-
Notifications
You must be signed in to change notification settings - Fork 5
Expand file tree
/
Copy pathencoder.py
More file actions
executable file
·152 lines (137 loc) · 5.55 KB
/
Copy pathencoder.py
File metadata and controls
executable file
·152 lines (137 loc) · 5.55 KB
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
from typing import Optional
import torch
import torch.nn as nn
from ding.torch_utils import ResFCBlock, ResBlock
from ding.utils import SequenceType
class ConvEncoder(nn.Module):
r"""
Overview:
The ``Convolution Encoder`` used in models. Used to encoder raw 2-dim observation.
Interfaces:
``__init__``, ``forward``
"""
def __init__(
self,
obs_shape: SequenceType,
hidden_size_list: SequenceType = [32, 64, 64, 128],
activation: Optional[nn.Module] = nn.ReLU(),
norm_type: Optional[str] = None
) -> None:
r"""
Overview:
Init the Convolution Encoder according to arguments.
Arguments:
- obs_shape (:obj:`SequenceType`): Sequence of ``in_channel``, some ``output size``
- hidden_size_list (:obj:`SequenceType`): The collection of ``hidden_size``
- activation (:obj:`nn.Module`):
The type of activation to use in the conv ``layers`` and ``ResBlock``,
if ``None`` then default set to ``nn.ReLU()``
- norm_type (:obj:`str`):
The type of normalization to use, see ``ding.torch_utils.ResBlock`` for more details
"""
super(ConvEncoder, self).__init__()
self.obs_shape = obs_shape
self.act = activation
self.hidden_size_list = hidden_size_list
layers = []
kernel_size = [8, 4, 3]
stride = [4, 2, 1]
input_size = obs_shape[0] # in_channel
for i in range(len(kernel_size)):
layers.append(nn.Conv2d(input_size, hidden_size_list[i], kernel_size[i], stride[i]))
layers.append(self.act)
input_size = hidden_size_list[i]
assert len(set(hidden_size_list[3:-1])) <= 1, "Please indicate the same hidden size for res block parts"
for i in range(3, len(self.hidden_size_list) - 1):
layers.append(ResBlock(self.hidden_size_list[i], activation=self.act, norm_type=norm_type))
layers.append(nn.Flatten())
self.main = nn.Sequential(*layers)
flatten_size = self._get_flatten_size()
self.mid = nn.Linear(flatten_size, hidden_size_list[-1])
def _get_flatten_size(self) -> int:
r"""
Overview:
Get the encoding size after ``self.main`` to get the number of ``in-features`` to feed to ``nn.Linear``.
Arguments:
- x (:obj:`torch.Tensor`): Encoded Tensor after ``self.main``
Returns:
- outputs (:obj:`torch.Tensor`): Size int, also number of in-feature
"""
test_data = torch.randn(1, *self.obs_shape)
with torch.no_grad():
output = self.main(test_data)
return output.shape[1]
def forward(self, x: torch.Tensor) -> torch.Tensor:
r"""
Overview:
Return embedding tensor of the env observation
Arguments:
- x (:obj:`torch.Tensor`): Env raw observation
Returns:
- outputs (:obj:`torch.Tensor`): Embedding tensor
"""
x = self.main(x)
x = self.mid(x)
return x
class FCEncoder(nn.Module):
r"""
Overview:
The ``FCEncoder`` used in models. Used to encoder raw 1-dim observation.
Interfaces:
``__init__``, ``forward``
"""
def __init__(
self,
obs_shape: int,
hidden_size_list: SequenceType,
res_block: bool = False,
activation: Optional[nn.Module] = nn.ReLU(),
norm_type: Optional[str] = None
) -> None:
r"""
Overview:
Init the FC Encoder according to arguments.
Arguments:
- obs_shape (:obj:`int`): Observation shape
- hidden_size_list (:obj:`SequenceType`): The collection of ``hidden_size``
- res_block (:obj:`bool`): Whether use ``res_block``.
- activation (:obj:`nn.Module`):
The type of activation to use in the ``ResFCBlock``,
if ``None`` then default set to ``nn.ReLU()``
- norm_type (:obj:`str`):
The type of normalization to use, see ``ding.torch_utils.ResFCBlock`` for more details
"""
super(FCEncoder, self).__init__()
self.obs_shape = obs_shape
self.act = activation
self.init = nn.Linear(obs_shape, hidden_size_list[0])
if res_block:
assert len(set(hidden_size_list)) == 1, "Please indicate the same hidden size for res block parts"
if len(hidden_size_list) == 1:
self.main = ResFCBlock(hidden_size_list[0], activation=self.act, norm_type=norm_type)
else:
layers = []
for i in range(len(hidden_size_list)):
layers.append(ResFCBlock(hidden_size_list[0], activation=self.act, norm_type=norm_type))
self.main = nn.Sequential(*layers)
else:
layers = []
for i in range(len(hidden_size_list) - 1):
layers.append(nn.Linear(hidden_size_list[i], hidden_size_list[i + 1]))
layers.append(self.act)
self.main = nn.Sequential(*layers)
def forward(self, x: torch.Tensor) -> torch.Tensor:
r"""
Overview:
Return embedding tensor of the env observation
Arguments:
- x (:obj:`torch.Tensor`): Env raw observation
Returns:
- outputs (:obj:`torch.Tensor`): Embedding tensor
"""
x = self.act(self.init(x))
x = self.main(x)
return x
class StructEncoder(nn.Module):
# TODO(nyz)
pass