/
random_shift.py
66 lines (50 loc) · 2.61 KB
/
random_shift.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
# =============================================================================
# MIT License
# Copyright (c) 2023 Reinforcement Learning Evolution Foundation
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of 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.
# =============================================================================
import torch as th
from torch.nn import functional as F
from rllte.common.prototype import BaseAugmentation
class RandomShift(BaseAugmentation):
"""Random shift operation for processing image-based observations.
Args:
pad (int): Padding size.
Returns:
Augmented images.
"""
def __init__(self, pad: int = 4) -> None:
super().__init__()
self.pad = pad
def forward(self, x: th.Tensor) -> th.Tensor:
n, c, h, w = x.size()
assert h == w
padding = tuple([self.pad] * 4)
x = F.pad(x, padding, "replicate")
eps = 1.0 / (h + 2 * self.pad)
arange = th.linspace(-1.0 + eps, 1.0 - eps, h + 2 * self.pad, device=x.device, dtype=x.dtype)[:h]
arange = arange.unsqueeze(0).repeat(h, 1).unsqueeze(2)
base_grid = th.cat([arange, arange.transpose(1, 0)], dim=2)
base_grid = base_grid.unsqueeze(0).repeat(n, 1, 1, 1)
# TODO: simplify this
try:
shift = th.randint(0, 2 * self.pad + 1, size=(n, 1, 1, 2), device=x.device, dtype=x.dtype)
except Exception:
shift = th.randint(0, 2 * self.pad + 1, size=(n, 1, 1, 2), dtype=x.dtype).to(x.device) # for npu device
shift *= 2.0 / (h + 2 * self.pad)
grid = base_grid + shift
return F.grid_sample(x, grid, padding_mode="zeros", align_corners=False)