-
Notifications
You must be signed in to change notification settings - Fork 2
/
stretchbev.py
467 lines (394 loc) · 21.2 KB
/
stretchbev.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
import torch
import torch.nn as nn
from stretchbev.models import model_utils
from stretchbev.models.decoder import Decoder
from stretchbev.models.encoder import Encoder
from stretchbev.models.future_prediction import FuturePrediction
from stretchbev.models.res_models import SmallEncoder, SmallDecoder, ConvNet
from stretchbev.utils.geometry import calculate_birds_eye_view_parameters, VoxelsSumming
from stretchbev.utils.network import pack_sequence_dim, unpack_sequence_dim, set_bn_momentum
class StretchBEV(nn.Module):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
bev_resolution, bev_start_position, bev_dimension = calculate_birds_eye_view_parameters(
self.cfg.LIFT.X_BOUND, self.cfg.LIFT.Y_BOUND, self.cfg.LIFT.Z_BOUND
)
self.bev_resolution = nn.Parameter(bev_resolution, requires_grad=False)
self.bev_start_position = nn.Parameter(bev_start_position, requires_grad=False)
self.bev_dimension = nn.Parameter(bev_dimension, requires_grad=False)
self.encoder_downsample = self.cfg.MODEL.ENCODER.DOWNSAMPLE
self.encoder_out_channels = self.cfg.MODEL.ENCODER.OUT_CHANNELS
self.min_log_sigma = self.cfg.MODEL.DISTRIBUTION.MIN_LOG_SIGMA
self.max_log_sigma = self.cfg.MODEL.DISTRIBUTION.MAX_LOG_SIGMA
self.skipco = self.cfg.MODEL.SMALL_ENCODER.SKIPCO
self.frustum = self.create_frustum()
self.depth_channels, _, _, _ = self.frustum.shape
if self.cfg.TIME_RECEPTIVE_FIELD == 1:
assert self.cfg.MODEL.TEMPORAL_MODEL.NAME == 'identity'
# temporal block
self.receptive_field = self.cfg.TIME_RECEPTIVE_FIELD
self.n_future = self.cfg.N_FUTURE_FRAMES
self.latent_dim = self.cfg.MODEL.DISTRIBUTION.LATENT_DIM
if self.cfg.MODEL.SUBSAMPLE:
assert self.cfg.DATASET.NAME == 'lyft'
self.receptive_field = 3
self.n_future = 5
# Spatial extent in bird's-eye view, in meters
self.spatial_extent = (self.cfg.LIFT.X_BOUND[1], self.cfg.LIFT.Y_BOUND[1])
self.bev_size = (self.bev_dimension[0].item(), self.bev_dimension[1].item())
# Encoder
self.encoder = Encoder(cfg=self.cfg.MODEL.ENCODER, D=self.depth_channels)
self.srvp_encoder = SmallEncoder(self.cfg.MODEL.ENCODER.OUT_CHANNELS, self.cfg.MODEL.ENCODER.OUT_CHANNELS,
self.cfg.MODEL.SMALL_ENCODER.FILTER_SIZE)
self.srvp_decoder = SmallDecoder(self.cfg.MODEL.ENCODER.OUT_CHANNELS, self.cfg.MODEL.ENCODER.OUT_CHANNELS,
self.cfg.MODEL.SMALL_ENCODER.FILTER_SIZE, self.skipco)
# in_channels: y*conditioning_frames, out_channels: y*2 (mean, sigma), num_layers
self.q_y = ConvNet(self.cfg.MODEL.ENCODER.OUT_CHANNELS * self.receptive_field,
self.cfg.MODEL.ENCODER.OUT_CHANNELS * 2,
self.cfg.MODEL.FIRST_STATE.NUM_LAYERS)
# residual update predictor
self.dynamics = ConvNet(self.cfg.MODEL.ENCODER.OUT_CHANNELS + self.cfg.MODEL.DISTRIBUTION.LATENT_DIM,
self.cfg.MODEL.ENCODER.OUT_CHANNELS,
self.cfg.MODEL.DYNAMICS.NUM_LAYERS) # in_channels: y+z, out_channels: y, num_layers
# inference of z, this is for processing posterior samples before sampling distribution parameters
# q_z = MLP(LSTM(x)) in srvp, this is the LSTM
self.inf_z = FuturePrediction(self.cfg.MODEL.ENCODER.OUT_CHANNELS, self.cfg.MODEL.ENCODER.OUT_CHANNELS + 6,
n_gru_blocks=2, n_res_layers=1)
# SpatialGRU(self.cfg.MODEL.ENCODER.OUT_CHANNELS+6, self.cfg.MODEL.ENCODER.OUT_CHANNELS)
# posterior sampling
# in_channels: y, out_channels: z*2, num_layers
self.q_z = ConvNet(self.cfg.MODEL.ENCODER.OUT_CHANNELS, self.cfg.MODEL.DISTRIBUTION.LATENT_DIM * 2,
self.cfg.MODEL.DISTRIBUTION.POSTERIOR_LAYERS)
# in_channels: y, out_channels: z*2, num_layers
self.p_z = ConvNet(self.cfg.MODEL.ENCODER.OUT_CHANNELS, self.cfg.MODEL.DISTRIBUTION.LATENT_DIM * 2,
self.cfg.MODEL.DISTRIBUTION.PRIOR_LAYERS)
# Decoder
self.decoder = Decoder(
in_channels=self.cfg.MODEL.ENCODER.OUT_CHANNELS,
n_classes=len(self.cfg.SEMANTIC_SEG.WEIGHTS),
predict_future_flow=self.cfg.INSTANCE_FLOW.ENABLED,
)
set_bn_momentum(self, self.cfg.MODEL.BN_MOMENTUM)
def create_frustum(self):
# Create grid in image plane
h, w = self.cfg.IMAGE.FINAL_DIM
downsampled_h, downsampled_w = h // self.encoder_downsample, w // self.encoder_downsample
# Depth grid
depth_grid = torch.arange(*self.cfg.LIFT.D_BOUND, dtype=torch.float)
depth_grid = depth_grid.view(-1, 1, 1).expand(-1, downsampled_h, downsampled_w)
n_depth_slices = depth_grid.shape[0]
# x and y grids
x_grid = torch.linspace(0, w - 1, downsampled_w, dtype=torch.float)
x_grid = x_grid.view(1, 1, downsampled_w).expand(n_depth_slices, downsampled_h, downsampled_w)
y_grid = torch.linspace(0, h - 1, downsampled_h, dtype=torch.float)
y_grid = y_grid.view(1, downsampled_h, 1).expand(n_depth_slices, downsampled_h, downsampled_w)
# Dimension (n_depth_slices, downsampled_h, downsampled_w, 3)
# containing data points in the image: left-right, top-bottom, depth
frustum = torch.stack((x_grid, y_grid, depth_grid), -1)
return nn.Parameter(frustum, requires_grad=False)
def forward(self, image, intrinsics, extrinsics, future_egomotion, future_distribution_inputs=None, noise=None,
nt=None):
output = {}
# Only process features from the past and present
# Lifting features and project to bird's-eye view
x = self.calculate_birds_eye_view_features(image, intrinsics, extrinsics)
srvp_x, skips = self.srvp_encode(x) # this will encode features into a more meaningful (?) space
y_0, q_y0_params = self.infer_first_state(
srvp_x[:, :self.receptive_field].contiguous()) # this will create the first state
ys, zs, q_z_params, p_z_params = self.srvp_generate(y_0, srvp_x, future_distribution_inputs, nt=nt)
# above line will return intermediate ys, zs, distribution parameters and residual changes
generated_srvp_x = self.srvp_decode(ys, skips) # this will decode intermediate states into LSS feature space
bev_output = self.decoder(generated_srvp_x) # this will decode LSS features into FIERY outputs
output = {
'bev_output': bev_output,
'generated_srvp_x': generated_srvp_x,
'lss_outs': x,
'q_z_params': q_z_params,
'p_z_params': p_z_params,
'q_y0_params': q_y0_params,
}
return output
def multi_sample_inference(self, image, intrinsics, extrinsics, future_egomotion, num_samples=1,
future_distribution_inputs=None, noise=None, nt=None):
image = image[:, :self.receptive_field].contiguous()
intrinsics = intrinsics[:, :self.receptive_field].contiguous()
extrinsics = extrinsics[:, :self.receptive_field].contiguous()
future_egomotion = future_egomotion[:, :self.receptive_field].contiguous()
# Lifting features and project to bird's-eye view
x = self.calculate_birds_eye_view_features(image, intrinsics, extrinsics)
srvp_x, skips = self.srvp_encode(x) # this will encode features into a more meaningful (?) space
y_0, q_y0_params = self.infer_first_state(
srvp_x[:, :self.receptive_field].contiguous()) # this will create the first state
hx_z = self.inf_z(torch.cat([srvp_x, future_distribution_inputs[:, :self.receptive_field].contiguous()],
dim=2)) # encodes srvp_encoder's output temporally
states = []
for s in range(num_samples):
states.append(self.inference_srvp_generate(y_0, hx_z, nt))
states = torch.cat(states, 0)
skips = [s.expand(num_samples, *s.shape[1:]) for s in skips]
generated_srvp_x = self.srvp_decode(states[:, self.receptive_field:], skips)
bev_output = self.decoder(generated_srvp_x) # this will decode LSS features into FIERY outputs
res = [{} for _ in range(num_samples)]
for key, val in bev_output.items():
for s in range(num_samples):
res[s][key] = val[s:s + 1]
return res
def inference_srvp_generate(self, y_0, hx_z, nt=0):
total_time = self.receptive_field + self.n_future if nt is None else nt
ys = [y_0]
y_tm1 = y_0
for t in range(1, total_time):
if t < self.receptive_field:
# observations are avaliable
z_t, q_z_t_params = self.infer_z(hx_z[:, t])
else:
assert not self.training
p_z_t_params = self.p_z(y_tm1)
z_t = model_utils.rsample_normal(p_z_t_params, max_log_sigma=self.max_log_sigma,
min_log_sigma=self.min_log_sigma)
y_t = self.residual_step(y_tm1, z_t)
# Update previous latent state
y_tm1 = y_t
ys.append(y_t)
return torch.stack(ys, 1)
def srvp_decode(self, x, skip):
# content vector is missing.
"""
Decodes SRVP intermediate states into LSS output space
x: SRVP intermediate states, torch.Tensor, [batch, seq_len, channels, height, width]
Returns:
torch.Tensor: [batch, seq_len, channels, height, width]
"""
b, t, c, h, w = x.shape
_x = x.reshape(b * t, c, h, w)
if skip:
skip = [s.unsqueeze(1).expand(b, t, *s.shape[1:]) for s in skip]
skip = [s.reshape(t * b, *s.shape[2:]) for s in skip]
x_out = self.srvp_decoder(_x, skip=skip)
return x_out.view(b, t, *x_out.shape[1:])
def srvp_encode(self, x):
"""
Encodes LSS's outputs
x: LSS encoded features, torch.Tensor, [batch, seq_len, channels, height, width]
Returns:
torch.Tensor: [batch, seq_len, channels, height, width]
"""
b, t, c, h, w = x.shape
_x = x.view(b * t, c, h, w)
hx, skips = self.srvp_encoder(_x, return_skip=True)
hx = hx.view(b, t, *hx.shape[1:])
if self.skipco:
if self.training:
# When training, take a random frame to compute the skip connections
tx = torch.randint(t, size=(b,)).to(hx.device)
index = torch.arange(b).to(hx.device)
skips = [s.view(b, t, *s.shape[1:])[index, tx] for s in skips]
else:
# When testing, choose the last frame
skips = [s.view(b, t, *s.shape[1:])[:, -1] for s in skips]
else:
skips = None
return hx, skips
def infer_first_state(self, x, deterministic=False):
"""
Creates the first state from the conditioning observation
x: encoded features, torch.Tensor, [batch, seq_len, channels, height, width]
deterministic: If true, it will return only the means otherwise a sample from Normal distribution
Returns:
First state of the model -> y, torch.Tensor, [batch, channels, height, width]
"""
# Q1: will the first state be stochastic?
# Q2: are we going to sample a different noise for each position
b, t, c, h, w = x.shape
_x = x.view(b, t * c, h, w)
q_y0_params = self.q_y(_x)
y_0 = model_utils.rsample_normal(q_y0_params, max_log_sigma=self.max_log_sigma,
min_log_sigma=self.min_log_sigma)
return y_0, q_y0_params
def infer_z(self, hx):
"""
Infers z from the SpatialGRU's output
hx: output of SpatialGRU, torch.Tensor, [batch, channels, height, width]
Returns:
z: torch.Tensor, [batch, channels, height, width]
qz_params: torch.Tensor, [batch, channels*2, height, width]
"""
q_z_params = self.q_z(hx)
z = model_utils.rsample_normal(q_z_params, max_log_sigma=self.max_log_sigma, min_log_sigma=self.min_log_sigma)
return z, q_z_params
def residual_step(self, y, z, dt=1):
# QUESTION: should we keep euler steps thing?
# Are we going to create states for not integer timesteps
"""
y: intermediate state, torch.Tensor, [batch, channels, height, width]
z: latent noise, torch.Tensor, [batch, channels, height, width]
Returns:
updated y: torch.Tensor, [batch, channels, height, width]
"""
res_step = self.dynamics(torch.cat([y, z], 1))
y_forward = y + dt * res_step
return y_forward
def srvp_generate(self, y_0, x, future_inputs, nt=None):
"""
Generates intermediate states with residual updates
y_0: created first state, [batch, seq_len, channels, height, width]
x: encoded features, [batch, seq_len, channels, height, width]
Returns:
ys: intermediate states, [batch, seq_len, channels, height, width] # dim might change
zs: used gaussian noise, [batch, seq_len - 1, channels, height, width] # dim might change
q_z_params: posterior distribution parameters, [batch, seq_len - 1, 2*channels, height, width] # dim might change
p_z_params: prior distribution parameters, [batch, seq_len - 1, 2*channels, height, width] # dim might change
residuals: residual changes, [batch, seq_len - 1, channels, height, width] # dim might change
"""
total_time = self.receptive_field + self.n_future if nt is None else nt
ys = [y_0]
z, q_z_params, p_z_params = [], [], []
hx_z = self.inf_z(torch.cat([x, future_inputs], dim=2)) # encodes srvp_encoder's output temporally
y_tm1 = y_0
for t in range(1, total_time):
# prior distribution
p_z_t_params = self.p_z(y_tm1)
p_z_params.append(p_z_t_params)
z_t, q_z_t_params = self.infer_z(hx_z[:, t])
q_z_params.append(q_z_t_params)
if self.training or t < self.receptive_field:
# observations are available
pass
else:
assert not self.training
z_t = model_utils.rsample_normal(p_z_t_params, max_log_sigma=self.max_log_sigma,
min_log_sigma=self.min_log_sigma)
# Residual step
y_t = self.residual_step(y_tm1, z_t)
# Update previous latent state
y_tm1 = y_t
ys.append(y_t)
y = torch.stack(ys, 1)
z = torch.stack(z, 1) if len(z) > 0 else None
q_z_params = torch.stack(q_z_params, 1) if len(q_z_params) > 0 else None
p_z_params = torch.stack(p_z_params, 1) if len(p_z_params) > 0 else None
return y, z, q_z_params, p_z_params
def get_geometry(self, intrinsics, extrinsics):
"""
Calculate the (x, y, z) 3D position of the features.
"""
rotation, translation = extrinsics[..., :3, :3], extrinsics[..., :3, 3]
B, N, _ = translation.shape
# Add batch, camera dimension, and a dummy dimension at the end
points = self.frustum.unsqueeze(0).unsqueeze(0).unsqueeze(-1)
# Camera to ego reference frame
points = torch.cat((points[:, :, :, :, :, :2] * points[:, :, :, :, :, 2:3], points[:, :, :, :, :, 2:3]), 5)
combined_transformation = rotation.matmul(torch.inverse(intrinsics))
points = combined_transformation.view(B, N, 1, 1, 1, 3, 3).matmul(points).squeeze(-1)
points += translation.view(B, N, 1, 1, 1, 3)
# The 3 dimensions in the ego reference frame are: (forward, sides, height)
return points
def encoder_forward(self, x):
# batch, n_cameras, channels, height, width
b, n, c, h, w = x.shape
x = x.view(b * n, c, h, w)
x = self.encoder(x)
x = x.view(b, n, *x.shape[1:])
x = x.permute(0, 1, 3, 4, 5, 2)
return x
def projection_to_birds_eye_view(self, x, geometry):
"""Adapted from https://github.com/nv-tlabs/lift-splat-shoot/blob/master/src/models.py#L200"""
# batch, n_cameras, depth, height, width, channels
batch, n, d, h, w, c = x.shape
output = torch.zeros(
(batch, c, self.bev_dimension[0], self.bev_dimension[1]), dtype=torch.float, device=x.device
)
# Number of 3D points
N = n * d * h * w
for b in range(batch):
# flatten x
x_b = x[b].reshape(N, c)
# Convert positions to integer indices
geometry_b = ((geometry[b] - (self.bev_start_position - self.bev_resolution / 2.0)) / self.bev_resolution)
geometry_b = geometry_b.view(N, 3).long()
# Mask out points that are outside the considered spatial extent.
mask = (
(geometry_b[:, 0] >= 0)
& (geometry_b[:, 0] < self.bev_dimension[0])
& (geometry_b[:, 1] >= 0)
& (geometry_b[:, 1] < self.bev_dimension[1])
& (geometry_b[:, 2] >= 0)
& (geometry_b[:, 2] < self.bev_dimension[2])
)
x_b = x_b[mask]
geometry_b = geometry_b[mask]
# Sort tensors so that those within the same voxel are consecutives.
ranks = (
geometry_b[:, 0] * (self.bev_dimension[1] * self.bev_dimension[2])
+ geometry_b[:, 1] * (self.bev_dimension[2])
+ geometry_b[:, 2]
)
ranks_indices = ranks.argsort()
x_b, geometry_b, ranks = x_b[ranks_indices], geometry_b[ranks_indices], ranks[ranks_indices]
# Project to bird's-eye view by summing voxels.
x_b, geometry_b = VoxelsSumming.apply(x_b, geometry_b, ranks)
bev_feature = torch.zeros((self.bev_dimension[2], self.bev_dimension[0], self.bev_dimension[1], c),
device=x_b.device)
bev_feature[geometry_b[:, 2], geometry_b[:, 0], geometry_b[:, 1]] = x_b
# Put channel in second position and remove z dimension
bev_feature = bev_feature.permute((0, 3, 1, 2))
bev_feature = bev_feature.squeeze(0)
output[b] = bev_feature
return output
def calculate_birds_eye_view_features(self, x, intrinsics, extrinsics):
b, s, n, c, h, w = x.shape
# Reshape
x = pack_sequence_dim(x)
intrinsics = pack_sequence_dim(intrinsics)
extrinsics = pack_sequence_dim(extrinsics)
geometry = self.get_geometry(intrinsics, extrinsics)
x = self.encoder_forward(x)
x = self.projection_to_birds_eye_view(x, geometry)
x = unpack_sequence_dim(x, b, s)
return x
def distribution_forward(self, present_features, future_distribution_inputs=None, noise=None):
"""
Parameters
----------
present_features: 5-D output from dynamics module with shape (b, 1, c, h, w)
future_distribution_inputs: 5-D tensor containing labels shape (b, s, cfg.PROB_FUTURE_DIM, h, w)
noise: a sample from a (0, 1) gaussian with shape (b, s, latent_dim). If None, will sample in function
Returns
-------
sample: sample taken from present/future distribution, broadcast to shape (b, s, latent_dim, h, w)
present_distribution_mu: shape (b, s, latent_dim)
present_distribution_log_sigma: shape (b, s, latent_dim)
future_distribution_mu: shape (b, s, latent_dim)
future_distribution_log_sigma: shape (b, s, latent_dim)
"""
b, s, _, h, w = present_features.size()
assert s == 1
present_mu, present_log_sigma = self.present_distribution(present_features)
future_mu, future_log_sigma = None, None
if future_distribution_inputs is not None:
# Concatenate future labels to z_t
future_features = future_distribution_inputs[:, 1:].contiguous().view(b, 1, -1, h, w)
future_features = torch.cat([present_features, future_features], dim=2)
future_mu, future_log_sigma = self.future_distribution(future_features)
if noise is None:
if self.training:
noise = torch.randn_like(present_mu)
else:
noise = torch.zeros_like(present_mu)
if self.training:
mu = future_mu
sigma = torch.exp(future_log_sigma)
else:
mu = present_mu
sigma = torch.exp(present_log_sigma)
sample = mu + sigma * noise
# Spatially broadcast sample to the dimensions of present_features
sample = sample.view(b, s, self.latent_dim, 1, 1).expand(b, s, self.latent_dim, h, w)
output_distribution = {
'present_mu': present_mu,
'present_log_sigma': present_log_sigma,
'future_mu': future_mu,
'future_log_sigma': future_log_sigma,
}
return sample, output_distribution