/
lstm.py
308 lines (255 loc) · 12.9 KB
/
lstm.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
import itertools
import copy
import numpy as np
import torch
import trajnetplusplustools
from .modules import Hidden2Normal, InputEmbedding
from .. import augmentation
from .utils import center_scene
NAN = float('nan')
def drop_distant(xy, r=6.0):
"""
Drops pedestrians more than r meters away from primary ped
"""
distance_2 = np.sum(np.square(xy - xy[:, 0:1]), axis=2)
mask = np.nanmin(distance_2, axis=0) < r**2
return xy[:, mask], mask
class LSTM(torch.nn.Module):
def __init__(self, embedding_dim=64, hidden_dim=128, pool=None, pool_to_input=True, goal_dim=None, goal_flag=False):
""" Initialize the LSTM forecasting model
Attributes
----------
embedding_dim : Embedding dimension of location coordinates
hidden_dim : Dimension of hidden state of LSTM
pool : interaction module
pool_to_input : Bool
if True, the interaction vector is concatenated to the input embedding of LSTM [preferred]
if False, the interaction vector is added to the LSTM hidden-state
goal_dim : Embedding dimension of the unit vector pointing towards the goal
goal_flag: Bool
if True, the embedded goal vector is concatenated to the input embedding of LSTM
"""
super(LSTM, self).__init__()
self.hidden_dim = hidden_dim
self.embedding_dim = embedding_dim
self.pool = pool
self.pool_to_input = pool_to_input
## Location
scale = 4.0
self.input_embedding = InputEmbedding(2, self.embedding_dim, scale)
## Goal
self.goal_flag = goal_flag
self.goal_dim = goal_dim or embedding_dim
self.goal_embedding = InputEmbedding(2, self.goal_dim, scale)
goal_rep_dim = self.goal_dim if self.goal_flag else 0
## Pooling
pooling_dim = 0
if pool is not None and self.pool_to_input:
pooling_dim = self.pool.out_dim
## LSTMs
self.encoder = torch.nn.LSTMCell(self.embedding_dim + goal_rep_dim + pooling_dim, self.hidden_dim)
self.decoder = torch.nn.LSTMCell(self.embedding_dim + goal_rep_dim + pooling_dim, self.hidden_dim)
# Predict the parameters of a multivariate normal:
# mu_vel_x, mu_vel_y, sigma_vel_x, sigma_vel_y, rho
self.hidden2normal = Hidden2Normal(self.hidden_dim)
def step(self, lstm, hidden_cell_state, obs1, obs2, goals, batch_split):
"""Do one step of prediction: two inputs to one normal prediction.
Parameters
----------
lstm: torch nn module [Encoder / Decoder]
The module responsible for prediction
hidden_cell_state : tuple (hidden_state, cell_state)
Current hidden_cell_state of the pedestrians
obs1 : Tensor [num_tracks, 2]
Previous x-y positions of the pedestrians
obs2 : Tensor [num_tracks, 2]
Current x-y positions of the pedestrians
goals : Tensor [num_tracks, 2]
Goal coordinates of the pedestrians
Returns
-------
hidden_cell_state : tuple (hidden_state, cell_state)
Updated hidden_cell_state of the pedestrians
normals : Tensor [num_tracks, 5]
Parameters of a multivariate normal of the predicted position
with respect to the current position
"""
num_tracks = len(obs2)
# mask for pedestrians absent from scene (partial trajectories)
# consider only the hidden states of pedestrains present in scene
track_mask = (torch.isnan(obs1[:, 0]) + torch.isnan(obs2[:, 0])) == 0
## Masked Hidden Cell State
hidden_cell_stacked = [
torch.stack([h for m, h in zip(track_mask, hidden_cell_state[0]) if m], dim=0),
torch.stack([c for m, c in zip(track_mask, hidden_cell_state[1]) if m], dim=0),
]
## Mask current velocity & embed
curr_velocity = obs2 - obs1
curr_velocity = curr_velocity[track_mask]
input_emb = self.input_embedding(curr_velocity)
## Mask Goal direction & embed
if self.goal_flag:
## Get relative direction to goals (wrt current position)
norm_factors = (torch.norm(obs2 - goals, dim=1))
goal_direction = (obs2 - goals) / norm_factors.unsqueeze(1)
goal_direction[norm_factors == 0] = torch.tensor([0., 0.], device=obs1.device)
goal_direction = goal_direction[track_mask]
goal_emb = self.goal_embedding(goal_direction)
input_emb = torch.cat([input_emb, goal_emb], dim=1)
## Mask & Pool per scene
if self.pool is not None:
hidden_states_to_pool = torch.stack(hidden_cell_state[0]).clone() # detach?
batch_pool = []
## Iterate over scenes
for (start, end) in zip(batch_split[:-1], batch_split[1:]):
## Mask for the scene
scene_track_mask = track_mask[start:end]
## Get observations and hidden-state for the scene
prev_position = obs1[start:end][scene_track_mask]
curr_position = obs2[start:end][scene_track_mask]
curr_hidden_state = hidden_states_to_pool[start:end][scene_track_mask]
## Provide track_mask to the interaction encoders
## Everyone absent by default. Only those visible in current scene are present
interaction_track_mask = torch.zeros(num_tracks, device=obs1.device).bool()
interaction_track_mask[start:end] = track_mask[start:end]
self.pool.track_mask = interaction_track_mask
## Pool
pool_sample = self.pool(curr_hidden_state, prev_position, curr_position)
batch_pool.append(pool_sample)
pooled = torch.cat(batch_pool)
if self.pool_to_input:
input_emb = torch.cat([input_emb, pooled], dim=1)
else:
hidden_cell_stacked[0] += pooled
# LSTM step
hidden_cell_stacked = lstm(input_emb, hidden_cell_stacked)
normal_masked = self.hidden2normal(hidden_cell_stacked[0])
# unmask [Update hidden-states and next velocities of pedestrians]
normal = torch.full((track_mask.size(0), 5), NAN, device=obs1.device)
mask_index = [i for i, m in enumerate(track_mask) if m]
for i, h, c, n in zip(mask_index,
hidden_cell_stacked[0],
hidden_cell_stacked[1],
normal_masked):
hidden_cell_state[0][i] = h
hidden_cell_state[1][i] = c
normal[i] = n
return hidden_cell_state, normal
def forward(self, observed, goals, batch_split, prediction_truth=None, n_predict=None):
"""Forecast the entire sequence
Parameters
----------
observed : Tensor [obs_length, num_tracks, 2]
Observed sequences of x-y coordinates of the pedestrians
goals : Tensor [num_tracks, 2]
Goal coordinates of the pedestrians
batch_split : Tensor [batch_size + 1]
Tensor defining the split of the batch.
Required to identify the tracks of to the same scene
prediction_truth : Tensor [pred_length - 1, num_tracks, 2]
Prediction sequences of x-y coordinates of the pedestrians
Helps in teacher forcing wrt neighbours positions during training
n_predict: Int
Length of sequence to be predicted during test time
Returns
-------
rel_pred_scene : Tensor [pred_length, num_tracks, 5]
Predicted velocities of pedestrians as multivariate normal
i.e. positions relative to previous positions
pred_scene : Tensor [pred_length, num_tracks, 2]
Predicted positions of pedestrians i.e. absolute positions
"""
assert ((prediction_truth is None) + (n_predict is None)) == 1
if n_predict is not None:
# -1 because one prediction is done by the encoder already
prediction_truth = [None for _ in range(n_predict - 1)]
# initialize: Because of tracks with different lengths and the masked
# update, the hidden state for every LSTM needs to be a separate object
# in the backprop graph. Therefore: list of hidden states instead of
# a single higher rank Tensor.
num_tracks = observed.size(1)
hidden_cell_state = (
[torch.zeros(self.hidden_dim, device=observed.device) for _ in range(num_tracks)],
[torch.zeros(self.hidden_dim, device=observed.device) for _ in range(num_tracks)],
)
## Reset LSTMs of Interaction Encoders.
if self.pool is not None:
self.pool.reset(num_tracks, device=observed.device)
# list of predictions
normals = [] # predicted normal parameters for both phases
positions = [] # true (during obs phase) and predicted positions
if len(observed) == 2:
positions = [observed[-1]]
# encoder
for obs1, obs2 in zip(observed[:-1], observed[1:]):
##LSTM Step
hidden_cell_state, normal = self.step(self.encoder, hidden_cell_state, obs1, obs2, goals, batch_split)
# concat predictions
normals.append(normal)
positions.append(obs2 + normal[:, :2]) # no sampling, just mean
# initialize predictions with last position to form velocity. DEEP COPY !!!
prediction_truth = copy.deepcopy(list(itertools.chain.from_iterable(
(observed[-1:], prediction_truth)
)))
# decoder, predictions
for obs1, obs2 in zip(prediction_truth[:-1], prediction_truth[1:]):
if obs1 is None:
obs1 = positions[-2].detach() # DETACH!!!
else:
for primary_id in batch_split[:-1]:
obs1[primary_id] = positions[-2][primary_id].detach() # DETACH!!!
if obs2 is None:
obs2 = positions[-1].detach()
else:
for primary_id in batch_split[:-1]:
obs2[primary_id] = positions[-1][primary_id].detach() # DETACH!!!
hidden_cell_state, normal = self.step(self.decoder, hidden_cell_state, obs1, obs2, goals, batch_split)
# concat predictions
normals.append(normal)
positions.append(obs2 + normal[:, :2]) # no sampling, just mean
# Pred_scene: Tensor [seq_length, num_tracks, 2]
# Absolute positions of all pedestrians
# Rel_pred_scene: Tensor [seq_length, num_tracks, 5]
# Velocities of all pedestrians
rel_pred_scene = torch.stack(normals, dim=0)
pred_scene = torch.stack(positions, dim=0)
return rel_pred_scene, pred_scene
class LSTMPredictor(object):
def __init__(self, model):
self.model = model
def save(self, state, filename):
with open(filename, 'wb') as f:
torch.save(self, f)
# # during development, good for compatibility across API changes:
# # Save state for optimizer to continue training in future
with open(filename + '.state', 'wb') as f:
torch.save(state, f)
@staticmethod
def load(filename):
with open(filename, 'rb') as f:
return torch.load(f)
def __call__(self, paths, scene_goal, n_predict=12, modes=1, predict_all=True, obs_length=9, start_length=0, args=None):
self.model.eval()
# self.model.train()
with torch.no_grad():
xy = trajnetplusplustools.Reader.paths_to_xy(paths)
# xy = augmentation.add_noise(xy, thresh=args.thresh, ped=args.ped_type)
batch_split = [0, xy.shape[1]]
if args.normalize_scene:
xy, rotation, center, scene_goal = center_scene(xy, obs_length, goals=scene_goal)
xy = torch.Tensor(xy) #.to(self.device)
scene_goal = torch.Tensor(scene_goal) #.to(device)
batch_split = torch.Tensor(batch_split).long()
multimodal_outputs = {}
for num_p in range(modes):
# _, output_scenes = self.model(xy[start_length:obs_length], scene_goal, batch_split, xy[obs_length:-1].clone())
_, output_scenes = self.model(xy[start_length:obs_length], scene_goal, batch_split, n_predict=n_predict)
output_scenes = output_scenes.numpy()
if args.normalize_scene:
output_scenes = augmentation.inverse_scene(output_scenes, rotation, center)
output_primary = output_scenes[-n_predict:, 0]
output_neighs = output_scenes[-n_predict:, 1:]
## Dictionary of predictions. Each key corresponds to one mode
multimodal_outputs[num_p] = [output_primary, output_neighs]
## Return Dictionary of predictions. Each key corresponds to one mode
return multimodal_outputs