File size: 5,643 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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

from ding.torch_utils import Swish


class StandardScaler(nn.Module):

    def __init__(self, input_size: int):
        super(StandardScaler, self).__init__()
        self.register_buffer('std', torch.ones(1, input_size))
        self.register_buffer('mu', torch.zeros(1, input_size))

    def fit(self, data: torch.Tensor):
        std, mu = torch.std_mean(data, dim=0, keepdim=True)
        std[std < 1e-12] = 1
        self.std.data.mul_(0.0).add_(std)
        self.mu.data.mul_(0.0).add_(mu)

    def transform(self, data: torch.Tensor):
        return (data - self.mu) / self.std

    def inverse_transform(self, data: torch.Tensor):
        return self.std * data + self.mu


class EnsembleFC(nn.Module):
    __constants__ = ['in_features', 'out_features']
    in_features: int
    out_features: int
    ensemble_size: int
    weight: torch.Tensor

    def __init__(self, in_features: int, out_features: int, ensemble_size: int, weight_decay: float = 0.) -> None:
        super(EnsembleFC, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.ensemble_size = ensemble_size
        self.weight = nn.Parameter(torch.zeros(ensemble_size, in_features, out_features))
        self.weight_decay = weight_decay
        self.bias = nn.Parameter(torch.zeros(ensemble_size, 1, out_features))

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        assert input.shape[0] == self.ensemble_size and len(input.shape) == 3
        return torch.bmm(input, self.weight) + self.bias  # w times x + b

    def extra_repr(self) -> str:
        return 'in_features={}, out_features={}, ensemble_size={}, weight_decay={}'.format(
            self.in_features, self.out_features, self.ensemble_size, self.weight_decay
        )


class EnsembleModel(nn.Module):

    def __init__(
        self,
        state_size,
        action_size,
        reward_size,
        ensemble_size,
        hidden_size=200,
        learning_rate=1e-3,
        use_decay=False
    ):
        super(EnsembleModel, self).__init__()

        self.use_decay = use_decay
        self.hidden_size = hidden_size
        self.output_dim = state_size + reward_size

        self.nn1 = EnsembleFC(state_size + action_size, hidden_size, ensemble_size, weight_decay=0.000025)
        self.nn2 = EnsembleFC(hidden_size, hidden_size, ensemble_size, weight_decay=0.00005)
        self.nn3 = EnsembleFC(hidden_size, hidden_size, ensemble_size, weight_decay=0.000075)
        self.nn4 = EnsembleFC(hidden_size, hidden_size, ensemble_size, weight_decay=0.000075)
        self.nn5 = EnsembleFC(hidden_size, self.output_dim * 2, ensemble_size, weight_decay=0.0001)
        self.max_logvar = nn.Parameter(torch.ones(1, self.output_dim).float() * 0.5, requires_grad=False)
        self.min_logvar = nn.Parameter(torch.ones(1, self.output_dim).float() * -10, requires_grad=False)
        self.swish = Swish()

        def init_weights(m: nn.Module):

            def truncated_normal_init(t, mean: float = 0.0, std: float = 0.01):
                torch.nn.init.normal_(t, mean=mean, std=std)
                while True:
                    cond = torch.logical_or(t < mean - 2 * std, t > mean + 2 * std)
                    if not torch.sum(cond):
                        break
                    t = torch.where(cond, torch.nn.init.normal_(torch.ones(t.shape), mean=mean, std=std), t)
                return t

            if isinstance(m, nn.Linear) or isinstance(m, EnsembleFC):
                input_dim = m.in_features
                truncated_normal_init(m.weight, std=1 / (2 * np.sqrt(input_dim)))
                m.bias.data.fill_(0.0)

        self.apply(init_weights)

        self.optimizer = torch.optim.Adam(self.parameters(), lr=learning_rate)

    def forward(self, x: torch.Tensor, ret_log_var: bool = False):
        x = self.swish(self.nn1(x))
        x = self.swish(self.nn2(x))
        x = self.swish(self.nn3(x))
        x = self.swish(self.nn4(x))
        x = self.nn5(x)

        mean, logvar = x.chunk(2, dim=2)
        logvar = self.max_logvar - F.softplus(self.max_logvar - logvar)
        logvar = self.min_logvar + F.softplus(logvar - self.min_logvar)

        if ret_log_var:
            return mean, logvar
        else:
            return mean, torch.exp(logvar)

    def get_decay_loss(self):
        decay_loss = 0.
        for m in self.modules():
            if isinstance(m, EnsembleFC):
                decay_loss += m.weight_decay * torch.sum(torch.square(m.weight)) / 2.
        return decay_loss

    def loss(self, mean: torch.Tensor, logvar: torch.Tensor, labels: torch.Tensor):
        """
        mean, logvar: Ensemble_size x N x dim
        labels: Ensemble_size x N x dim
        """
        assert len(mean.shape) == len(logvar.shape) == len(labels.shape) == 3
        inv_var = torch.exp(-logvar)
        # Average over batch and dim, sum over ensembles.
        mse_loss_inv = (torch.pow(mean - labels, 2) * inv_var).mean(dim=(1, 2))
        var_loss = logvar.mean(dim=(1, 2))
        with torch.no_grad():
            # Used only for logging.
            mse_loss = torch.pow(mean - labels, 2).mean(dim=(1, 2))
        total_loss = mse_loss_inv.sum() + var_loss.sum()
        return total_loss, mse_loss

    def train(self, loss: torch.Tensor):
        self.optimizer.zero_grad()

        loss += 0.01 * torch.sum(self.max_logvar) - 0.01 * torch.sum(self.min_logvar)
        if self.use_decay:
            loss += self.get_decay_loss()

        loss.backward()

        self.optimizer.step()