forked from harvardnlp/seq2seq-attn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
prune.lua
123 lines (106 loc) · 3.16 KB
/
prune.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
require 'nn'
require 'string'
require 'nngraph'
require 's2sa.models'
require 's2sa.data'
require 's2sa.plinear'
local cmd = torch.CmdLine()
-- file location
cmd:option('-model', 'model.t7','model file')
cmd:option('-savefile', 'model-pruned.t7','pruned model file')
cmd:option('-gpuid', -1, [[Which gpu to use. -1 = use CPU]])
cmd:option('-ratio', 0.2, [[ratio of the parameters to prune]])
cmd:option('-prune', 'blind', [[type of pruning strategy: (class) blind, uniform]])
local opt = cmd:parse(arg)
local allparams=torch.Tensor()
-- find n-th percentile of a tensor without sorting the tensor
local function getKth(k, t)
-- define 1000 buckets to avoid complete sorting for interval i/1000;i+1/1000
local buckets={}
for _=1,1000 do
table.insert(buckets,{})
end
-- fill the bucket with the idx
for i=1,t:size(1) do
local nbucket=math.floor(t[i]*1000)
if nbucket>999 then nbucket=999 end
table.insert(buckets[nbucket+1],i)
end
-- find the bucket of interest
local idx=1
while k>0 do
k=k-#buckets[idx]
idx=idx+1
end
-- sort the bucket
k=k+#buckets[idx-1]
local sbucket=torch.Tensor(#buckets[idx-1])
for i=1,#buckets[idx-1] do
sbucket[i]=t[buckets[idx-1][i]]
end
return torch.sort(sbucket)[k]
end
-- count total number of parameters
local function countParameters(m)
local classname=torch.typename(m)
if classname=='nn.Linear' or classname=='nn.LinearNoBias' then
local p=m:getParameters()
if allparams:dim()==0 then
allparams=p
else
allparams=torch.cat(allparams, p)
end
end
end
local function prune(m, gthreshold, locopt)
local p=m:getParameters()
local classname=torch.typename(m)
if classname=='nn.Linear' or classname=='nn.LinearNoBias' then
local lthreshold=getKth(math.floor(locopt.ratio*p:size(1))+1,torch.abs(p))
local threshold=gthreshold
if locopt.prune == 'uniform' then
threshold=lthreshold
end
local pruned,total=m:prune(threshold)
print(m.name,pruned,total)
end
end
local function main()
if not(opt.prune=='blind' or opt.prune=='uniform') then
print('ERROR - prune type should be "blind" or "uniform"')
return
end
if opt.gpuid >= 0 then
print('using CUDA on GPU ' .. opt.gpuid .. '...')
require 'cutorch'
require 'cunn'
cutorch.setDevice(opt.gpuid)
end
print('loading model ' .. opt.model)
local checkpoint = torch.load(opt.model)
local model, model_opt = checkpoint[1], checkpoint[2]
if model_opt.prune ~= nil then
print('cannot prune pruned model')
return
end
-- count all parameters
for i = 1, #model do
if model[i].apply then
model[i]:apply(countParameters)
end
end
-- sort by magnitude
local gthreshold=getKth(math.floor(opt.ratio*allparams:size(1))+1,torch.abs(allparams))
print('#parameters:', allparams:size(1), 'global threshold:',gthreshold, 'prune method:',opt.prune)
-- apply pruning
for i = 1, #model do
model[i]:apply(function(m) prune(m, gthreshold, opt) end)
end
if model_opt.pruning == nil then
model_opt.pruning = {}
end
table.insert(model_opt.pruning, opt)
print('saving model to ' .. opt.savefile)
torch.save(opt.savefile, {model, model_opt})
end
main()