File size: 9,276 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from typing import Tuple, Dict

import torch
import numpy as np
import torch.nn.functional as F


class Pd(object):
    """
    Overview:
        Abstract class for parameterizable probability distributions and sampling functions.
    Interfaces:
        ``neglogp``, ``entropy``, ``noise_mode``, ``mode``, ``sample``

    .. tip::

        In dereived classes, `logits` should be an attribute member stored in class.
    """

    def neglogp(self, x: torch.Tensor) -> torch.Tensor:
        """
        Overview:
            Calculate cross_entropy between input x and logits
        Arguments:
            - x (:obj:`torch.Tensor`): the input tensor
        Return:
            - cross_entropy (:obj:`torch.Tensor`): the returned cross_entropy loss
        """
        raise NotImplementedError

    def entropy(self) -> torch.Tensor:
        """
        Overview:
            Calculate the softmax entropy of logits
        Arguments:
            - reduction (:obj:`str`): support [None, 'mean'], default set to 'mean'
        Returns:
            - entropy (:obj:`torch.Tensor`): the calculated entropy
        """
        raise NotImplementedError

    def noise_mode(self):
        """
        Overview:
            Add noise to logits. This method is designed for randomness
        """
        raise NotImplementedError

    def mode(self):
        """
        Overview:
            Return logits argmax result. This method is designed for deterministic.
        """
        raise NotImplementedError

    def sample(self):
        """
        Overview:
            Sample from logits's distribution by using softmax. This method is designed for multinomial.
        """
        raise NotImplementedError


class CategoricalPd(Pd):
    """
    Overview:
        Catagorical probility distribution sampler
    Interfaces:
        ``__init__``, ``neglogp``, ``entropy``, ``noise_mode``, ``mode``, ``sample``
    """

    def __init__(self, logits: torch.Tensor = None) -> None:
        """
        Overview:
            Init the Pd with logits
        Arguments:
            - logits (:obj:torch.Tensor): logits to sample from
        """
        self.update_logits(logits)

    def update_logits(self, logits: torch.Tensor) -> None:
        """
        Overview:
            Updata logits
        Arguments:
            - logits (:obj:`torch.Tensor`): logits to update
        """
        self.logits = logits

    def neglogp(self, x, reduction: str = 'mean') -> torch.Tensor:
        """
        Overview:
            Calculate cross_entropy between input x and logits
        Arguments:
            - x (:obj:`torch.Tensor`): the input tensor
            - reduction (:obj:`str`): support [None, 'mean'], default set to mean
        Return:
            - cross_entropy (:obj:`torch.Tensor`): the returned cross_entropy loss
        """
        return F.cross_entropy(self.logits, x, reduction=reduction)

    def entropy(self, reduction: str = 'mean') -> torch.Tensor:
        """
        Overview:
            Calculate the softmax entropy of logits
        Arguments:
            - reduction (:obj:`str`): support [None, 'mean'], default set to mean
        Returns:
            - entropy (:obj:`torch.Tensor`): the calculated entropy
        """
        a = self.logits - self.logits.max(dim=-1, keepdim=True)[0]
        ea = torch.exp(a)
        z = ea.sum(dim=-1, keepdim=True)
        p = ea / z
        entropy = (p * (torch.log(z) - a)).sum(dim=-1)
        assert (reduction in [None, 'mean'])
        if reduction is None:
            return entropy
        elif reduction == 'mean':
            return entropy.mean()

    def noise_mode(self, viz: bool = False) -> Tuple[torch.Tensor, Dict[str, np.ndarray]]:
        """
        Overview:
            add noise to logits
        Arguments:
            - viz (:obj:`bool`): Whether to return numpy from of logits, noise and noise_logits; \
                Short for ``visualize`` . (Because tensor type cannot visualize in tb or text log)
        Returns:
            - result (:obj:`torch.Tensor`): noised logits
            - viz_feature (:obj:`Dict[str, np.ndarray]`): ndarray type data for visualization.
        """
        u = torch.rand_like(self.logits)
        u = -torch.log(-torch.log(u))
        noise_logits = self.logits + u
        result = noise_logits.argmax(dim=-1)
        if viz:
            viz_feature = {}
            viz_feature['logits'] = self.logits.data.cpu().numpy()
            viz_feature['noise'] = u.data.cpu().numpy()
            viz_feature['noise_logits'] = noise_logits.data.cpu().numpy()
            return result, viz_feature
        else:
            return result

    def mode(self, viz: bool = False) -> Tuple[torch.Tensor, Dict[str, np.ndarray]]:
        """
        Overview:
            return logits argmax result
        Arguments:
            - viz (:obj:`bool`): Whether to return numpy from of logits, noise and noise_logits;
                Short for ``visualize`` . (Because tensor type cannot visualize in tb or text log)
        Returns:
            - result (:obj:`torch.Tensor`): the logits argmax result
            - viz_feature (:obj:`Dict[str, np.ndarray]`): ndarray type data for visualization.
        """
        result = self.logits.argmax(dim=-1)
        if viz:
            viz_feature = {}
            viz_feature['logits'] = self.logits.data.cpu().numpy()
            return result, viz_feature
        else:
            return result

    def sample(self, viz: bool = False) -> Tuple[torch.Tensor, Dict[str, np.ndarray]]:
        """
        Overview:
            Sample from logits's distribution by using softmax
        Arguments:
            - viz (:obj:`bool`): Whether to return numpy from of logits, noise and noise_logits; \
                Short for ``visualize`` . (Because tensor type cannot visualize in tb or text log)
        Returns:
            - result (:obj:`torch.Tensor`): the logits sampled result
            - viz_feature (:obj:`Dict[str, np.ndarray]`): ndarray type data for visualization.
        """
        p = torch.softmax(self.logits, dim=1)
        result = torch.multinomial(p, 1).squeeze(1)
        if viz:
            viz_feature = {}
            viz_feature['logits'] = self.logits.data.cpu().numpy()
            return result, viz_feature
        else:
            return result


class CategoricalPdPytorch(torch.distributions.Categorical):
    """
    Overview:
        Wrapped ``torch.distributions.Categorical``

    Interfaces:
        ``__init__``, ``update_logits``, ``update_probs``, ``sample``, ``neglogp``, ``mode``, ``entropy``
    """

    def __init__(self, probs: torch.Tensor = None) -> None:
        """
        Overview:
            Initialize the CategoricalPdPytorch object.
        Arguments:
            - probs (:obj:`torch.Tensor`): The tensor of probabilities.
        """
        if probs is not None:
            self.update_probs(probs)

    def update_logits(self, logits: torch.Tensor) -> None:
        """
        Overview:
            Updata logits
        Arguments:
            - logits (:obj:`torch.Tensor`): logits to update
        """
        super().__init__(logits=logits)

    def update_probs(self, probs: torch.Tensor) -> None:
        """
        Overview:
            Updata probs
        Arguments:
            - probs (:obj:`torch.Tensor`): probs to update
        """
        super().__init__(probs=probs)

    def sample(self) -> torch.Tensor:
        """
        Overview:
            Sample from logits's distribution by using softmax
        Return:
            - result (:obj:`torch.Tensor`): the logits sampled result
        """
        return super().sample()

    def neglogp(self, actions: torch.Tensor, reduction: str = 'mean') -> torch.Tensor:
        """
        Overview:
            Calculate cross_entropy between input x and logits
        Arguments:
            - actions (:obj:`torch.Tensor`): the input action tensor
            - reduction (:obj:`str`): support [None, 'mean'], default set to mean
        Return:
            - cross_entropy (:obj:`torch.Tensor`): the returned cross_entropy loss
        """
        neglogp = super().log_prob(actions)
        assert (reduction in ['none', 'mean'])
        if reduction == 'none':
            return neglogp
        elif reduction == 'mean':
            return neglogp.mean(dim=0)

    def mode(self) -> torch.Tensor:
        """
        Overview:
            Return logits argmax result
        Return:
            - result(:obj:`torch.Tensor`): the logits argmax result
        """
        return self.probs.argmax(dim=-1)

    def entropy(self, reduction: str = None) -> torch.Tensor:
        """
        Overview:
            Calculate the softmax entropy of logits
        Arguments:
            - reduction (:obj:`str`): support [None, 'mean'], default set to mean
        Returns:
            - entropy (:obj:`torch.Tensor`): the calculated entropy
        """
        entropy = super().entropy()
        assert (reduction in [None, 'mean'])
        if reduction is None:
            return entropy
        elif reduction == 'mean':
            return entropy.mean()