-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathdatasets.py
152 lines (139 loc) · 6.97 KB
/
datasets.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
"""
Non-commercial Use License
Copyright (c) 2021 Siemens Technology
This software, along with associated documentation files (the "Software"), is
provided for the sole purpose of providing Proof of Concept. Any commercial
uses of the Software including, but not limited to, the rights to sublicense,
and/or sell copies of the Software are prohibited and are subject to a
separate licensing agreement with Siemens. This software may be proprietary
to Siemens and may be covered by patent and copyright laws. Processes
controlled by the Software are patent pending.
The above copyright notice and this permission notice shall remain attached
to the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
"""
from torch.utils.data import Dataset
import torch
import os
from pytorch_lightning import seed_everything
from systems.bouncing_point_masses import BouncingPointMasses
def rel_err(x, y):
return (((x - y) ** 2).sum() / ((x + y) ** 2).sum()).sqrt()
class RigidBodyDataset(Dataset):
def __init__(
self,
root_dir=os.path.dirname(os.path.abspath(__file__)),
body=BouncingPointMasses(),
n_traj=100,
mode="train",
dtype=torch.float32,
chunk_len=5,
regen=False,
noise_std=0.0,
separate=False,
):
super().__init__()
self.mode = mode
self.body = body
self.noise_std = noise_std
self.separate = separate
filename = os.path.join(
root_dir, f"traj_{body}_N{n_traj}_{mode}.pt"
)
if os.path.exists(filename) and not regen:
ts, zs, is_clds = torch.load(filename)
else:
print(f"generating trajectories (mode: {mode}), this might take a while...")
seed_everything(0)
ts, zs, is_clds = self.generate_trajectory_data(n_traj, separate=separate)
os.makedirs(root_dir, exist_ok=True)
torch.save((ts, zs, is_clds), filename)
seed_everything(0)
ts, zs, is_clds = self.chunk_training_data(ts, zs, is_clds, chunk_len)
self.ts, self.zs, self.is_clds = ts.to(dtype=dtype), zs.to(dtype=dtype), is_clds
print(f"{is_clds.sum()} out of {len(is_clds)} trajectories contains collision.")
def __len__(self):
return self.zs.shape[0]
def __getitem__(self, idx):
return (self.zs[idx, 0], self.ts[idx]), self.zs[idx], self.is_clds[idx]
def generate_trajectory_data(self, n_traj, separate=False):
"""
return ts, zs
ts: n_traj, traj_len
zs: n_traj, traj_len, z_dim
"""
z0s = self.body.sample_initial_conditions(n_traj)
ts = torch.arange(
0, self.body.integration_time, self.body.dt, device=z0s.device, dtype=z0s.dtype
)
if not separate:
zs, is_clds = self.body.integrate(z0s, ts)
ts = ts.repeat(n_traj, 1)
return ts, zs, is_clds
else:
zs, is_clds = [], []
for i in range(n_traj):
z, is_cld = self.body.integrate(z0s[i:i+1], ts)
zs.append(z)
is_clds.append(is_cld)
ts = ts.repeat(n_traj, 1)
return ts, torch.cat(zs, dim=0), torch.cat(is_clds, dim=0)
def chunk_training_data(self, ts, zs, is_clds, chunk_len, p_cld=0.5):
""" Randomly samples chunks of trajectory data, returns tensors shaped for training.
Inputs: [ts (bs, T)] [zs (bs, T, *z_dim)] [is_clds (bs, T)]
outputs: [chosen_ts (bs, chunk_len)] [chosen_zs (bs, chunk_len, *z_dim)]"""
bs, T, *z_dim = zs.shape
n_chunks = (T - chunk_len + 1)
# Cut each trajectory into non-overlapping chunks
chunked_ts = torch.stack([ts[:, i:i+chunk_len] for i in range(n_chunks)], dim=0) # n_chunks, bs, chunk_len
chunked_zs = torch.stack([zs[:, i:i+chunk_len] for i in range(n_chunks)], dim=0) # n_chunks, bs, chunk_len, *z_dim
chunked_is_clds = torch.stack([is_clds[:, i:i+chunk_len] for i in range(n_chunks)], dim=0) # n_chunks, bs, chunk_len
is_clds_t0 = chunked_is_clds[..., 0] # n_chunks, bs
is_clds_chunk = chunked_is_clds.sum(-1) > 0 # n_chunks, bs
is_cld = chunked_is_clds.sum(dim=-1) # n_chunks, bs
# From each trajectory, we choose a single chunk randomly
# we make sure that the initial condition is not during collision
chosen_ts = torch.zeros(bs, chunk_len, dtype=ts.dtype, device=ts.device)
chosen_zs = torch.zeros(bs, chunk_len, *chunked_zs.shape[3:], dtype=zs.dtype, device=zs.device)
is_cld_in_chosen = torch.zeros(bs, dtype=torch.bool, device=zs.device)
# we make sure there are roughly p_cld trajectories that contains collision
is_cld_T = is_clds.sum(-1) > 0
cld_ratio = (is_cld_T).sum().true_divide(bs)
if cld_ratio < p_cld:
contains_cld = is_cld_T
else:
cld_idx = torch.nonzero(is_cld_T, as_tuple=False)[:, 0]
rand_idx = torch.rand(len(cld_idx)) < p_cld / cld_ratio
contains_cld = torch.zeros(bs, dtype=torch.bool, device=zs.device)
contains_cld[cld_idx[rand_idx]] = True
for i in range(bs):
no_cld0_no_cld_chunk_idx = torch.nonzero(
torch.logical_and(is_clds_t0[:, i] == 0, is_clds_chunk[:, i] == 0),
as_tuple=False
)[:, 0]
no_cld0_cld_chunk_idx = torch.nonzero(
torch.logical_and(is_clds_t0[:, i] == 0, is_clds_chunk[:, i] == 1),
as_tuple=False
)[:, 0]
if (contains_cld[i] and len(no_cld0_cld_chunk_idx) > 0) or len(no_cld0_no_cld_chunk_idx) == 0:
rand_idx = torch.randint(0, len(no_cld0_cld_chunk_idx), (1,), device=zs.device)[0]
chunk_idx = no_cld0_cld_chunk_idx[rand_idx]
else:
rand_idx = torch.randint(0, len(no_cld0_no_cld_chunk_idx), (1,), device=zs.device)[0]
chunk_idx = no_cld0_no_cld_chunk_idx[rand_idx]
chosen_ts[i, :] = chunked_ts[chunk_idx, i]
chosen_zs[i, :] = chunked_zs[chunk_idx, i]
is_cld_in_chosen[i] = chunked_is_clds[chunk_idx, i].sum() > 0
if self.body.__class__.__name__ == "ChainPendulumWithContact":
chosen_q_qdot = self.body.global_cartesian_to_angle(chosen_zs)
chosen_q_qdot = chosen_q_qdot + torch.randn(*chosen_q_qdot.shape) * self.noise_std
noisy_zs = self.body.angle_to_global_cartesian(chosen_q_qdot)
else:
noisy_zs = chosen_zs + torch.randn(*chosen_zs.shape) * self.noise_std
return chosen_ts, noisy_zs, is_cld_in_chosen