-
Notifications
You must be signed in to change notification settings - Fork 0
/
data.py
201 lines (174 loc) · 9.11 KB
/
data.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
import torch
from torch.nn.utils.rnn import pad_sequence
import numpy as np
"""
this module performs preprocessing on colledted data and used to
generate pytorch datasets
Check JointLandmarkDataset which is used alongside MMCNP model
"""
def map_to_unit_circle( x, y, fac = None):
"""
normalizes the collected landmark trajectories to unit circle
during testing, use the average fac coming from the training data
"""
x = x - x[:,0, ...].unsqueeze(-1)
y = y - y[:, 0, ...].unsqueeze(-1)
if fac == None:
fac = x**2 + y **2
fac = (fac).max(dim = -1)[0]
fac = torch.sqrt(fac)
x = x/ fac.unsqueeze(-1)
y = y/fac.unsqueeze(-1)
xy = torch.cat([x.unsqueeze(-1),y.unsqueeze(-1)], dim = -1)
return xy, fac
def FIR(landmark) :
"""
applies a low pass filter to landmarks
depending on the number of steps, filter coefficients may change
number of timesteps are hardcoded, filter properties may change with number of timesteps
landmark dimension:
[action, timesteps, landmark_shape]
"""
for l in range(landmark.shape[-1 ]):
for act in range(landmark.shape[0]):
z = torch.cat([landmark[[act], :, l], landmark[act, [-1], l].reshape(1,1) ], dim = -1)
for i in range(6):
z = torch.cat([z, landmark[act, [-1], l].reshape(1,1) ], dim = -1)
k = torch.nn.functional.conv1d(z, torch.tensor([0.2, 0.2, 0.2,0.2,0.2]).reshape(1,1,5), bias=None, stride=1, padding='same', dilation=1, groups=1)
#depending on num timesteps, this assignment may change
landmark[act, 5:, l] = k[0, 5:30]
return landmark
class CNPDemonstrationDataset:
def __init__(self) -> None:
self.N = 0
self.data = []
def get_sample(self, batch_size=1, max_context=10, max_target=10, R_g= None):
context_all, target_all, context_mask, target_mask = [], [], [], []
for _ in range(batch_size):
n_context = torch.randint(1, max_context, ())
n_target = torch.randint(1, max_target, ())
idx = torch.randint(0, self.N, ())
traj = self.data[idx]
if R_g is None:
R = torch.randperm(traj.shape[0])
else:
R = R_g[_]
context = traj[R[:n_context]]
target = traj[R[:(n_context+n_target)]]
context_all.append(context)
target_all.append(target)
context_mask.append(torch.ones(context.shape[0]))
target_mask.append(torch.ones(target.shape[0]))
context_all = pad_sequence(context_all, batch_first=True)
target_all = pad_sequence(target_all, batch_first=True)
context_mask = pad_sequence(context_mask, batch_first=True)
target_mask = pad_sequence(target_mask, batch_first=True)
return context_all, target_all, context_mask, target_mask
class LandmarkDataset(CNPDemonstrationDataset):
def __init__(self, data_path):
landmark = torch.load(data_path)
n_traj, n_length, n_landmark, n_dim = landmark.shape
self.N = n_traj
self.data = landmark.reshape(n_traj, n_length, n_landmark*n_dim)
time = torch.linspace(0, 1, n_length).repeat(n_traj, 1).unsqueeze(2)
self.data = torch.cat([time, self.data], dim=-1)
class JointDataset(CNPDemonstrationDataset):
def __init__(self, data_path):
self.data = torch.load(data_path)
self.N = self.data.shape[0]
time = torch.linspace(0, 1, self.data.shape[1]).repeat(self.N, 1).unsqueeze(2)
self.data = torch.cat([time, self.data], dim=-1)
"""
dataset definition containing both joints and landmarks
phase is used when synthetic data is used
use phase 0 for training data and some other value for validation data
"""
class JointLandmarkDataset(CNPDemonstrationDataset):
def __init__(self, joint_path, landmark_path, phase = 0, synthetic = False):
if synthetic:
# generate synthetic data
landmark= torch.zeros((6, 30,1 , 2))
self.joint = torch.zeros((6,30,5))
for i in range(landmark.shape[0]):
ts = torch.arange(landmark.shape[1]) /landmark.shape[1]
act = i % 7
landmark[i,: ,0, 0] = ts * np.cos(act * np.pi/6 + phase)
landmark[i,: ,0, 1] = ts * np.sin(act * np.pi/6 + phase)
for i in range(landmark.shape[0]):
for k in range(self.joint.shape[-1]):
self.joint[i,:,k] = np.sin(ts *10 * np.pi/6 +act * np.pi/8 +phase + k)
else:
landmark = torch.load(landmark_path)[:,:,20,:]
self.joint= torch.load(joint_path)[:,:,:5]
landmark = FIR(landmark)
landmark, self.fac = map_to_unit_circle(landmark[:,:,0], landmark[:,:,1])
landmark = landmark.unsqueeze(-2)
n_traj, n_length, n_landmark, n_dim = landmark.shape
self.N = n_traj
self.data = landmark.reshape(n_traj, n_length, n_landmark*n_dim)
time = torch.linspace(0, 1, n_length).repeat(n_traj, 1).unsqueeze(2)
self.landmark = torch.cat([time, self.data], dim=-1)
# Currenytly 5 joints are used out of 7
self.N = self.data.shape[0]
self.idx = 0
time = torch.linspace(0, 1, self.data.shape[1]).repeat(self.N, 1).unsqueeze(2)
self.joint = torch.cat([time, self.joint], dim=-1)
"""
generates samples for training
each sample consisting of a 4 tuple:
each element in this 4 tuple consists of a 2 tuple: one for landmark and one for joints
elements are in the following:
(joint_context_all, landmark_context_all), : observation data, timestep at first index
(joint_target_all, landmark_target_all),: prediction timesteps for each embodiment
(joint_context_mask, landmark_context_mask), : observation mask for each embodiment
(joint_target_mask, landmark_target_mask) : prediction mask for each embodiment
"""
def get_sample(self, batch_size=1, max_context=10, max_target=10, R_g= None):
joint_context_all, joint_target_all, joint_context_mask, joint_target_mask = [], [], [], []
landmark_context_all, landmark_target_all, landmark_context_mask, landmark_target_mask = [], [], [], []
for _ in range(batch_size):
# randomly select context and target sizes
n_context = torch.randint(1, max_context, ())
n_target = torch.randint(1, max_target, ())
# select an action
# increment id with each sample, or randomly select
# idx = torch.randint(0, self.N, ())
idx = self.idx
self.idx = (idx + 1) %self.landmark.shape[0]
joint_traj = self.joint[idx]
landmark_traj = self.landmark[idx]
R = torch.randperm(joint_traj.shape[0])
context_joint = joint_traj[R[:n_context]]
target_joint = joint_traj[R[:(n_context+n_target)]]
context_landmark= landmark_traj[R[:n_context]]
target_landmark = landmark_traj[R[:(n_context+n_target)]]
joint_context_all.append(context_joint)
joint_target_all.append(target_joint)
joint_context_mask.append(torch.ones(context_joint.shape[0]))
joint_target_mask.append(torch.ones(target_joint.shape[0]))
landmark_context_all.append(context_landmark)
landmark_target_all.append(target_landmark)
landmark_context_mask.append(torch.ones(context_landmark.shape[0]))
landmark_target_mask.append(torch.ones(target_landmark.shape[0]))
joint_context_all = pad_sequence(joint_context_all, batch_first=True)
joint_target_all = pad_sequence(joint_target_all, batch_first=True)
joint_context_mask = pad_sequence(joint_context_mask, batch_first=True)
joint_target_mask = pad_sequence(joint_target_mask, batch_first=True)
landmark_context_all = pad_sequence(landmark_context_all, batch_first=True)
landmark_target_all = pad_sequence(landmark_target_all, batch_first=True)
landmark_context_mask = pad_sequence(landmark_context_mask, batch_first=True)
landmark_target_mask = pad_sequence(landmark_target_mask, batch_first=True)
return (joint_context_all, landmark_context_all), (joint_target_all, landmark_target_all),\
(joint_context_mask, landmark_context_mask), (joint_target_mask, landmark_target_mask)
def unequal_collate(batch):
context_all, target_all, context_mask, target_mask = [], [], [], []
for context, target in batch:
context_all.append(context)
target_all.append(target)
context_mask.append(torch.ones(context.shape[0]))
target_mask.append(torch.ones(target.shape[0]))
context_all = pad_sequence(context_all, batch_first=True)
target_all = pad_sequence(target_all, batch_first=True)
context_mask = pad_sequence(context_mask, batch_first=True)
target_mask = pad_sequence(target_mask, batch_first=True)
return context_all, target_all, context_mask, target_mask