File size: 9,414 Bytes
d4ab5ac
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Parts of this file have been adapted from
https://github.com/nicola-decao/diffmask/blob/master/diffmask/models/gates.py
"""

import torch
import torch.nn as nn

from torch import Tensor
from typing import Optional
from utils.distributions import RectifiedStreched, BinaryConcrete


class MLPGate(nn.Module):
    def __init__(self, input_size: int, hidden_size: int, bias: bool = True):
        """
        This is an MLP with the following structure;
        Linear(input_size, hidden_size), Tanh(), Linear(hidden_size, 1)
        The bias of the last layer is set to 5.0 to start with high probability
        of keeping states (fundamental for good convergence as the initialized
        DiffMask has not learned what to mask yet).

        Args:
            input_size (int): the number of input features
            hidden_size (int): the number of hidden units
            bias (bool): whether to use a bias term
        """
        super().__init__()

        self.f = nn.Sequential(
            nn.utils.weight_norm(nn.Linear(input_size, hidden_size)),
            nn.Tanh(),
            nn.utils.weight_norm(nn.Linear(hidden_size, 1, bias=bias)),
        )

        if bias:
            self.f[-1].bias.data[:] = 5.0

    def forward(self, *args: Tensor) -> Tensor:
        return self.f(torch.cat(args, -1))


class MLPMaxGate(nn.Module):
    def __init__(
        self,
        input_size: int,
        hidden_size: int,
        mul_activation: float = 10.0,
        add_activation: float = 5.0,
        bias: bool = True,
    ):
        """
        This is an MLP with the following structure;
        Linear(input_size, hidden_size), Tanh(), Linear(hidden_size, 1)
        The bias of the last layer is set to 5.0 to start with high probability
        of keeping states (fundamental for good convergence as the initialized
        DiffMask has not learned what to mask yet).
        It also uses a scaler for the output of the activation function.

        Args:
            input_size (int): the number of input features
            hidden_size (int): the number of hidden units
            mul_activation (float): the scaler for the output of the activation function
            add_activation (float): the offset for the output of the activation function
            bias (bool): whether to use a bias term
        """
        super().__init__()

        self.f = nn.Sequential(
            nn.utils.weight_norm(nn.Linear(input_size, hidden_size)),
            nn.Tanh(),
            nn.utils.weight_norm(nn.Linear(hidden_size, 1, bias=bias)),
            nn.Tanh(),
        )
        self.add_activation = nn.Parameter(torch.tensor(add_activation))
        self.mul_activation = mul_activation

    def forward(self, *args: Tensor) -> Tensor:
        return self.f(torch.cat(args, -1)) * self.mul_activation + self.add_activation


class DiffMaskGateInput(nn.Module):
    def __init__(
        self,
        hidden_size: int,
        hidden_attention: int,
        num_hidden_layers: int,
        max_position_embeddings: int,
        gate_fn: nn.Module = MLPMaxGate,
        mul_activation: float = 10.0,
        add_activation: float = 5.0,
        gate_bias: bool = True,
        placeholder: bool = False,
        init_vector: Tensor = None,
    ):
        """This is a DiffMask module that masks the input of the first layer.

        Args:
            hidden_size (int): the size of the hidden representations
            hidden_attention (int) the amount of units in the gate's hidden (bottleneck) layer
            num_hidden_layers (int): the number of hidden layers (and thus gates to use)
            max_position_embeddings (int): the amount of placeholder embeddings to learn for the masked positions
            gate_fn (nn.Module): the PyTorch module to use as a gate
            mul_activation (float): the scaler for the output of the activation function
            add_activation (float): the offset for the output of the activation function
            gate_bias (bool): whether to use a bias term
            placeholder (bool): whether to use placeholder embeddings or a zero vector
            init_vector (Tensor): the initial vector to use for the placeholder embeddings
        """
        super().__init__()

        # Create a ModuleList with the gates
        self.g_hat = nn.ModuleList(
            [
                gate_fn(
                    hidden_size * 2,
                    hidden_attention,
                    mul_activation,
                    add_activation,
                    gate_bias,
                )
                for _ in range(num_hidden_layers)
            ]
        )

        if placeholder:
            # Use a placeholder embedding for the masked positions
            self.placeholder = nn.Parameter(
                nn.init.xavier_normal_(
                    torch.empty((1, max_position_embeddings, hidden_size))
                )
                if init_vector is None
                else init_vector.view(1, 1, hidden_size).repeat(
                    1, max_position_embeddings, 1
                )
            )
        else:
            # Use a zero vector for the masked positions
            self.register_buffer(
                "placeholder",
                torch.zeros((1, 1, hidden_size)),
            )

    def forward(
        self, hidden_states: tuple[Tensor], layer_pred: Optional[int]
    ) -> tuple[tuple[Tensor], Tensor, Tensor, Tensor, Tensor]:
        # Concatenate the output of all the gates
        logits = torch.cat(
            [
                self.g_hat[i](hidden_states[0], hidden_states[i])
                for i in range(
                    (layer_pred + 1) if layer_pred is not None else len(hidden_states)
                )
            ],
            -1,
        )

        # Define a Hard Concrete distribution
        dist = RectifiedStreched(
            BinaryConcrete(torch.full_like(logits, 0.2), logits),
            l=-0.2,
            r=1.0,
        )

        # Calculate the expectation for the full gate probabilities
        # These act as votes for the masked positions
        gates_full = dist.rsample().cumprod(-1)
        expected_L0_full = dist.log_expected_L0().cumsum(-1)

        # Extract the probabilities from the last layer, which acts
        # as an aggregation of the votes per position
        gates = gates_full[..., -1]
        expected_L0 = expected_L0_full[..., -1]

        return (
            hidden_states[0] * gates.unsqueeze(-1)
            + self.placeholder[:, : hidden_states[0].shape[-2]]
            * (1 - gates).unsqueeze(-1),
            gates,
            expected_L0,
            gates_full,
            expected_L0_full,
        )


# class DiffMaskGateHidden(nn.Module):
#     def __init__(
#         self,
#         hidden_size: int,
#         hidden_attention: int,
#         num_hidden_layers: int,
#         max_position_embeddings: int,
#         gate_fn: nn.Module = MLPMaxGate,
#         gate_bias: bool = True,
#         placeholder: bool = False,
#         init_vector: Tensor = None,
#     ):
#         super().__init__()
#
#         self.g_hat = nn.ModuleList(
#             [
#                 gate_fn(hidden_size, hidden_attention, bias=gate_bias)
#                 for _ in range(num_hidden_layers)
#             ]
#         )
#
#         if placeholder:
#             self.placeholder = nn.ParameterList(
#                 [
#                     nn.Parameter(
#                         nn.init.xavier_normal_(
#                             torch.empty((1, max_position_embeddings, hidden_size))
#                         )
#                         if init_vector is None
#                         else init_vector.view(1, 1, hidden_size).repeat(
#                             1, max_position_embeddings, 1
#                         )
#                     )
#                     for _ in range(num_hidden_layers)
#                 ]
#             )
#         else:
#             self.register_buffer(
#                 "placeholder",
#                 torch.zeros((num_hidden_layers, 1, 1, hidden_size)),
#             )
#
#     def forward(
#         self, hidden_states: tuple[Tensor], layer_pred: Optional[int]
#     ) -> tuple[tuple[Tensor], Tensor, Tensor, Tensor, Tensor]:
#         if layer_pred is not None:
#             logits = self.g_hat[layer_pred](hidden_states[layer_pred])
#         else:
#             logits = torch.cat(
#                 [self.g_hat[i](hidden_states[i]) for i in range(len(hidden_states))], -1
#             )
#
#         dist = RectifiedStreched(
#             BinaryConcrete(torch.full_like(logits, 0.2), logits),
#             l=-0.2,
#             r=1.0,
#         )
#
#         gates_full = dist.rsample()
#         expected_L0_full = dist.log_expected_L0()
#
#         gates = gates_full if layer_pred is not None else gates_full[..., :1]
#         expected_L0 = (
#             expected_L0_full if layer_pred is not None else expected_L0_full[..., :1]
#         )
#
#         layer_pred = layer_pred or 0  # equiv to "layer_pred if layer_pred else 0"
#         return (
#             hidden_states[layer_pred] * gates
#             + self.placeholder[layer_pred][:, : hidden_states[layer_pred].shape[-2]]
#             * (1 - gates),
#             gates.squeeze(-1),
#             expected_L0.squeeze(-1),
#             gates_full,
#             expected_L0_full,
#         )