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
|
local MatrixParam = nerv.class('nerv.MatrixParam', 'nerv.Param')
local LinearTransParam = nerv.class('nerv.LinearTransParam', 'nerv.MatrixParam')
local BiasParam = nerv.class('nerv.BiasParam', 'nerv.MatrixParam')
local AffineLayer = nerv.class('nerv.AffineLayer', 'nerv.Layer')
function MatrixParam:read(handle)
self.trans = self.gconf.cumat_type.new_from_host(
nerv.MMatrixFloat.load(handle))
end
function MatrixParam:write(handle)
self.trans:new_to_host():save(handle)
end
function MatrixParam:train_init()
self.correction = self.trans:create()
self.correction:fill(0)
end
function MatrixParam:update(gradient)
local gconf = self.gconf
if gconf.momentum > 0 then
self.correction:add(self.correction, gradient, gconf.momentum, 1.0)
-- momentum gain
local mmt_gain = 1.0 / (1.0 - gconf.momentum)
local n = self.gconf.batch_size * mmt_gain
-- perform update
self.trans:add(self.trans, self.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
else
self.trans:add(self.trans, gradient, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / gconf.batch_size)
end
end
function MatrixParam:updateEI(err, input)
local gconf = self.gconf
if gconf.momentum > 0 then
self.correction:mul(input, err, 1.0, gconf.momentum, 'T', 'N')
-- momentum gain
local mmt_gain = 1.0 / (1.0 - gconf.momentum)
local n = self.gconf.batch_size * mmt_gain
-- perform update
self.trans:add(self.trans, self.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
else
self.trans:mul(input, err, - gconf.lrate / gconf.batch_size, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, 'T', 'N')
end
end
--[[ --these updates are the same
function LinearTransParam:update(gradient)
MatrixParam.update(self, gradient)
-- local gconf = self.gconf
-- weight decay(put into MatrixParam:update)
-- self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size)
end
function BiasParam:update(gradient)
MatrixParam.update(self, gradient)
--local gconf = self.gconf
-- weight decay
-- self.trans:add(self.trans, self.trans, 1.0, -gconf.lrate * gconf.wcost / gconf.batch_size)
end
]]--
function AffineLayer:__init(id, global_conf, layer_conf)
self.id = id
self.ltp = layer_conf.ltp
self.bp = layer_conf.bp
self.dim_in = layer_conf.dim_in
self.dim_out = layer_conf.dim_out
self.gconf = global_conf
self:check_dim_len(1, 1) -- exactly one input and one output
self.direct_update = layer_conf.direct_update or global_conf.direct_update
end
function AffineLayer:init(batch_size)
if self.ltp.trans:ncol() ~= self.bp.trans:ncol() then
nerv.error("mismatching dimensions of linear transform and bias paramter")
end
if self.dim_in[1] ~= self.ltp.trans:nrow() then
nerv.error("mismatching dimensions of linear transform parameter and input")
end
if self.dim_out[1] ~= self.ltp.trans:ncol() then
nerv.error("mismatching dimensions of linear transform parameter and output")
end
self.ltp_grad = self.ltp.trans:create()
self.ltp:train_init()
self.bp:train_init()
end
function AffineLayer:batch_resize(batch_size)
-- do nothing
end
function AffineLayer:update(bp_err, input, output)
if self.direct_update == true then
local gconf = self.gconf
if gconf.momentum > 0 then
self.ltp.correction:mul(input[1], bp_err[1], 1.0, gconf.momentum, 'T', 'N')
self.bp.correction:add(self.bp.correction, bp_err[1]:colsum(), gconf.momentum, 1)
-- momentum gain
local mmt_gain = 1.0 / (1.0 - gconf.momentum)
local n = self.gconf.batch_size * mmt_gain
-- perform update
self.ltp.trans:add(self.ltp.trans, self.ltp.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
self.bp.trans:add(self.bp.trans, self.bp.correction, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / n)
else
self.ltp.trans:mul(input[1], bp_err[1], - gconf.lrate / gconf.batch_size, 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, 'T', 'N')
self.bp.trans:add(self.bp.trans, bp_err[1]:colsum(), 1.0 - gconf.lrate * gconf.wcost / gconf.batch_size, - gconf.lrate / gconf.batch_size)
end
else
self.ltp:updateEI(bp_err[1], input[1])
self.bp:update(bp_err[1]:colsum())
end
end
function AffineLayer:propagate(input, output)
-- apply linear transform
output[1]:mul(input[1], self.ltp.trans, 1.0, 0.0, 'N', 'N')
-- add bias
output[1]:add_row(self.bp.trans, 1.0)
end
function AffineLayer:back_propagate(bp_err, next_bp_err, input, output)
next_bp_err[1]:mul(bp_err[1], self.ltp.trans, 1.0, 0.0, 'N', 'T')
end
function AffineLayer:get_params()
return nerv.ParamRepo({self.ltp, self.bp})
end
|