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
|
local SoftmaxCELayer = nerv.class("nerv.SoftmaxCELayer", "nerv.Layer")
function SoftmaxCELayer:__init(id, global_conf, layer_conf)
self.id = id
self.gconf = global_conf
if self.gconf.use_cpu then
self.mat_type = self.gconf.mmat_type
else
self.mat_type = self.gconf.cumat_type
end
self.dim_in = layer_conf.dim_in
self.dim_out = layer_conf.dim_out
self.compressed = layer_conf.compressed
if self.compressed == nil then
self.compressed = false
end
self:check_dim_len(2, -1) -- two inputs: nn output and label
end
function SoftmaxCELayer:init(batch_size, chunk_size)
if not self.compressed and (self.dim_in[1] ~= self.dim_in[2]) then
nerv.error("mismatching dimensions of previous network output and labels")
end
if chunk_size == nil then
chunk_size = 1
end
self.total_ce = 0.0
self.total_correct = 0
self.total_frames = 0
self.softmax = {}
self.ce = {}
for t = 1, chunk_size do
self.softmax[t] = self.mat_type(batch_size, self.dim_in[1])
self.ce[t] = self.softmax[t]:create()
end
end
function SoftmaxCELayer:batch_resize(batch_size, chunk_size)
if chunk_size == nil then
chunk_size = 1
end
for t = 1, chunk_size do
if self.softmax[t]:nrow() ~= batch_size then
self.softmax[t] = self.mat_type(batch_size, self.dim_in[1])
self.ce[t] = self.softmax:create()
end
end
end
function SoftmaxCELayer:update(bp_err, input, output, t)
-- no params, therefore do nothing
end
function SoftmaxCELayer:propagate(input, output, t)
if t == nil then
t = 1
end
local softmax = self.softmax[t]
local ce = self.ce[t]
local classified = softmax:softmax(input[1])
local label = input[2]
ce:log_elem(softmax)
if self.compressed then
label = label:decompress(input[1]:ncol())
end
ce:mul_elem(ce, label)
ce = ce:rowsum()
if output[1] ~= nil then
output[1]:copy_from(ce)
end
-- add total ce
self.total_ce = self.total_ce - ce:colsum()[0][0]
self.total_frames = self.total_frames + softmax:nrow()
-- TODO: add colsame for uncompressed label
if self.compressed then
self.total_correct = self.total_correct + classified:colsame(input[2])[0][0]
end
end
function SoftmaxCELayer:back_propagate(bp_err, next_bp_err, input, output, t)
-- softmax output - label
if t == nil then
t = 1
end
local label = input[2]
if self.compressed then
label = label:decompress(input[1]:ncol())
end
local nbe = next_bp_err[1]
nbe:add(self.softmax[t], label, 1.0, -1.0)
if bp_err[1] ~= nil then
nbe:scale_rows_by_col(bp_err[1])
end
end
function SoftmaxCELayer:get_params()
return nerv.ParamRepo({})
end
|