-
Notifications
You must be signed in to change notification settings - Fork 12
/
grad_check_seq.lua
219 lines (187 loc) · 6.68 KB
/
grad_check_seq.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
require 'nn'
require 'cunn'
require 'cudnn'
require 'cutorch'
require 'misc.ProcNets_bilstm_seq_guide'
local gradcheck = require 'misc.gradcheck'
local tests = {}
local tester = torch.Tester()
-- cutorch.manualSeed(123)
-- cutorch.setDevice(1)
-------------------------------------------------------------------------------
-- gradient check for the ts model
-------------------------------------------------------------------------------
local function gradCheckTS()
local dtype = 'torch.DoubleTensor'
local lmOpt = {}
lmOpt.input_encoding_size = 8
lmOpt.rnn_size = 9
lmOpt.dropout = 0
lmOpt.frames_per_video = 20 -- number of unrolled LSTMs
lmOpt.batch_size = 1
lmOpt.KTL = 3
lmOpt.KTU = 9
lmOpt.KTS = 2
lmOpt.CN = 10
local lm = nn.ProcNets(lmOpt)
local crit = nn.ProcNetsCriterion({lmOpt.KTL, lmOpt.KTU, lmOpt.KTS})
lm:type(dtype)
crit:type(dtype)
local ts = lm.ts
-- imgs = torch.randn(lmOpt.KTL, lmOpt.frames_per_video):type(dtype)
-- imgs = torch.Tensor(imgs_tmp:size(2),imgs_tmp:size(1)):copy(imgs_tmp:transpose(1,2))
-- imgs = torch.randn(lmOpt.CN,2):type(dtype)
imgs = torch.randn(lmOpt.frames_per_video,lmOpt.rnn_size):type(dtype)
-- evaluate the analytic gradient
local output = ts:forward(imgs)
local w = torch.randn(output[1]:size())
local w2 = torch.randn(output[2]:size())
-- generate random weighted sum criterion
local loss = torch.sum(torch.cmul(output[1], w))+torch.sum(torch.cmul(output[2], w2))
local gradOutput = {w, w2}
local gradInput = ts:backward(imgs, gradOutput)
-- create a loss function wrapper
local function f(x)
local output = ts:forward(x)
local loss = torch.sum(torch.cmul(output[1], w))+torch.sum(torch.cmul(output[2], w2))
return loss
end
local gradInput_num = gradcheck.numeric_gradient(f, imgs, 1, 1e-6)
--[[
print(gradInput)
print(gradInput_num)
local g = gradInput:view(-1)
local gn = gradInput_num:view(-1)
for i=1,g:nElement() do
local r = gradcheck.relative_error(g[i],gn[i])
print(i, g[i], gn[i], r)
end
]]--
tester:assertTensorEq(gradInput, gradInput_num, 1e-4)
tester:assertlt(gradcheck.relative_error(gradInput, gradInput_num, 1e-8), 1e-4)
end
-------------------------------------------------------------------------------
-- gradient check for the language model
-------------------------------------------------------------------------------
-- test just the language model alone (without the criterion)
local function gradCheckLM()
local dtype = 'torch.DoubleTensor'
local lmOpt = {}
lmOpt.input_encoding_size = 8
lmOpt.rnn_size = 15
lmOpt.dropout = 0
lmOpt.frames_per_video = 20 -- number of unrolled LSTMs
lmOpt.batch_size = 1
lmOpt.KTL = 3
lmOpt.KTU = 9
lmOpt.KTS = 2
lmOpt.CN = 5
lmOpt.mp_scale_h = 4
lmOpt.mp_scale_w = 4
lmOpt.gradcheck = true
local lm = nn.ProcNets(lmOpt)
local crit = nn.ProcNetsCriterion({lmOpt.KTL, lmOpt.KTU, lmOpt.KTS})
lm:type(dtype)
crit:type(dtype)
--[[
local seq = torch.LongTensor(lmOpt.seq_length, lmOpt.batch_size):random(lmOpt.vocab_size)
seq[{ {4, 7}, 1 }] = 0
seq[{ {5, 7}, 4 }] = 0
local imgs = torch.randn(lmOpt.batch_size, lmOpt.input_encoding_size):type(dtype)
]]--
imgs = torch.randn(lmOpt.frames_per_video,lmOpt.input_encoding_size):type(dtype)
segs = {{1,3},{5,8},{10,11},{13,17},{19,20}}
-- evaluate the analytic gradient
local output = lm:forward({imgs,segs})
local w = torch.randn(output[1]:size())
local w2 = torch.randn(output[2]:size())
local w3 = torch.randn(output[3]:size())
local w4 = torch.randn(output[4]:size())
-- generate random weighted sum criterion
local loss = torch.sum(torch.cmul(output[1], w)) + torch.sum(torch.cmul(output[2], w2)) + torch.sum(torch.cmul(output[3], w3)) + torch.sum(torch.cmul(output[4], w4))
local gradOutput = {w, w2, w3, w4}
local gradInput_t = lm:backward({imgs,segs}, gradOutput)
local gradInput = gradInput_t[1]
-- create a loss function wrapper
local function f(x)
local output = lm:forward({x,segs})
local loss = torch.sum(torch.cmul(output[1], w)) + torch.sum(torch.cmul(output[2], w2)) + torch.sum(torch.cmul(output[3], w3)) + torch.sum(torch.cmul(output[4], w4))
return loss
end
local gradInput_num = gradcheck.numeric_gradient(f, imgs, 1, 1e-6)
--[[
print(gradInput)
print(gradInput_num)
local g = gradInput:view(-1)
local gn = gradInput_num:view(-1)
for i=1,g:nElement() do
local r = gradcheck.relative_error(g[i],gn[i])
print(i, g[i], gn[i], r)
end
]]--
tester:assertTensorEq(gradInput, gradInput_num, 1e-4)
tester:assertlt(gradcheck.relative_error(gradInput, gradInput_num, 1e-8), 1e-4)
end
-------------------------------------------------------------------------------
-- gradient check
-------------------------------------------------------------------------------
-- test just the language model alone (without the criterion)
local function gradCheck()
local dtype = 'torch.DoubleTensor'
local lmOpt = {}
lmOpt.input_encoding_size = 8
lmOpt.vocab_size = 6
lmOpt.rnn_size = 15
lmOpt.dropout = 0
lmOpt.frames_per_video = 20 -- number of unrolled LSTMs
lmOpt.batch_size = 1
lmOpt.KTL = 3
lmOpt.KTU = 9
lmOpt.KTS = 2
lmOpt.CN = 5
lmOpt.mp_scale_h = 4
lmOpt.mp_scale_w = 4
lmOpt.gradcheck = true
local lm = nn.ProcNets(lmOpt)
local crit = nn.ProcNetsCriterion({lmOpt.KTL, lmOpt.KTU, lmOpt.KTS})
lm:type(dtype)
crit:type(dtype)
--[[
local seq = torch.LongTensor(lmOpt.seq_length, lmOpt.batch_size):random(lmOpt.vocab_size)
seq[{ {4, 7}, 1 }] = 0
seq[{ {5, 7}, 4 }] = 0
local imgs = torch.randn(lmOpt.batch_size, lmOpt.input_encoding_size):type(dtype)
]]--
imgs = torch.randn(lmOpt.frames_per_video,lmOpt.input_encoding_size):type(dtype)
label = {{1,3},{5,8},{10,11},{13,17},{19,20}}
-- evaluate the analytic gradient
local output = lm:forward({imgs,label})
local loss = crit:forward(output,label)
local gradOutput = crit:backward(output,label)
local gradInput_t = lm:backward({imgs,label}, gradOutput)
local gradInput = gradInput_t[1]
-- create a loss function wrapper
local function f(x)
local output = lm:forward({x,label})
local loss = crit:forward(output,label)
return loss
end
local gradInput_num = gradcheck.numeric_gradient(f, imgs, 1, 1e-6)
--[[
print(gradInput)
print(gradInput_num)
local g = gradInput:view(-1)
local gn = gradInput_num:view(-1)
for i=1,g:nElement() do
local r = gradcheck.relative_error(g[i],gn[i])
print(i, g[i], gn[i], r)
end
--]]
tester:assertTensorEq(gradInput, gradInput_num, 1e-4)
tester:assertlt(gradcheck.relative_error(gradInput, gradInput_num, 1e-8), 1e-4)
end
tests.gradCheckTS = gradCheckTS
tests.gradCheckLM = gradCheckLM
tests.gradCheck = gradCheck
tester:add(tests)
tester:run()