-
Notifications
You must be signed in to change notification settings - Fork 109
/
phase1_inference.py
447 lines (357 loc) · 21 KB
/
phase1_inference.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
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
# -*- coding: utf-8 -*-
"""
@date: 2023.03.29-31 week13
@func: PhaseI inference code.
"""
import argparse
import datetime
import torch
import torchvision
import torch.nn as nn
import torch.optim as optim
import numpy as np
import pickle
import os
import os.path as osp
import torchvision
import torchvision.transforms as transforms
from torchvision.models.feature_extraction import create_feature_extractor, get_graph_node_names
import torch.nn.functional as F
import thinplate as tps
import time
from renderer.cloth_renderer import ClothRenderer
import matplotlib.pyplot as plt
from PIL import Image
import importlib
import random
from utils.frequency import extract_ampl_phase
from utils.binary_function import Binarize
from utils.tvl_loss import TVLoss, TVMaskLoss
from tqdm import tqdm
import json
from pytorch3d.io import load_obj, save_obj
import cv2
from itertools import chain
from pytorch3d.structures import Meshes
from pytorch3d.transforms import RotateAxisAngle
from pytorch3d.loss import (
mesh_edge_loss,
mesh_laplacian_smoothing,
mesh_normal_consistency,
)
from lib.deformation_graph import DeformationGraph
from lib.mesh_sampling import generate_transform_matrices_coma
from lib.utils_dg import to_edge_index, to_sparse, get_vert_connectivity, scipy_to_torch_sparse
from models import DeformGraphModel
from torch_geometric.transforms import FaceToEdge
from torch_geometric.data import Data
from psbody.mesh import Mesh
from torch_geometric.io import read_ply
class Trainer(object):
def __init__(self, objfile, savedir, resolution=512, focal_distance=2, verts_num=9648, scale_factor=1.0):
self.device = torch.device("cuda")
#set mesh and visualizer----------------------
self.cloth_renderer = ClothRenderer(objfile, resolution, focal_distance, scale_factor)
if os.path.exists(os.path.join("experiments", savedir)):
pass
else:
os.makedirs(os.path.join("experiments", savedir))
self.savedir = savedir
self.uv = torch.ones((1, 512, 512, 3)).cuda()
self.uv.requires_grad = True
self.optimizer = optim.Adam([self.uv], lr=5e-3, betas=(0.5, 0.999))
# define loss
self.criterion = nn.MSELoss() # nn.L1Loss() nn.MSELoss()
self.mse = nn.MSELoss()
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# You can choose TVMaskLoss and test if it is suitable for your case.
self.tvl_loss = TVLoss(weight=1) # TVMaskLoss(weight=1) or TVLoss(weight=1)
# self.tvl_loss = TVMaskLoss(weight=1)
self.canonical_mesh = self.cloth_renderer.canonical_mesh
self.deform_verts = self.cloth_renderer.verts.to(self.device)
self.deform_verts.requires_grad = False
self.deform_graph = DeformationGraph(vert_number=verts_num)
self.mask_silhouette = None
self.w_photo = 100
self.w_tex = 10
self.w_perp = 2
self.w_ssim = 1
self.w_tvl = 1
_tmp_model1 = torchvision.models.resnext50_32x4d(pretrained=True).to(self.device)
self.feature_extractor_coarse = create_feature_extractor(_tmp_model1, {'layer1': 'feat1', 'layer4': 'feat2'})
_tmp_model2 = torchvision.models.vgg16(pretrained=True).to(self.device)
self.feature_extractor_fine = create_feature_extractor(_tmp_model2, {'features.12': 'feat1', 'features.30': 'feat2'})
self.binarization = Binarize.apply
self.face2edge = FaceToEdge(remove_faces=False)
def create_graph(self, garment_type, verts, faces, std_lst):
print("Start Graph Creation ...")
self.deform_graph.construct_graph(garment_type, verts.cpu(), faces.cpu())
num_nodes = self.deform_graph.nodes_idx.shape[0]
self.opt_d_rotations = torch.nn.Parameter(torch.zeros(1, num_nodes, 3), requires_grad=True) # axis angle
self.opt_d_translations = torch.nn.Parameter(torch.zeros(1, num_nodes, 3), requires_grad=True)
self.dg = DeformGraphModel(deform_graph=self.deform_graph,
renderer=self.cloth_renderer,
binarization=self.binarization,
canonical_mesh=self.canonical_mesh,
std_lst=std_lst,
lr_rate=1e-2,
savedir=self.savedir)
print("Finish Graph Creation !!!")
def iterative_mesh(self,
garment_type,
batch_id,
std_lst,
vertex_number,
inputs,
contours,
times=1001):
# Weight for mask & keypoint
w_mask = 10.0
w_kp = 0.005
# Weight for mesh edge loss
w_edge = 1.0
# Weight for mesh normal consistency
w_normal = 0.01
w_normal = 0.1
# Weight for mesh laplacian smoothing
w_laplacian = 0.1
loop = tqdm(range(401))
inputs_front, inputs_back = inputs[0].to(self.device).float(), inputs[1].to(self.device).float()
landmark_front, landmark_back= contours[0].to(self.device).float(), contours[1].to(self.device).float() # landmark (2023.02.15)
# A1 step1: Align top neckline edges.
for i in loop:
render_mask, specific_verts_2d = self.cloth_renderer.render_silhouette(self.deform_verts, side='back', landmark=True, vertex_number=vertex_number)
render_mask = render_mask[..., 3]
render_mask = torch.clip(render_mask * 2, 0, 1)
masks = inputs_back[0]
distance = specific_verts_2d[0][0].detach() - landmark_back.squeeze()[0] # 对齐领口左侧.
# image plane up&down
if abs(distance[1]) > 1:
if distance[1] > 0:
self.deform_verts = self.deform_verts + 0.003 * torch.tensor([0, 1, 0]).to(self.deform_verts.device)
else:
self.deform_verts = self.deform_verts - 0.003 * torch.tensor([0, 1, 0]).to(self.deform_verts.device)
if i % 200 == 0:
aaa = render_mask.unsqueeze(-1).detach().cpu().numpy() * 255.
bbb = masks[0].unsqueeze(-1).cpu().numpy() * 255.
ccc = inputs_front[0][0].unsqueeze(-1).cpu().numpy() * 255.
cv2.imwrite("experiments/{0}/iterative_{1}_step1_{2}.jpg".format(self.savedir, batch_id, i), cv2.hconcat([(aaa.astype(np.uint8)), bbb.astype(np.uint8), ccc.astype(np.uint8)]))
# A2 step2: deformation graph.
self.cloth_renderer.verts = self.deform_verts.detach()
_verts = self.deform_verts.detach()
_verts.requires_grad = False
_faces = self.cloth_renderer.faces
_faces.requires_grad = False
self.create_graph(garment_type, _verts, _faces, std_lst)
final_vertices, opt_d_rot, opt_d_trans = self.dg.iterative_deformgraph(
batch_id,
vertex_number,
inputs,
contours,
self.deform_verts.detach(),
self.opt_d_rotations,
self.opt_d_translations,
times=1001)
print("deformation graph finish!")
self.cloth_renderer.verts = final_vertices.detach()
def iterative_optimize(self,
garment_mask,
batch_id,
vertex_number,
inputs,
input_masks,
times=1001):
inputs_front, inputs_back = inputs[0].to(self.device).float(), inputs[1].to(self.device).float()
masks_front, masks_back = input_masks[0].to(self.device).float(), input_masks[1].to(self.device).float()
loop = tqdm(range(times))
for n in loop:
t1 = time.time()
self.optimizer.zero_grad()
#=======================Finetune the UV map=====================
rendered_imgs, rendered_masks = self.cloth_renderer.render_image(self.uv[0].unsqueeze(0))
#============================== Loss Definition ===========================================
outputs = rendered_imgs[:, :, :, :3].permute(0, 3, 1, 2).contiguous()
render_mask = rendered_masks[..., 3]
render_mask_out = self.binarization(render_mask)
# 1) Photometric Loss
l_photo = nn.L1Loss()(outputs[0].unsqueeze(0), inputs_front) + nn.L1Loss()(outputs[1].unsqueeze(0), inputs_back)
l_perp = 0.0
# 2) TVL Loss
l_tvl = self.tvl_loss(self.uv[0].unsqueeze(0).permute(0, 3, 1, 2))
loss = self.w_photo*l_photo + self.w_tvl*l_tvl
#=============================== Backward and optimization =======================================
loss.backward()
self.optimizer.step()
if n % 500 == 0:
with torch.no_grad():
rendered_imgs_all, rendered_masks_all = self.cloth_renderer.render_image(self.uv[0].unsqueeze(0))
rendered_imgs_all_finish = rendered_imgs_all[:, :, :, :3].permute(0, 3, 1, 2).contiguous()
mask_front_r, mask_back_r = rendered_masks_all[0, ..., 3], rendered_masks_all[1, ..., 3]
mask_front_r, mask_back_r = self.binarization(mask_front_r), self.binarization(mask_back_r)
rendered_imgs_front, rendered_imgs_back = rendered_imgs_all_finish[0].unsqueeze(0), rendered_imgs_all_finish[1].unsqueeze(0),
out_front = (np.clip(rendered_imgs_front.detach().cpu()[0].permute(1,2,0).numpy(), 0.0, 1.0) * 255.)
gaga_front = inputs_front[0].cpu().permute(1,2,0).numpy() * 255. # * mask_front_r.permute(1,2,0).cpu().numpy()
out_back = (np.clip(rendered_imgs_back.detach().cpu()[0].permute(1,2,0).numpy(), 0.0, 1.0) * 255.)
gaga_back = inputs_back[0].cpu().permute(1,2,0).numpy() * 255. # * mask_back_r.permute(1,2,0).cpu().numpy()
cv2.imwrite("experiments/{0}/{1}_texture_{2}.jpg".format(self.savedir, batch_id, n), cv2.hconcat([gaga_front.astype(np.uint8), out_front.astype(np.uint8), gaga_back.astype(np.uint8), out_back.astype(np.uint8)]))
out_uv = np.clip(self.uv[0].detach().cpu().numpy()[:, :, :], 0.0, 1.0)
out_uv = out_uv / out_uv.max()
out_uv = (out_uv * 255.)
if garment_mask is not None:
garment_mask_img = cv2.imread(garment_mask)
cv2.imwrite("experiments/{0}/{1}_texture_uv_{2}.jpg".format(self.savedir, batch_id, n), (out_uv * garment_mask_img / 255. + (255 - garment_mask_img)).astype(np.uint8))
loop.set_description("iter {0}, loss: l_photo {1:.3f}, l_tvl: {2:.3f}".format(n, self.w_photo*l_photo, self.w_tvl*l_tvl))
print('Finished Phase I Optimization')
torch.cuda.empty_cache()
def main(category,
savedir,
scale,
steps_one,
steps_two):
img_transform = transforms.ToTensor()
category = category
savedir = savedir
scale = scale
steps_one = steps_one
steps_two = steps_two
mapping_dict = {"1_wy": "wy.jpg",
"2_Polo": "polo.jpg",
"3_Tshirt": "Tshirt.jpg",
"4_shorts": "shorts.jpg",
"5_trousers": "trousers.jpg",
"6_zipup": "zipup.jpg",
"7_windcoat": "windcoat.jpg",
"9_jacket": "jacket.jpg",
"11_skirt": "skirt.jpg"}
garment_mask = "template/uv_mask/{}".format(mapping_dict.get(category))
landmark_order_dict = {"1_wy": ['ln', 'rn', 'lh', 'rh', 'lso', 'rso', 'lpo', 'lpi', 'rpo', 'rpi'],
"2_Polo": ['ln', 'rn', 'lh', 'rh', 'lso', 'rso', 'lpo', 'lpi', 'rpo', 'rpi'],
"3_Tshirt": ['ln', 'rn', 'lh', 'rh', 'lso', 'rso', 'lpo', 'lpi', 'rpo', 'rpi'],
"4_shorts": ['lw', 'mw', 'rw', 'cc', 'llo', 'lli', 'rli', 'rlo'],
"5_trousers": ['lw', 'mw', 'rw', 'cc', 'llo', 'lli', 'rli', 'rlo'],
"6_zipup": ['ln', 'rn', 'lh', 'rh', 'lso', 'rso', 'lpo', 'lpi', 'rpo', 'rpi'],
"7_windcoat": ['ln', 'rn', 'lh', 'rh', 'lso', 'rso', 'lpo', 'lpi', 'rpo', 'rpi'],
"9_jacket": ['ln', 'rn', 'lh', 'rh', 'lso', 'rso', 'lpo', 'lpi', 'rpo', 'rpi'],
"11_skirt": ['lw', 'mw', 'rw', 'll', 'ml', 'rl']}
per_vertex_dict = {"1_wy": 9648,
"2_Polo": 8922,
"3_Tshirt": 8523,
"4_shorts": 8767,
"5_trousers": 9323,
"6_zipup": 8537,
"7_windcoat": 9881,
"9_jacket": 8168,
"11_skirt": 6116}
# NOTICE: we do not release auto scaling code here and release a pre-defined scale coefficient parameters.
per_scale_dict = {"1_wy": 1.1,
"2_Polo": 0.8, # default 0.8
"3_Tshirt": 0.9, # default 0.7
"4_shorts": 0.75, # # default 0.7
"5_trousers": 0.75,
"6_zipup": 1.1,
"7_windcoat": 0.65,
"9_jacket": 1.0,
"11_skirt": 1.0}
vertex_number_dict = {"1_wy": [[9613, 9628, 9310, 9341, 5829, 5701, 8916, 8899, 8772, 8755, 9325, 9583],
[9572, 9558, 9004, 8974, 5829, 5701, 8916, 8899, 8772, 8755, 8990]],
"2_Polo": [[4318, 4332, 1031, 849, 8485, 91, 8812, 8794, 8871, 8858, 1050, 24],
[4321, 4330, 4710, 4845, 4759, 4796, 8812, 8794, 8871, 8858, 4691]],
"3_Tshirt": [[4427, 658, 4497, 4526, 7828, 8190, 40, 7799, 324, 333, 4511, 4415],
[842, 856, 762, 728, 7830, 8192, 7814, 7802, 8172, 8164, 744]],
"4_shorts": [[8215, 8198, 8179, 1772, 5092, 5000, 3398, 3490],
[7639, 7657, 7673, 5018, 5092, 5000, 3398, 3490]],
"5_trousers": [[8502, 9081, 8536, 6639, 4384, 4404, 2476, 2331],
[8502, 8518, 8536, 4454, 6538, 6685, 2476, 2331]],
"6_zipup": [[3325, 3210, 8305, 8335, 65, 15, 8019, 8012, 8494, 8487, 8319],
[3325, 3211, 8305, 8335, 65, 15, 8019, 8012, 8494, 8487, 8076]],
"7_windcoat": [[2378, 2405, 7177, 5531, 8806, 9411, 8835, 8771, 9380, 9372, 5514],
[2378, 2405, 7177, 5532, 8806, 9411, 8835, 8771, 9380, 9372, 2282]],
"9_jacket": [[7488, 8022, 7892, 7380, 669, 31, 8120, 8136, 7840, 7856, 5506, 8077],
[7488, 7498, 7892, 7380, 2514, 31, 5436, 5414, 5347, 5325, 7587]],
"11_skirt": [[216, 206, 193, 3301, 3348, 507],
[216, 45, 193, 3301, 556, 507]]}
for idx in range(0,2):
trainer = Trainer(objfile="template/reference/{}/mesh/mesh.obj".format(category),
savedir=savedir,
resolution=512,
focal_distance=1.7,
verts_num=per_vertex_dict[category],
scale_factor=scale)
if os.path.exists("template/reference/{0}/square/{1}_1.jpg".format(category, idx)) is False:
continue
ref_img = cv2.imread("template/reference/{0}/bg/{1}_1.jpg".format(category, idx))
ref_img_back = cv2.imread("template/reference/{0}/bg/{1}_2.jpg".format(category, idx))
H, W = ref_img.shape[:2]
H1, W1 = ref_img_back.shape[:2]
with open("template/reference/{0}/square/{1}_1.json".format(category, idx)) as f:
result_kp1=json.load(f)
with open("template/reference/{0}/square/{1}_2.json".format(category, idx)) as f:
result_kp2=json.load(f)
std_lst = landmark_order_dict[category].copy()
tmp_count_f, tmp_count_b = len(result_kp1['shapes']), len(result_kp2['shapes'])
c_src_front = [result_kp1['shapes'][_]['points'][:2] for _ in range(tmp_count_f)]
c_src_front = np.array(c_src_front)
c_src_front_label = [result_kp1['shapes'][_]['label'] for _ in range(tmp_count_f)]
c_src_back = [result_kp2['shapes'][_]['points'][:2] for _ in range(tmp_count_b)]
c_src_back = np.array(c_src_back)
c_src_back_label = [result_kp2['shapes'][_]['label'] for _ in range(tmp_count_b)]
c_src_idx, c_dst_idx = [], []
print("#"*30)
print(idx, std_lst)
print("#"*30)
for tmp_idx in std_lst:
c_src_idx.append(c_src_front_label.index(tmp_idx))
c_dst_idx.append(c_src_back_label.index(tmp_idx))
c_src_front = c_src_front[c_src_idx]
c_src_back = c_src_back[c_dst_idx]
# add mh point.
if "cc" not in std_lst and "skirt" not in category:
std_lst.append("mh")
tmp_mh = np.expand_dims((c_src_front[2] + c_src_front[3]) / 2, 0)
c_src_front = np.append(c_src_front, tmp_mh, axis=0)
tmp_mh = np.expand_dims((c_src_back[2] + c_src_back[3]) / 2, 0)
c_src_back = np.append(c_src_back, tmp_mh, axis=0)
# wy, polo, T needs mn
std_lst_front = std_lst.copy()
std_lst_back = std_lst.copy()
if category in ["1_wy", "2_Polo", "3_Tshirt", "9_jacket"]:
for mn_idx in range(tmp_count_f):
if result_kp1['shapes'][mn_idx]['label'] == "mn" or result_kp1['shapes'][mn_idx]['label'] == "lmn":
tmp_mn = np.expand_dims(np.asarray(result_kp1['shapes'][mn_idx]['points'][:2]), 0)
c_src_front = np.append(c_src_front, tmp_mn, axis=0)
std_lst_front.append("mn")
print("[mn] appending success!")
break
inputs = img_transform(ref_img).unsqueeze(0)
inputs_back = img_transform(ref_img_back).unsqueeze(0)
c_src_front, c_src_back = torch.from_numpy(c_src_front).unsqueeze(0), torch.from_numpy(c_src_back).unsqueeze(0)
mask_front = cv2.imread("template/reference/{0}/mask/{1}_1.jpg".format(category, idx), 0)
mask_back = cv2.imread("template/reference/{0}/mask/{1}_2.jpg".format(category, idx), 0)
mask_front, mask_back = np.where(mask_front > 10, 255, 0), np.where(mask_back > 10, 255, 0)
mask_front, mask_back = img_transform(mask_front).unsqueeze(0), img_transform(mask_back).unsqueeze(0)
mask_front, mask_back = mask_front / 255., mask_back / 255.
# optimize
trainer.iterative_mesh(category,
idx,
[std_lst_front, std_lst_back],
vertex_number_dict[category],
[mask_front, mask_back],
[c_src_front, c_src_back],
times=steps_one)
trainer.iterative_optimize(garment_mask,
idx,
vertex_number_dict[category],
[inputs, inputs_back],
[mask_front, mask_back],
times=steps_two)
break
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--g', '--garment', help = 'garment type: 1_wy, 2_Polo, 3_Tshirt, 4_shorts, 5_trousers, 6_zipup, 7_windcoat, 9_jacket, 11_skirt', type = str, default = "1_wy")
parser.add_argument('--d', '--dstdir', help = 'dst dir', type = str, default = "{}".format(datetime.date.today().strftime("%Y-%m-%d")))
parser.add_argument('--s', '--scale', help = 'scale coefficient', type = float, default = 1.0)
parser.add_argument('--steps_one', help = 'optimizer 1 steps', type = int, default = 501)
parser.add_argument('--steps_two', help = 'optimizer 2 steps', type = int, default = 1001)
args = parser.parse_args()
print(args)
main(args.g, args.d, args.s, args.steps_one, args.steps_two)