-
Notifications
You must be signed in to change notification settings - Fork 508
/
VanillaRNN.lua
223 lines (179 loc) · 5.62 KB
/
VanillaRNN.lua
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
require 'torch'
require 'nn'
local layer, parent = torch.class('nn.VanillaRNN', 'nn.Module')
--[[
Vanilla RNN with tanh nonlinearity that operates on entire sequences of data.
The RNN has an input dim of D, a hidden dim of H, operates over sequences of
length T and minibatches of size N.
On the forward pass we accept a table {h0, x} where:
- h0 is initial hidden states, of shape (N, H)
- x is input sequence, of shape (N, T, D)
The forward pass returns the hidden states at each timestep, of shape (N, T, H).
SequenceRNN_TN swaps the order of the time and minibatch dimensions; this is
very slightly faster, but probably not worth it since it is more irritating to
work with.
--]]
function layer:__init(input_dim, hidden_dim)
parent.__init(self)
local D, H = input_dim, hidden_dim
self.input_dim, self.hidden_dim = D, H
self.weight = torch.Tensor(D + H, H)
self.gradWeight = torch.Tensor(D + H, H)
self.bias = torch.Tensor(H)
self.gradBias = torch.Tensor(H)
self:reset()
self.h0 = torch.Tensor()
self.remember_states = false
self.buffer1 = torch.Tensor()
self.buffer2 = torch.Tensor()
self.grad_h0 = torch.Tensor()
self.grad_x = torch.Tensor()
self.gradInput = {self.grad_h0, self.grad_x}
end
function layer:reset(std)
if not std then
std = 1.0 / math.sqrt(self.hidden_dim + self.input_dim)
end
self.bias:zero()
self.weight:normal(0, std)
return self
end
function layer:resetStates()
self.h0 = self.h0.new()
end
function layer:_unpack_input(input)
local h0, x = nil, nil
if torch.type(input) == 'table' and #input == 2 then
h0, x = unpack(input)
elseif torch.isTensor(input) then
x = input
else
assert(false, 'invalid input')
end
return h0, x
end
local function check_dims(x, dims)
assert(x:dim() == #dims)
for i, d in ipairs(dims) do
assert(x:size(i) == d)
end
end
function layer:_get_sizes(input, gradOutput)
local h0, x = self:_unpack_input(input)
local N, T = x:size(1), x:size(2)
local H, D = self.hidden_dim, self.input_dim
check_dims(x, {N, T, D})
if h0 then
check_dims(h0, {N, H})
end
if gradOutput then
check_dims(gradOutput, {N, T, H})
end
return N, T, D, H
end
--[[
Input: Table of
- h0: Initial hidden state of shape (N, H)
- x: Sequence of inputs, of shape (N, T, D)
Output:
- h: Sequence of hidden states, of shape (N, T, H)
--]]
function layer:updateOutput(input)
self.recompute_backward = true
local h0, x = self:_unpack_input(input)
local N, T, D, H = self:_get_sizes(input)
self._return_grad_h0 = (h0 ~= nil)
if not h0 then
h0 = self.h0
if h0:nElement() == 0 or not self.remember_states then
h0:resize(N, H):zero()
elseif self.remember_states then
local prev_N, prev_T = self.output:size(1), self.output:size(2)
assert(prev_N == N, 'batch sizes must be constant to remember states')
h0:copy(self.output[{{}, prev_T}])
end
end
local bias_expand = self.bias:view(1, H):expand(N, H)
local Wx = self.weight[{{1, D}}]
local Wh = self.weight[{{D + 1, D + H}}]
self.output:resize(N, T, H):zero()
local prev_h = h0
for t = 1, T do
local cur_x = x[{{}, t}]
local next_h = self.output[{{}, t}]
next_h:addmm(bias_expand, cur_x, Wx)
next_h:addmm(prev_h, Wh)
next_h:tanh()
prev_h = next_h
end
return self.output
end
-- Normally we don't implement backward, and instead just implement
-- updateGradInput and accGradParameters. However for an RNN, separating these
-- two operations would result in quite a bit of repeated code and compute;
-- therefore we'll just implement backward and update gradInput and
-- gradients with respect to parameters at the same time.
function layer:backward(input, gradOutput, scale)
self.recompute_backward = false
scale = scale or 1.0
assert(scale == 1.0, 'scale must be 1')
local N, T, D, H = self:_get_sizes(input, gradOutput)
local h0, x = self:_unpack_input(input)
if not h0 then h0 = self.h0 end
local grad_h = gradOutput
local Wx = self.weight[{{1, D}}]
local Wh = self.weight[{{D + 1, D + H}}]
local grad_Wx = self.gradWeight[{{1, D}}]
local grad_Wh = self.gradWeight[{{D + 1, D + H}}]
local grad_b = self.gradBias
local grad_h0 = self.grad_h0:resizeAs(h0):zero()
local grad_x = self.grad_x:resizeAs(x):zero()
local grad_next_h = self.buffer1:resizeAs(h0):zero()
for t = T, 1, -1 do
local next_h, prev_h = self.output[{{}, t}], nil
if t == 1 then
prev_h = h0
else
prev_h = self.output[{{}, t - 1}]
end
grad_next_h:add(grad_h[{{}, t}])
local grad_a = grad_h0:resizeAs(h0)
grad_a:fill(1):addcmul(-1.0, next_h, next_h):cmul(grad_next_h)
grad_x[{{}, t}]:mm(grad_a, Wx:t())
grad_Wx:addmm(scale, x[{{}, t}]:t(), grad_a)
grad_Wh:addmm(scale, prev_h:t(), grad_a)
grad_next_h:mm(grad_a, Wh:t())
self.buffer2:resize(1, H):sum(grad_a, 1)
grad_b:add(scale, self.buffer2)
end
grad_h0:copy(grad_next_h)
if self._return_grad_h0 then
self.gradInput = {self.grad_h0, self.grad_x}
else
self.gradInput = self.grad_x
end
return self.gradInput
end
function layer:updateGradInput(input, gradOutput)
if self.recompute_backward then
self:backward(input, gradOutput, 1.0)
end
return self.gradInput
end
function layer:accGradParameters(input, gradOutput, scale)
if self.recompute_backward then
self:backward(input, gradOutput, scale)
end
end
function layer:clearState()
self.buffer1:set()
self.buffer2:set()
self.grad_h0:set()
self.grad_x:set()
self.output:set()
end
function layer:__tostring__()
local name = torch.type(self)
local din, dout = self.input_dim, self.hidden_dim
return string.format('%s(%d -> %d)', name, din, dout)
end