-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.lua
262 lines (253 loc) · 12.2 KB
/
model.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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
--
-- Copyright (c) 2014, Facebook, Inc.
-- All rights reserved.
--
-- This source code is licensed under the BSD-style license found in the
-- LICENSE file in the root directory of this source tree. An additional grant
-- of patent rights can be found in the PATENTS file in the same directory.
--
require 'nn'
require 'cunn'
require 'cudnn'
local model_utils = require 'util.model_utils'
paths.dofile('trainUtil.lua')
require 'optim'
--[[
1. Create Model
2. Create Criterion
3. If preloading option is set, preload weights from existing models appropriately
4. Convert model to CUDA
]]--
-- 1. Create Network
--local config = opt.netType
local config = opt.netType .. '_' .. opt.backend
if opt.retrain == 'train' then
paths.dofile('models/' .. config .. '.lua')
end
-- local dbg = require('debugger')
-- 2. Create Criterion
if opt.lossType==0 then -- binary classification
criterion = nn.BCECriterion();
opt.labelOffset=-1;
elseif opt.lossType==1 then -- multi-class
criterion = nn.ClassNLLCriterion()
opt.labelOffset=0;
end
-- 3. If preloading option is set, preload weights from existing models appropriately
if opt.retrain == 'train' then
print('=> Creating model from file: models/' .. config .. '.lua')
-- train from scratch
if string.sub(opt.netType,1,2)=="sm" then
cutorch.setDevice(opt.GPUs[1])
model = createModel(opt.vnC,opt.nClasses)
if opt.netType=='smD' then
model.modules[1].p = opt.Mdropout[1]
end
else
model = createModel(opt.GPUs,opt.vnF*opt.vnC,opt.nClasses) -- for the model creation code, check the models/ folder
-- change dropout rate
if not string.match(opt.netType,'resnet') then --for vgg-fc-bn
if string.match(opt.netType,'F_ucfP') then --for vgg-fc-bn
model.modules[2].modules[5].p=opt.Mdropout[1];
model.modules[2].modules[9].p=opt.Mdropout[2];
end
if string.match(opt.netType,'F_nin') then --for vgg-fc-bn
model.modules[2].modules[4].p=opt.Mdropout[1];
model.modules[2].modules[8].p=opt.Mdropout[2];
end
end
end
if opt.wInit == 'kaiming' then
for indx,module in pairs(model:findModules('cudnn.SpatialConvolution')) do
module.weight:normal(0,math.sqrt(2/(module.kW*module.kH*module.nOutputPlane)))
end
elseif opt.wInit == 'xavier' then
for indx,module in pairs(model:findModules('cudnn.SpatialConvolution')) do
module.weight:normal(0,math.sqrt(1/(module.kW*module.kH*module.nOutputPlane)))
end
end
for indx,module in pairs(model:findModules('Linear')) do
modules.weight:copy(torch.randn(modules.weight:size()):mul(0.001))
modules.bias:fill(0) -- small uniform numbers
end
if opt.nEpochs ==0 then -- save the initial model
model:clearState()
saveDataParallel(paths.concat(opt.save, 'model_0.t7'), model) -- defined in trainUtil.lua
os.exit()
end
else
-- retrain
print('===> Loading model from file: ')
print(opt.retrain)
if opt.retrainOpt=='finetuneCBN' or opt.retrainOpt=='finetuneC' or opt.retrainOpt=='finetuneMD' or opt.retrainOpt=='finetuneMP' or opt.retrainOpt=='finetuneP' then
require 'loadcaffe'
model0 = loadcaffe.load(opt.retrain[1],opt.retrain[2],opt.retrain[3])
-- local model0 = loadcaffe.load(opt.retrain[1],opt.retrain[2],opt.retrain[3])
model_utils.caffeToTorch(model0,opt.retrainOpt,opt.Mdropout)
if string.match(opt.netType,'PT') then -- regular
model = createModel(opt.GPUs,opt.vnF,opt.vnC,nClasses,opt.timeMP)
else
model = createModel(opt.GPUs,opt.vnF*opt.vnC,nClasses)
end
-- copy from preload
local numP1
local offset=0;
local countBN=0;
if string.match(opt.netType, "pair2") then offset=-1;end
if #opt.GPUs==1 then
numP1 = #model.modules[1]['modules']
for j=1,numP1 do
if torch.type(model.modules[1]['modules'][j]):find('Convolution') then
model.modules[1]['modules'][j].weight:copy(model0.modules[j+offset-countBN].weight)
model.modules[1]['modules'][j].bias:copy(model0.modules[j+offset-countBN].bias)
if string.match(opt.netType,"bnC") then countBN = countBN+1;end
end
end
else
numP1 = #model.modules[1]['modules'][1]['modules']
for j=1,numP1 do
if torch.type(model.modules[1]['modules'][1]['modules'][j]):find('Convolution') then
--print(j,offset,numP1)
--print(#model.modules[1]['modules'][1]['modules'][j].weight)
model.modules[1]['modules'][1]['modules'][j].weight:copy(model0.modules[j+offset-countBN].weight)
model.modules[1]['modules'][1]['modules'][j].bias:copy(model0.modules[j+offset-countBN].bias)
if string.match(opt.netType,"bnC") then countBN = countBN+1;end
end
end
end
local numP2 = #model.modules[2]['modules']
if (not string.match(opt.netType, "pair2")) and (not string.match(opt.netType, "nin")) then
for j=1,numP2 do
if torch.type(model.modules[2]['modules'][j]):find('Linear') then
model.modules[2]['modules'][j].weight:copy(model0.modules[numP1+j].weight)
model.modules[2]['modules'][j].bias:copy(model0.modules[numP1+j].bias)
elseif torch.type(model.modules[2]['modules'][j]):find('Dropout') then --dropout
model.modules[2]['modules'][j].p=model0.modules[numP1+j].p
end
end
end
elseif opt.retrainOpt=='test' or opt.retrainOpt=='finetune' or opt.retrainOpt=='finetuneL' or opt.retrainOpt=='finetuneM' or opt.retrainOpt == 'trainL' or opt.retrainOpt == 'trainA' then
-- finetune
print('=> finetune or test from previous model: ' .. opt.netType)
if type(opt.retrain)=="table" then
require 'loadcaffe'
-- finetune from caffe
model = loadcaffe.load(opt.retrain[1],opt.retrain[2],opt.retrain[3])
-- local dbg = require('debugger');dbg()
-- local parameters, gradParameters = model:getParameters()
-- modify last layer ...
model_utils.caffeToTorch(model,opt.retrainOpt,opt.Mdropout)
else
-- finetune from t7
assert(paths.filep(opt.retrain), 'File not found: ' .. opt.retrain)
model = loadDataParallel(opt.retrain,#opt.GPUs)
-- change dropout rate
if string.match(opt.netType,'F_ucfP') then --for vgg-fc-bn
model.modules[2].modules[5].p=opt.Mdropout[1];
model.modules[2].modules[9].p=opt.Mdropout[2];
end
end
end
end
-- only last layer
if opt.retrainOpt == 'trainA' or opt.retrainOpt == 'trainL' or opt.retrainOpt == 'finetuneL' then
print('=> only retrain last layer: ' .. opt.netType)
local last_id = 22; -- alexnet
if string.match(opt.netType, "vggbnCF_nin2pair2") then
--AoT need to change last layer
if opt.retrainOpt ~= 'trainA' then
-- first conv
for j=1,#model.modules[1].modules[1].modules do
if torch.type(model.modules[1].modules[1].modules[j]):find('Linear') or torch.type(model.modules[1].modules[1].modules[j]):find('Convolution') then
print('Freeze conv-layers: '..j..','..torch.type(model.modules[1].modules[1].modules[j]))
model.modules[1].modules[1].modules[j].accGradParameters = function() end
model.modules[1].modules[1].modules[j].updateParameters = function() end
end
end
-- fc
if opt.netType ~= "vggbnCF_nin2pair2fc" then
for j=1,#model.modules[2].modules do
if torch.type(model.modules[2].modules[j]):find('Linear') or torch.type(model.modules[2].modules[j]):find('Convolution') then
print('Freeze fc-layers: '..j..','..torch.type(model.modules[2].modules[j]))
model.modules[2].modules[j].accGradParameters = function() end
model.modules[2].modules[j].updateParameters = function() end
end
end
end
end
-- first time, change the sturcture
-- should move to model_utils
if string.match(opt.retrainOpt , 'train') then
if opt.netType == "vggbnCF_nin2pair2" or opt.netType == "vggbnCF_nin2pair2fc" then
-- dropout
model.modules[2].modules[5].p = opt.Mdropout[1]
model.modules[2].modules[9].p = opt.Mdropout[2]
print('change dropout: ',opt.Mdropout)
model.modules[2].modules[12]=nn.Linear(1024, opt.nClasses)
model.modules[2].modules[12].weight:copy(torch.randn(model.modules[2].modules[12].weight:size()):mul(0.001))
model.modules[2].modules[12].bias:fill(0) -- small uniform numbers
model.modules[2].modules[13]=nn.LogSoftMax()
elseif opt.netType == "vggbnCF_nin2pair2NP" then
model.modules[2].modules[10]=nn.View(-1,200704)
model.modules[2].modules[11]=nn.Linear(200704, opt.nClasses)
model.modules[2].modules[11].weight:copy(torch.randn(model.modules[2].modules[11].weight:size()):mul(0.001))
model.modules[2].modules[11].bias:fill(0) -- small uniform numbers
model.modules[2].modules[12]=nn.LogSoftMax()
model.modules[2].modules[13]=nil
elseif opt.netType == "resnetbnCF_nin2pair2" then
model.modules[2].modules[5]=nn.Linear(2048, opt.nClasses)
model.modules[2].modules[5].weight:copy(torch.randn(model.modules[2].modules[5].weight:size()):mul(0.001))
model.modules[2].modules[5].bias:fill(0) -- small uniform numbers
model.modules[2].modules[6]=nn.LogSoftMax()
end
end
else -- no need to change structure
if #opt.GPUs==1 then -- for imNet
if opt.retrainOpt ~= 'trainA' then
if string.match(opt.netType, "vgg_ucfP") then
last_id=38
if opt.netType=="vgg_ucfPp5" or opt.netType=="vgg_ucfPfc" then
last_id=30
end
end
print('Freeze layers: '..opt.netType..','..last_id)
for j=1,last_id do
if torch.type(model.modules[j]):find('Linear') or torch.type(model.modules[j]):find('Convolution') then
model.modules[j].accGradParameters = function() end
model.modules[j].updateParameters = function() end
end
end
end
else -- for rand
if opt.retrainOpt ~= 'trainA' then
for j=1,#model.modules[1].modules[1].modules do
if torch.type(model.modules[1].modules[1].modules[j]):find('Linear') or torch.type(model.modules[1].modules[1].modules[j]):find('Convolution') then
model.modules[1].modules[1].modules[j].accGradParameters = function() end
model.modules[1].modules[1].modules[j].updateParameters = function() end
end
end
-- fc
if not string.match(opt.netType, "ucfPfc") then
for j=1,#model.modules[2].modules-2 do
if torch.type(model.modules[2].modules[j]):find('Linear') or torch.type(model.modules[2].modules[j]):find('Convolution') then
model.modules[2].modules[j].accGradParameters = function() end
model.modules[2].modules[j].updateParameters = function() end
end
end
end
end
end
end
end
if opt.backend == 'cudnn' then
model = model:cuda()
cudnn.convert(model, cudnn)
elseif opt.backend == 'cunn' then
model = model:cuda()
elseif opt.backend ~= 'nn' then
error'Unsupported backend'
end
if opt.backend ~= 'nn' then
criterion:cuda()
end
collectgarbage()