-
Notifications
You must be signed in to change notification settings - Fork 8
/
init.lua
62 lines (51 loc) · 1.84 KB
/
init.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
require 'nn'
local ffi = require 'ffi'
caffegraph = {}
ffi.cdef[[
struct params { int num_params; THFloatTensor** params; };
void loadModel(void** handle, const char* prototxt, const char* caffemodel);
void buildModel(void** handle, const char* lua_path);
void getParams(void** handle, THFloatTensor*** params);
void freeModel(void** handle);
]]
caffegraph.C = ffi.load(package.searchpath('libcaffegraph', package.cpath))
caffegraph.load = function(prototxt, caffemodel)
local handle = ffi.new('void*[1]')
-- load the caffemodel into a graph structure
local initHandle = handle[1]
caffegraph.C.loadModel(handle, prototxt, caffemodel)
if handle[1] == initHandle then
error('Unable to load model.')
end
-- serialize the graph and write it out
local luaModel = path.splitext(caffemodel)
luaModel = luaModel..'.lua'
caffegraph.C.buildModel(handle, luaModel)
-- -- bring the model into lua world
local model, modmap = dofile(luaModel)
-- transfer the parameters
local noData = torch.FloatTensor():zero():cdata()
local module_params = {}
for i,nodes in ipairs(modmap) do
local params = {}
for i=1,#nodes do
local module = nodes[i].data.module
module:float()
if torch.isTypeOf(module, nn.BatchNormalization) then
module.weight:fill(1)
module.bias:zero()
params[(i-1)*2+1] = module.running_mean:cdata()
params[i*2] = module.running_var:cdata()
else
params[(i-1)*2+1] = module.weight and module.weight:cdata() or noData
params[i*2] = module.bias and module.bias:cdata() or noData
end
end
module_params[i] = ffi.new('THFloatTensor*['..#params..']', params)
end
local cParams = ffi.new('THFloatTensor**['..#module_params..']', module_params)
caffegraph.C.getParams(handle, cParams)
caffegraph.C.freeModel(handle)
return model
end
return caffegraph