-
Notifications
You must be signed in to change notification settings - Fork 2
/
trainset.py
executable file
·275 lines (256 loc) · 11.7 KB
/
trainset.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
import torch
import numpy as np
import glob, random
import open3d as o3d
from tqdm import tqdm
from utils.utils import *
from utils.r_eval import *
from knn_search import knn_module
from backbone_fcgf import load_model
from utils.misc import extract_features
import multiprocessing as mp
class generate_trainset:
def __init__(self):
self.trainseq = [0,1,2,3,4,5]
self.valseq = [6,7]
self.basedir = f'./data/origin_data/kitti_train'
self.feat_train_dir = f'./data/cache/train'
make_non_exists_dir(self.feat_train_dir)
self.load_model()
self.G = np.load(f'./group_related/Rotation_8.npy')
self.knn = knn_module.KNN(1)
self.batchsize = 64
def loadset(self):
self.train = {}
for i in range(8):
seq = {
'pc':[],
'pair':{}
}
pair_fns = glob.glob(f'{self.basedir}/icp/{i}_*') #for 555
for fn in pair_fns:
trans = np.load(fn)
pair = str.split(fn,'/')[-1][:-4]
pair = str.split(pair,'_')
assert int(pair[0]) == i
seq['pair'][f'{pair[1]}-{pair[2]}'] = trans
if not pair[1] in seq['pc']:
seq['pc'].append(pair[1])
if not pair[2] in seq['pc']:
seq['pc'].append(pair[2])
self.train[f'{i}'] = seq
def gt_match(self):
for seqs in [self.trainseq, self.valseq]:
for i in seqs:
seq = self.train[f'{i}']
savedir = f'{self.feat_train_dir}/{i}/gt_match'
make_non_exists_dir(savedir)
for pair,trans in tqdm(seq['pair'].items()):
id0,id1=str.split(pair,'-')
pc0 = o3d.io.read_point_cloud(f'{self.basedir}/{i}/PointCloud/cloud_bin_{id0}.ply')
pc1 = o3d.io.read_point_cloud(f'{self.basedir}/{i}/PointCloud/cloud_bin_{id1}.ply')
pc0 = np.array(pc0.points)
pc1 = np.array(pc1.points)
key0 = np.loadtxt(f'{self.basedir}/{i}/Keypoints/cloud_bin_{id0}Keypoints.txt').astype(np.int)
key1 = np.loadtxt(f'{self.basedir}/{i}/Keypoints/cloud_bin_{id1}Keypoints.txt').astype(np.int)
key0 = pc0[key0]
key1 = pc1[key1]
key0 = apply_transform(key0, trans) #align
# pair with the filtered keypoints: index in keys
dist = np.sum(np.square(key0[:,None,:]-key1[None,:,:]),axis=-1)
# match
thres = 0.3*1.5
d_min = np.min(dist,axis=1)
arg_min = np.argmin(dist,axis=1)
m0 = np.arange(d_min.shape[0])[d_min<thres*thres]
m1 = arg_min[d_min<thres*thres]
pair = np.concatenate([m0[:,None],m1[:,None]],axis=1) #pairnum*2
save_fn = f'{savedir}/{id0}_{id1}.npy'
np.save(save_fn, pair)
def load_model(self):
checkpoint = torch.load('./model/Backbone/best_val_checkpoint.pth')
config = checkpoint['config']
Model = load_model(config.model)
num_feats = 1
self.model = Model(
num_feats,
config.model_n_out,
bn_momentum=0.05,
normalize_feature=config.normalize_feature,
conv1_kernel_size=config.conv1_kernel_size,
D=3)
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.model.to(self.device)
self.model.load_state_dict(checkpoint['state_dict'])
self.model.eval()
def generate_scan_gfeats(self,pc,key):
feats = []
if pc.shape[0]>40000:
index = np.arange(pc.shape[0])
np.random.shuffle(index)
pc = pc[index[0:40000]]
for gid in range(self.G.shape[0]):
feats_g = []
g = self.G[gid]
#rot the point cloud
pc_g = [email protected]
key_g = [email protected]
with torch.no_grad():
pc_g_down, feature_g = extract_features(
self.model,
xyz=pc_g,
voxel_size=0.3,
device=self.device,
skip_check=True)
feature_g=feature_g.cpu().numpy()
xyz_down_pcd = o3d.geometry.PointCloud()
xyz_down_pcd.points = o3d.utility.Vector3dVector(pc_g_down)
pcd_tree = o3d.geometry.KDTreeFlann(xyz_down_pcd)
for k in range(key_g.shape[0]):
[_, idx, _] = pcd_tree.search_knn_vector_3d(key_g[k], 1)
feats_g.append(feature_g[idx[0]][None,:])
feats_g=np.concatenate(feats_g,axis=0)#kn*32
feats.append(feats_g[:,:,None])
feats = np.concatenate(feats, axis=-1)#kn*32*8
return feats
def R2DR_id(self,R):
min_diff=180
best_id=0
for R_id in range(self.G.shape[0]):
R_diff=compute_R_diff(self.G[R_id],R)
if R_diff<min_diff:
min_diff=R_diff
best_id=R_id
return best_id
def DeltaR(self,R,index):
R_anchor=self.G[index]#3*3
#R=Rres@Ranc->[email protected]
deltaR=R@R_anchor.T
return quaternion_from_matrix(deltaR)
def generate_batches(self, start = 0):
batchsavedir = f'{self.feat_train_dir}/train_val_batch/trainset'
make_non_exists_dir(batchsavedir)
batch_i = start
for i in self.trainseq:
seq = self.train[f'{i}']
for pair, trans in tqdm(seq['pair'].items()):
id0,id1=str.split(pair,'-')
pc0 = np.load(f'{self.basedir}/{i}/PointCloud/cloud_bin_{id0}.npy')
pc1 = np.load(f'{self.basedir}/{i}/PointCloud/cloud_bin_{id1}.npy')
key_idx0 = np.loadtxt(f'{self.basedir}/{i}/Keypoints/cloud_bin_{id0}Keypoints.txt').astype(np.int64)
key_idx1 = np.loadtxt(f'{self.basedir}/{i}/Keypoints/cloud_bin_{id1}Keypoints.txt').astype(np.int64)
key0 = pc0[key_idx0]
key1 = pc1[key_idx1]
R_base = random_rotation_zgroup()
# gt alignment
pc0 = apply_transform(pc0, trans)
key0 = apply_transform(key0, trans)
# 1.random z rotation to pc0&pc1 2.group rot to pc1 3.residual rot to pc1
R_z = random_z_rotation(180)
R_45 = random_z_rotation(45)
pc0 = pc0@R_z.T
pc1 = ((pc1@R_z.T)@R_base.T)@R_45.T
key0 = key0@R_z.T
key1 = ((key1@R_z.T)@R_base.T)@R_45.T
# added rot
R = R_45@R_base
R_index = self.R2DR_id(R)
R_residual = self.DeltaR(R,R_index)
batch_Rs, batch_Rids, batch_Rres = [],[],[]
for b in range(self.batchsize):
batch_Rs.append(R[None,:,:])
batch_Rids.append(R_index)
batch_Rres.append(R_residual[None,:])
batch_Rs = np.concatenate(batch_Rs, axis=0)
batch_Rids = np.array(batch_Rids)
batch_Rres = np.concatenate(batch_Rres, axis=0)
#gennerate rot feats
feats0 = self.generate_scan_gfeats(pc0, key0) #5000*32*8
feats1 = self.generate_scan_gfeats(pc1, key1)
pt_pair = np.load(f'{self.feat_train_dir}/{i}/gt_match/{id0}_{id1}.npy').astype(np.int32)
index = np.arange(pt_pair.shape[0])
np.random.shuffle(index)
index = index[0:self.batchsize]
pt_pair = pt_pair[index]
# paired feats
feats0 = feats0[pt_pair[:,0],:,:] #64*32*8
feats1 = feats1[pt_pair[:,1],:,:]
item={
'feats0':torch.from_numpy(feats0.astype(np.float32)), #before enhanced rot
'feats1':torch.from_numpy(feats1.astype(np.float32)), #after enhanced rot
'R':torch.from_numpy(batch_Rs.astype(np.float32)),
'true_idx':torch.from_numpy(batch_Rids.astype(np.int)),
'deltaR':torch.from_numpy(batch_Rres.astype(np.float32))
}
# save
torch.save(item,f'{batchsavedir}/{batch_i}.pth',_use_new_zipfile_serialization=False)
batch_i += 1
def generate_val_batches(self, vallen = 3000):
batchsavedir = f'{self.feat_train_dir}/train_val_batch/valset'
make_non_exists_dir(batchsavedir)
# generate matches
matches = []
for i in self.valseq:
seq = self.train[f'{i}']
for pair, trans in tqdm(seq['pair'].items()):
id0,id1=str.split(pair,'-')
pair = np.load(f'{self.feat_train_dir}/{i}/gt_match/{id0}_{id1}.npy').astype(np.int32)
for p in range(pair.shape[0]):
matches.append((i,id0,id1,pair[p][0],pair[p][1],trans))
random.shuffle(matches)
batch_i=0
for batch_i in tqdm(range(vallen)):
tup = matches[batch_i]
scene, id0, id1, pt0, pt1, trans = tup
pc0 = np.load(f'{self.basedir}/{scene}/PointCloud/cloud_bin_{id0}.npy')
pc1 = np.load(f'{self.basedir}/{scene}/PointCloud/cloud_bin_{id1}.npy')
key_idx0 = np.loadtxt(f'{self.basedir}/{scene}/Keypoints/cloud_bin_{id0}Keypoints.txt').astype(np.int64)
key_idx1 = np.loadtxt(f'{self.basedir}/{scene}/Keypoints/cloud_bin_{id1}Keypoints.txt').astype(np.int64)
key0 = pc0[key_idx0]
key1 = pc1[key_idx1]
R_base = random_rotation_zgroup()
# gt alignment
pc0 = apply_transform(pc0, trans) #align
key0 = apply_transform(key0, trans)
# 1.random z rotation to pc0&pc1 2.group rot to pc1 3.residual rot to pc1
R_z = random_z_rotation(180)
R_45 = random_z_rotation(45)
pc0 = pc0@R_z.T
pc1 = ((pc1@R_z.T)@R_base.T)@R_45.T
key0 = key0@R_z.T
key1 = ((key1@R_z.T)@R_base.T)@R_45.T
# added rot
R = R_45@R_base
R_index = self.R2DR_id(R)
R_residual = self.DeltaR(R,R_index)
#gennerate rot feats
feats0 = self.generate_scan_gfeats(pc0, key0) #5000*32*8
feats1 = self.generate_scan_gfeats(pc1, key1)
feats0 = feats0[int(pt0)] #32*8
feats1 = feats1[int(pt1)]
# joint to be a batch
item={
'feats0':torch.from_numpy(feats0.astype(np.float32)), #before enhanced rot
'feats1':torch.from_numpy(feats1.astype(np.float32)), #after enhanced rot
'R':torch.from_numpy(R.astype(np.float32)),
'true_idx':torch.from_numpy(np.array([R_index]).astype(np.int)),
'deltaR':torch.from_numpy(R_residual.astype(np.float32))
}
# save
torch.save(item,f'{batchsavedir}/{batch_i}.pth',_use_new_zipfile_serialization=False)
batch_i += 1
def trainval_list(self):
traindir = f'{self.feat_train_dir}/train_val_batch/trainset'
valdir = f'{self.feat_train_dir}/train_val_batch/valset'
trainlist = glob.glob(f'{traindir}/*.pth')
vallist = glob.glob(f'{valdir}/*.pth')
save_pickle(range(len(trainlist)), f'{self.feat_train_dir}/train_val_batch/train.pkl')
save_pickle(range(len(vallist)), f'{self.feat_train_dir}/train_val_batch/val.pkl')
if __name__=='__main__':
generator = generate_trainset()
generator.loadset()
generator.gt_match()
for i in range(3):
generator.generate_batches(start = len(glob.glob(f'{generator.feat_train_dir}/train_val_batch/trainset/*.pth')))
generator.generate_val_batches()
generator.trainval_list()