-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
502 lines (440 loc) · 19.6 KB
/
model.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
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
import numpy as np
import torch
from torch import nn
from torch.nn import functional as F
from util import *
from torch.optim import Adam,SGD
from random import choice,sample
from torch.utils.data.sampler import BatchSampler, SubsetRandomSampler
from util import writeSummary
from torchvision.models import resnet18
from copy import deepcopy
try:
from torch.utils.tensorboard import SummaryWriter
except:
pass
def Quat2Rotation(x,y,z,w):
l1 = torch.stack([1 - 2 * y ** 2 - 2 * z ** 2, 2 * x * y + 2 * w * z, 2 * x * z - 2 * w * y],dim=0)
l2 = torch.stack([2 * x * y - 2 * w * z, 1 - 2 * x ** 2 - 2 * z ** 2, 2 * y * z + 2 * w * x],dim=0)
l3 = torch.stack([2 * x * z + 2 * w * y, 2 * y * z - 2 * w * x, 1 - 2 * x ** 2 - 2 * y ** 2], dim=0)
T_w = torch.stack([l1,l2,l3],dim=0)
return T_w
def Rotation2Quat(pose):
m11,m22,m33 = pose[0][0],pose[1][1],pose[2][2]
m12,m13,m21,m23,m31,m32 = pose[0][1],pose[0][2],pose[1][0],pose[1][2],pose[2][0],pose[2][1]
x,y,z,w = torch.sqrt(m11-m22-m33+1)/2,torch.sqrt(-m11+m22-m33+1)/2,torch.sqrt(-m11-m22+m33+1)/2,torch.sqrt(m11+m22+m33+1)/2
Quat_ = torch.tensor([
[x,(m12+m21)/(4*x),(m13+m31)/(4*x),(m23-m32)/(4*x)],
[(m12+m21)/(4*y),y,(m23+m32)/(4*y),(m31-m13)/(4*y)],
[(m13 + m31) / (4 * z), (m23 + m32) / (4 * z), z,(m12 - m21) / (4 * z)],
[(m23 - m32) / (4 * w), (m31 - m13) / (4 * w), (m12 - m21) / (4 * w),w]
], dtype=torch.float32)
_,index = torch.tensor([x,y,z,w]).max(dim=0)
Quat = Quat_[index.item()]
return Quat
def axis_angle_to_quaternion(axis_angle):
"""
Convert rotations given as axis/angle to quaternions.
Args:
axis_angle: Rotations given as a vector in axis angle form,
as a tensor of shape (..., 3), where the magnitude is
the angle turned anticlockwise in radians around the
vector's direction.
Returns:
quaternions with real part first, as tensor of shape (..., 4).
"""
angles = torch.norm(axis_angle, p=2, dim=-1, keepdim=True)
half_angles = 0.5 * angles
eps = 1e-6
small_angles = angles.abs() < eps
sin_half_angles_over_angles = torch.empty_like(angles)
sin_half_angles_over_angles[~small_angles] = (
torch.sin(half_angles[~small_angles]) / angles[~small_angles]
)
# for x small, sin(x/2) is about x/2 - (x/2)^3/6
# so sin(x/2)/x is about 1/2 - (x*x)/48
sin_half_angles_over_angles[small_angles] = (
0.5 - (angles[small_angles] * angles[small_angles]) / 48
)
quaternions = torch.cat(
[torch.cos(half_angles), axis_angle * sin_half_angles_over_angles], dim=-1
)
return quaternions
def axis_angle_to_matrix(axis_angle):
"""
Convert rotations given as axis/angle to rotation matrices.
Args:
axis_angle: Rotations given as a vector in axis angle form,
as a tensor of shape (..., 3), where the magnitude is
the angle turned anticlockwise in radians around the
vector's direction.
Returns:
Rotation matrices as tensor of shape (..., 3, 3).
"""
return quaternion_to_matrix(axis_angle_to_quaternion(axis_angle))
def quaternion_to_matrix(quaternions):
"""
Convert rotations given as quaternions to rotation matrices.
Args:
quaternions: quaternions with real part first,
as tensor of shape (..., 4).
Returns:
Rotation matrices as tensor of shape (..., 3, 3).
"""
r, i, j, k = torch.unbind(quaternions, -1)
two_s = 2.0 / (quaternions * quaternions).sum(-1)
o = torch.stack(
(
1 - two_s * (j * j + k * k),
two_s * (i * j - k * r),
two_s * (i * k + j * r),
two_s * (i * j + k * r),
1 - two_s * (i * i + k * k),
two_s * (j * k - i * r),
two_s * (i * k - j * r),
two_s * (j * k + i * r),
1 - two_s * (i * i + j * j),
),
-1,
)
return o.reshape(quaternions.shape[:-1] + (3, 3))
def iden(x):
return x
class NeRF_pi(nn.Module):
def __init__(self, input_dim=3, W=64, pos_multires=10, dir_multires=4):
super(NeRF_pi,self).__init__()
self.input_dim = input_dim * (pos_multires * 2 + 1)
self.input_dir_dim = input_dim * (dir_multires * 2 + 1)
self.pos_freq_bands = (2. ** torch.linspace(0., pos_multires - 1, steps=pos_multires)) * torch.pi
self.dir_freq_bands = (2. ** torch.linspace(0., dir_multires - 1, steps=dir_multires)) * torch.pi
self.W = W
self.part1 = nn.Sequential(
nn.Linear(self.input_dim, W),
nn.ReLU(),
nn.Linear(W, W),
nn.ReLU()
)
self.part2 = nn.Sequential(
nn.Linear(self.input_dim+W, W),
nn.ReLU(),
nn.Linear(W, W),
nn.ReLU(),
)
self.part3 = nn.Sequential(
nn.Linear(W + self.input_dir_dim, W),
nn.ReLU(),
)
self.alpha_linear = nn.Sequential(
nn.Linear(W, 1),
nn.ReLU(),
)
self.rgb_linear = nn.Sequential(
nn.Linear(W, 3),
nn.ReLU()
)
def get_embed(self,x, freq_bands):
with torch.no_grad():
x_ = torch.cat([x*freq for freq in freq_bands], -1)
x_ = torch.cat([fn(x_) for fn in [torch.sin, torch.cos]], -1).float()
return torch.cat([x, x_], -1).float()
def forward(self,pts, viewdirs):
N_ray, N_sample, _ = pts.shape
viewdirs = viewdirs[:, None].expand(pts.shape).reshape(-1, 3)
pts = pts.view(-1, 3)
gamma = self.get_embed(pts, self.pos_freq_bands)
dirs = self.get_embed(viewdirs, self.dir_freq_bands)
out = self.part1(gamma)
out = torch.cat([out, gamma], -1)
out1 = self.part2(out)
alpha = self.alpha_linear(out1).view(N_ray, N_sample, -1)
out2 = self.part3(torch.cat([out1, dirs], -1)).view(N_ray, N_sample, -1)
rgb = self.rgb_linear(out2).view(N_ray, N_sample, -1)
return torch.cat([alpha, rgb], dim=2), out2
class Prd_Net(nn.Module):
def __init__(self):
super(Prd_Net, self).__init__()
self.atten1 = nn.MultiheadAttention(embed_dim=64*2, num_heads=2, dropout=0.0, batch_first=True)
self.feed1=nn.Sequential(
nn.Linear(64*2,64),
nn.LeakyReLU(0.1,True),
nn.Linear(64, 64),
nn.LeakyReLU(0.1, True),
)
self.layer_norm1 = nn.LayerNorm([1200,64])
self.atten2 = nn.MultiheadAttention(embed_dim=64*2, num_heads=2, dropout=0.0, batch_first=True)
self.feed2 = nn.Sequential(
nn.Linear(64*2, 64),
nn.LeakyReLU(0.1, True),
nn.Linear(64, 64),
nn.LeakyReLU(0.1, True),
)
self.layer_norm2 = nn.LayerNorm([1200, 64])
self.middle = nn.Sequential(
nn.Linear(64, 64),
nn.LeakyReLU(0.1, True)
)
self.fc = nn.Sequential(
nn.Linear(64 , 64),
nn.LeakyReLU(0.1, True),
nn.Linear(64, 32),
nn.LeakyReLU(0.1, True),
nn.Linear(32, 1)
)
def forward(self, prd_map):
x = F.avg_pool2d(prd_map, (3, 3), 3, 0).view(prd_map.shape[0], prd_map.shape[1], -1).transpose(1, 2) # [batch, 1200, 64]
x_ = torch.cat([x,x],-1)
aten1 = self.atten1(x_,x_, x_, need_weights=False)[0]
out1 = self.feed1((aten1+x_).reshape(-1,aten1.shape[-1])).reshape(aten1.shape[0],aten1.shape[1],-1)
out1 = self.layer_norm1(out1)
out1_ = torch.cat([out1, out1],-1)
aten2 = self.atten2(out1_, out1_, out1_, need_weights=False)[0] # [batch, 1200, 64]
out2 = self.feed2((aten2+out1_).reshape(-1, aten2.shape[-1])).reshape(aten2.shape[0], aten2.shape[1], -1)
out2 = self.layer_norm2(out2)
midd = self.middle(out2.mean(dim=1, keepdim=False))
pred = self.fc(midd)
return midd, pred
class bypath(nn.Module):
def __init__(self):
super(bypath, self).__init__()
self.resnet = resnet18(pretrained=False)
self.resnet.conv1 = nn.Conv2d(4, 64, (7, 7), (2, 2), (3, 3), bias=False)
self.linear = nn.Sequential(
nn.Linear(1000, 256),
nn.LeakyReLU(0.1,True),
)
def forward(self,x):
out = self.resnet(x)
return self.linear(out)
class E2E_model(nn.Module):
def __init__(self, action_space):
super(E2E_model,self).__init__()
self.model_name = 'E2E_model'
self.bypath = bypath()
self.linear_process = nn.Sequential(
nn.Linear(67, 64),
nn.LeakyReLU(0.1, True),
nn.Linear(64, 64),
nn.LeakyReLU(0.1, True),
)
self.pred_net = Prd_Net()
self.policy_net = nn.Sequential(
nn.Linear(256+64,256),
nn.LeakyReLU(0.1, True),
nn.Linear(256, 64),
nn.LeakyReLU(0.1, True),
nn.Linear(64, 32),
nn.LeakyReLU(0.1, True),
nn.Linear(32, action_space)
)
def forward(self, x, type = 'gathering', out_pred=None): # x:[b, 1, 180, 240]
if type == 'gathering':
a,b,c,d = out_pred.shape
pred_in = self.linear_process(out_pred.reshape(-1,out_pred.shape[-1])).view(a,b,c,-1)
fc,pred = self.pred_net(pred_in.permute([0,3,1,2]))
out_bypath = self.bypath(x)
pi = self.policy_net(torch.cat([out_bypath, fc], dim=1))
return F.softmax(pi, dim=1)
else:
a, b, c, d = out_pred.shape
pred_in = self.linear_process(out_pred.reshape(-1, out_pred.shape[-1])).view(a, b, c, -1)
fc, pred = self.pred_net(pred_in.permute([0, 3, 1, 2]))
out_bypath = self.bypath(x)
pi = self.policy_net(torch.cat([out_bypath, fc], dim=1))
return F.softmax(pi, dim=1), pred
class NeRF_proc():
def __init__(self, nerf_tmp, device, nerf_list, Camera_Intrinc='cameras.txt',N_sample=192):
super(NeRF_proc, self).__init__()
self.nerf = nerf_tmp
self.feature_t = None
self.N_sample = N_sample
self.half_dist = 1.0
self.jitter = True
self.device = device
self.nerf = self.nerf.to(self.device)
self.nerf_list = nerf_list
with open(Camera_Intrinc, 'r') as f:
K = f.readline()
K = K.split(' ')
self.H, self.W = int(K[-2]), int(K[-1])
self.K = np.array([
[float(K[0]), 0, float(K[2])],
[0, float(K[1]), float(K[3])],
[0, 0, 1]
], dtype=np.float32)
def change_device(self,device):
self.device = device
self.nerf = self.nerf.to(self.device)
def change_target(self,rgb_t):
with torch.no_grad():
self.feature_t = rgb_t[0:-1:2, 0:-1:2, :]
def memory_process(self, drgb, pose, lock, queue, step, nerf_batch=43200, other_device=None):
depth, rgb = drgb[:, 0:1], drgb[0, 1:, 0:-1:2, 0:-1:2].transpose(0, 1).transpose(1, 2) # depth [1,1, H,W], rgb [1,H,W,3]
depths, z_vals = generate_z_vals_and_depths(depth, self.N_sample, self.half_dist, self.jitter) #[H*W,1] [H*W,N_sample, 1]
pts, viewdirs = generate_rays_half(pose, self.H, self.W, self.K, z_vals) # pts: [N_rays, N_sample, 3],
if not len(self.nerf_list) == 0:
lock.acquire()
_state = self.nerf_list[-1]
self.nerf.load_state_dict(_state)
self.nerf = self.nerf.to(self.device)
self.nerf_list[:] = []
lock.release()
raw, out2 = minibatch(nerf_batch, self.nerf, pts, viewdirs)
pred = render_pred(raw, out2, z_vals, self.H // 2, self.W //2, is_flat=False)
if step % 3 == 0:
if other_device==None:
queue.put([pts, viewdirs, rgb.reshape(-1,3), depths, z_vals])
else:
queue.put([pts.to(other_device), viewdirs.to(other_device), rgb.reshape(-1,3).to(other_device), depths.to(other_device), z_vals.to(other_device)])
return torch.cat([pred, self.feature_t], dim=-1).unsqueeze(0)
def nerf_reset(nerf, lock, nerf_list):
for layer in nerf.modules():
if isinstance(layer, nn.Linear):
torch.nn.init.xavier_uniform_(layer.weight)
layer.bias.data.fill_(0)
lock.acquire()
nerf = nerf.cpu()
if not len(nerf_list) == 0:
nerf_list[:] = []
nerf_list.append(deepcopy(nerf.state_dict()))
lock.release()
def nerf_train(nerf, N_sample,device, lock, queue, nerf_list, reset_list, child_conn):
nerf = nerf.to(device)
nerf.train()
optimizer = Adam(nerf.parameters(), lr=0.001)
data_cache = [torch.zeros(0,N_sample,3),torch.zeros(0,3),
torch.zeros(0,3),torch.zeros(0,1),torch.zeros(0,N_sample,1),0] # pts, viewdirs, rgb, depths, z_vals, rays_total_nums
nerf_batch = 10800
count = 0
last_undate=0
H=90
W=120
while True:
if not queue.empty():
if data_cache[-1] == 0:
data_cache[0], data_cache[1],data_cache[2], data_cache[3], data_cache[4] = queue.get()
data_cache[-1] = data_cache[0].shape[0]
else:
pts, viewdirs, rgb, depths, z_vals = queue.get()
data_cache[0] = torch.cat([data_cache[0], pts],0) # N, N_sample, 3
data_cache[1] = torch.cat([data_cache[1], viewdirs], 0) # N, 3
data_cache[2] = torch.cat([data_cache[2], rgb], 0) # N, 3
data_cache[3] = torch.cat([data_cache[3], depths], 0) # N, 1
data_cache[4] = torch.cat([data_cache[4], z_vals], 0) # N, N_sample, 1
data_cache[-1] += pts.shape[0]
index = sample(list(range(H*W)), nerf_batch)
raw, out2 = minibatch(nerf_batch, nerf, data_cache[0][-H*W:][index], data_cache[1][-H*W:][index])
rgb_map, depth_map = render(raw, data_cache[4][-H*W:][index], H, W, is_flat=False)
loss_t, psnr = loss(rgb_map, depth_map, data_cache[2][-H*W:][index], data_cache[3][-H*W:][index], True, True)
# print(loss_t.cpu().item(),psnr)
optimizer.zero_grad()
loss_t.backward()
optimizer.step()
count += 1
if count-last_undate > 4:
lock.acquire()
nerf = nerf.cpu()
if not len(nerf_list) == 0:
nerf_list[:] = []
nerf_list.append(deepcopy(nerf.state_dict()))
lock.release()
nerf = nerf.to(device)
last_undate = count
if not data_cache[-1] == 0:
index = sample(list(range(data_cache[-1])), 1024)
raw, out2 = minibatch(1024, nerf, data_cache[0][index], data_cache[1][index])
rgb_map, depth_map = render(raw, data_cache[4][index], is_flat=True)
loss_t, psnr = loss(rgb_map, depth_map, data_cache[2][index], data_cache[3][index], False, False)
optimizer.zero_grad()
loss_t.backward()
optimizer.step()
count += 1
if data_cache[-1] > 10800*190:
data_cache[0] = data_cache[0][10800 * 30:]
data_cache[1] = data_cache[1][10800 * 30:]
data_cache[2] = data_cache[2][10800 * 30:]
data_cache[3] = data_cache[3][10800 * 30:]
data_cache[4] = data_cache[4][10800 * 30:]
torch.cuda.empty_cache()
data_cache[-1] = data_cache[0].shape[0]
if reset_list[-1]:
st = 'total train num:%d'%count
count =0
last_undate=0
del data_cache
data_cache = [torch.zeros(0, N_sample, 3), torch.zeros(0, 3),
torch.zeros(0, 3), torch.zeros(0, 1), torch.zeros(0, N_sample, 1),
0] # pts, viewdirs, rgb, depths, z_vals, rays_total_nums
while not queue.empty(): ### 清空queue
queue.get()
reset_list[-1] = False
# torch.cuda.empty_cache()
nerf_reset(nerf, lock, nerf_list)
nerf = nerf.to(device)
child_conn.send('reset '+st)
def model_train(model, device, lock, source, summary_path, model_path, init_lr,batch_Size,_flag):
model = model.to(device)
model.train()
lr=init_lr = init_lr
optimizer = Adam(model.parameters(), lr=init_lr)
writer = SummaryWriter(summary_path)
flag = False
num = 0
batchSize = batch_Size
total_p, total_t = 0, 0
_count = 0
episode = 0
stats = {'policy_loss':[], 'pred_loss':[], 'learning_rate':[]}
print(len(source) > 0)
while True:
if len(source) > 0:
lock.acquire()
_action = np.array([tt.action for tt in source], dtype=np.int64)
indexl = np.where(_action == 1)[0]
indexr = np.where(_action == 2)[0]
indexf = np.where(_action == 0)[0]
index_l_r = np.concatenate([indexl,indexr],0)
mean = ((len(indexl) + len(indexr)) * 2) // 5
index_l_r_ = np.random.choice(index_l_r, mean)
index_ = np.concatenate([indexl, indexr,indexf,index_l_r_], 0)
action = []
state = []
out_pred = []
label = []
for _i in index_:
action.append(source[_i].action)
state.append(source[_i].state)
out_pred.append(source[_i].prd_map)
label.append(source[_i].label)
action = torch.from_numpy(np.array(action, dtype=np.int64)).view(-1,1).to(device)
state = torch.from_numpy(np.concatenate(state, 0)).float().to(device)
out_pred = torch.cat(out_pred, dim=0).to(device)
label = torch.from_numpy(np.stack(label, 0)).to(device)
flag = True
num = label.shape[0]
print('length:', num)
source[:] = []
lock.release()
torch.cuda.empty_cache()
if flag:
_count=0
total_p, total_t = 0, 0
for index in BatchSampler(SubsetRandomSampler(range(num)), batchSize, False):
action_prob,pred = model(x=state[index], out_pred=out_pred[index], type='training')
action_loss = F.cross_entropy(action_prob, action[index].view(-1))
theta_loss = F.l1_loss(pred, label[index])
loss = action_loss + theta_loss
optimizer.zero_grad() # Delete old gradients
loss.backward() # Perform backward step to compute new gradients
nn.utils.clip_grad_norm_(model.parameters(), 0.6) # Clip gradients
optimizer.step() # Perform training step bas ed on gradients
total_p += action_loss
total_t += theta_loss
_count += 1
episode += 1
lr = adjust_learning_rate(init_lr,1e6,episode,optimizer)
stats['policy_loss'].append(total_p.cpu().item()/_count)
stats['pred_loss'].append(total_t.cpu().item()/_count)
stats['learning_rate'].append(lr)
writeSummary(writer,stats, episode)
if episode % 500 == 0:
torch.save(model.state_dict(), model_path + str(episode) + '.pkl')