpm_scheduler.step(pred, t, surfPos).prev_sample
#######################################
# STEP 1-2: remove duplicate surfaces #
#######################################
surfPos_deduplicate = []
surfMask_deduplicate = []
for ii in range(batch_size):
bboxes = np.round(surfPos[ii].unflatten(-1,torch.Size([2,3])).detach().cpu().numpy(), 4)
non_repeat = bboxes[:1]
for bbox_idx, bbox in enumerate(bboxes):
diff = np.max(np.max(np.abs(non_repeat - bbox),-1),-1)
same = diff < bbox_threshold
bbox_rev = bbox[::-1] # also test reverse bbox for matching
diff_rev = np.max(np.max(np.abs(non_repeat - bbox_rev),-1),-1)
same_rev = diff_rev < bbox_threshold
if same.sum()>=1 or same_rev.sum()>=1:
continue # repeat value
else:
non_repeat = np.concatenate([non_repeat, bbox[np.newaxis,:,:]],0)
bboxes = non_repeat.reshape(len(non_repeat),-1)
surf_mask = torch.zeros((1, len(bboxes))) == 1
bbox_padded = torch.concat([torch.FloatTensor(bboxes), torch.zeros(num_surfaces-len(bboxes),6)])
mask_padded = torch.concat([surf_mask, torch.zeros(1, num_surfaces-len(bboxes))==0], -1)
surfPos_deduplicate.append(bbox_padded)
surfMask_deduplicate.append(mask_padded)
surfPos = torch.stack(surfPos_deduplicate).cuda()
surfMask = torch.vstack(surfMask_deduplicate).cuda()
#################################
# STEP 1-3: generate surface z #
#################################
surfZ = randn_tensor((batch_size, num_surfaces, 48)).to(device)
pndm_scheduler.set_timesteps(200)
for t in tqdm(pndm_scheduler.timesteps):
timesteps = t.reshape(-1).cuda()
if class_label is not None:
_surfZ_ = surfZ.repeat(2,1,1)
_surfPos_ = surfPos.repeat(2,1,1)
_surfMask_ = surfMask.repeat(2,1)
pred = surfZ_model(_surfZ_, timesteps, _surfPos_, _surfMask_, class_label)
pred = pred[:batch_size] * (1+w) - pred[batch_size:] * w
else:
pred = surfZ_model(surfZ, timesteps, surfPos, surfMask, class_label)
surfZ = pndm_scheduler.step(pred, t, surfZ).prev_sample
########################################
# STEP 2-1: generate the edge position #
########################################
edgePos = randn_tensor((batch_size, num_surfaces, num_edges, 6)).cuda()
pndm_scheduler.set_timesteps(200)
for t in tqdm(pndm_scheduler.timesteps[:158]):
timesteps = t.reshape(-1).cuda()
if class_label is not None:
_surfZ_ = surfZ.repeat(2,1,1)
_surfPos_ = surfPos.repeat(2,1,1)
_surfMask_ = surfMask.repeat(2,1)
_edgePos_ = edgePos.repeat(2,1,1,1)
noise_pred = edgePos_model(_edgePos_, timesteps, _surfPos_, _surfZ_, _surfMask_, class_label)
noise_pred = noise_pred[:batch_size] * (1+w) - noise_pred[batch_size:] * w
else:
noise_pred = edgePos_model(edgePos, timesteps, surfPos, surfZ, surfMask, class_label)
edgePos = pndm_scheduler.step(noise_pred, t, edgePos).prev_sample
ddpm_scheduler.set_timesteps(1000)
for t in tqdm(ddpm_scheduler.timesteps[-250:]):
timesteps = t.reshape(-1).cuda()
if class_label is not None:
_surfZ_ = surfZ.repeat(2,1,1)
_surfPos_ = surfPos.repeat(2,1,1)
_surfMask_ = surfMask.repeat(2,1)
_edgePos_ = edgePos.repeat(2,1,1,1)
noise_pred = edgePos_model(_edgePos_, timesteps, _surfPos_, _surfZ_, _surfMask_, class_label)
noise_pred = noise_pred[:batch_size] * (1+w) - noise_pred[batch_size:] * w
else:
noise_pred = edgePos_model(edgePos, timesteps, surfPos, surfZ, surfMask, class_label)
edgePos = ddpm_scheduler.step(noise_pred, t, edgePos).prev_sample
####################################################
# STEP 2-2: remove duplicate edges per face (bbox) #
####################################################
edgeM = surfMask.unsqueeze(-1).repeat(1, 1, num_edges)
for ii in range(batch_size):
edge_bboxs = edgePos[ii][~surfMask[ii]].detach().cpu().numpy()
for surf_idx, bboxes in enumerate(edge_bboxs):
bboxes = bboxes.reshape(len(bboxes),2,3)
valid_bbox = bboxes[0:1]
for bbox_idx, bbox in enumerate(bboxes):
diff = np.max(np.max(np.abs(valid_bbox - bbox),-1),-1)
bbox_rev = bbox[::-1] # also test reverse bbox for matching
diff_rev = np.max(np.max(np.abs(valid_bbox - bbox_rev),-1),-1)
same = diff < bbox_threshold
same_rev = diff_rev < bbox_threshold
if same.sum()>=1 or same_rev.sum()>=1:
edgeM[ii, surf_idx, bbox_idx] = True
continue # repeat value
else:
valid_bbox = np.concatenate([valid_bbox, bbox[np.newaxis,:,:]],0)
edgeM[ii, surf_idx, 0] = False # set first one to False
##############################
# STEP 2-3: generate edge zv #
##############################
edgeZV = randn_tensor((batch_size, num_surfaces, num_edges, 18)).cuda()
pndm_scheduler.set_timesteps(200)
for t in tqdm(pndm_scheduler.timesteps):
timesteps = t.reshape(-1).cuda()
if class_label is not None:
_surfZ_ = surfZ.repeat(2,1,1)
_surfPos_ = surfPos.repeat(2,1,1)
_edgePos_ = edgePos.repeat(2,1,1,1)
_edgeM_ = edgeM.repeat(2,1,1)
_edgeZV_ = edgeZV.repeat(2,1,1,1)
noise_pred = edgeZ_model(_edgeZV_, timesteps, _edgePos_, _surfPos_, _surfZ_, _edgeM_, class_label)
noise_pred = noise_pred[:batch_size] * (1+w) - noise_pred[batch_size:] * w
else:
noise_pred = edgeZ_model(edgeZV, timesteps, edgePos, surfPos, surfZ, edgeM, class_label)
edgeZV = pndm_scheduler.step(noise_pred, t, edgeZV).prev_sample
edgeZV[edgeM] = 0 # set removed data to 0
edge_z = edgeZV[:,:,:,:12]
edgeV = edgeZV[:,:,:,12:].detach().cpu().numpy()
# Decode the surfaces
surf_ncs = surf_vae(surfZ.unflatten(-1,torch.Size([16,3])).flatten(0,1).permute(0,2,1).unflatten(-1,torch.Size([4,4])))
surf_ncs = surf_ncs.permute(0,2,3,1).unflatten(0, torch.Size([batch_size, num_surfaces])).detach().cpu().numpy()
# Decode the edges
edge_ncs = edge_vae(edge_z.unflatten(-1,torch.Size([4,3])).reshape(-1,4,3).permute(0,2,1))
edge_ncs = edge_ncs.permute(0,2,1).reshape(batch_size, num_surfaces, num_edges, 32, 3).detach().cpu().numpy()
edge_mask = edgeM.detach().cpu().numpy()
edge_pos = edgePos.detach().cpu().numpy() / 3.0
surfPos = surfPos.detach().cpu().numpy() / 3.0
#############################################
### STEP 3: Post-process (per-single CAD) ###
#############################################
for batch_idx in range(batch_size):
# Per cad (not including invalid faces)
surfMask_cad = surfMask[batch_idx].detach().cpu().numpy()
edge_mask_cad = edge_mask[batch_idx][~surfMask_cad]
edge_pos_cad = edge_pos[batch_idx][~surfMask_cad]
edge_ncs_cad = edge_ncs[batch_idx][~surfMask_cad]
edgeV_cad = edgeV[batch_idx][~surfMask_cad]
edge_z_cad = edge_z[batch_idx][~surfMask[batch_idx]].detach().cpu().numpy()[~edge_mask_cad]
surf_z_cad = surfZ[batch_idx][~surfMask[batch_idx]].detach().cpu().numpy()
surf_pos_cad = surfPos[batch_idx][~surfMask_cad]
# Retrieve vertices based on edge start/end
edgeV_bbox = []
for bbox, ncs, mask in zip(edge_pos_cad, edge_ncs_cad, edge_mask_cad):
epos = bbox[~mask]
edge = ncs[~mask]
bbox_startends = []
for bb, ee in zip(epos, edge):
bcenter, bsize = compute_bbox_center_and_size(bb[0:3], bb[3:])
wcs = ee*(bsize/2) + bcenter
bbox_start_end = wcs[[0,-1]]
bbox_start_end = bbox_start_end.reshape(2,3)
bbox_startends.append(bbox_start_end.reshape(1,2,3))
bbox_startends = np.vstack(bbox_startends)
edgeV_bbox.append(bbox_startends)
### 3-1: Detect shared vertices ###
try:
unique_vertices, new_vertex_dict = detect_shared_vertex(edgeV_cad, edge_mask_cad, edgeV_bbox)
except Exception as e:
print('Vertex detection failed...')
continue
### 3-2: Detect shared edges ###
try:
unique_faces, unique_edges, FaceEdgeAdj, EdgeVertexAdj = detect_shared_edge(unique_vertices, new_vertex_dict, edge_z_cad, surf_z_cad, z_threshold, edge_mask_cad)
except Exception as e:
print('Edge detection failed...')
continue
# Decode unique faces / edges
with torch.no_grad():
with torch.cuda.amp.autocast():
surf_ncs_cad = surf_vae(torch.FloatTensor(unique_faces).cuda().unflatten(-1,torch.Size([16,3])).permute(0,2,1).unflatten(-1,torch.Size([4,4])))
surf_ncs_cad = surf_ncs_cad.permute(0,2,3,1).detach().cpu().numpy()
edge_ncs_cad = edge_vae(torch.FloatTensor(unique_edges).cuda().unflatten(-1,torch.Size([4,3])).permute(0,2,1))
edge_ncs_cad = edge_ncs_cad.permute(0,2,1).detach().cpu().numpy()
#### 3-3: Joint Optimize ###
num_edge = len(edge_ncs_cad)
num_surf = len(surf_ncs_cad)
surf_wcs, edge_wcs = joint_optimize(surf_ncs_cad, edge_ncs_cad, surf_pos_cad, unique_vertices, EdgeVertexAdj, FaceEdgeAdj, num_edge, num_surf)
#### 3-4: Build the B-rep ###
try:
solid = construct_brep(surf_wcs, edge_wcs, FaceEdgeAdj, EdgeVertexAdj)
except Exception as e:
print('B-rep rebuild failed...')
continue
# Save CAD model
random_string = generate_random_string(15)
write_step_file(solid, f'{save_folder}/{random_string}_{batch_idx}.step')
write_stl_file(solid, f'{save_folder}/{random_string}_{batch_idx}.stl', linear_deflection=0.001, angular_deflection=0.5)
return
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--mode", type=str, choices=['abc', 'deepcad', 'furniture'], default='abc',
help="Choose between evaluation mode [abc/deepcad/furniture] (default: abc)")
args = parser.parse_args()
# Load evaluation config
with open('eval_config.yaml', 'r') as file:
config = yaml.safe_load(file)
eval_args = config[args.mode]
while(True):
sample(eval_args)