-
Notifications
You must be signed in to change notification settings - Fork 0
/
btplib.lua
225 lines (199 loc) · 7.41 KB
/
btplib.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
--------------------------------------
-- Optim Library, courtesy: A.Karpathy
--------------------------------------
function sgd(x, dx, lr)
x:add(-lr, dx)
end
function sgdm(x, dx, lr, alpha, state)
-- sgd with momentum, standard update
if not state.v then
state.v = x.new(#x):zero()
end
state.v:mul(alpha)
state.v:add(lr, dx)
x:add(-1, state.v)
end
function sgdmom(x, dx, lr, alpha, state)
-- sgd momentum, uses nesterov update (reference: http://cs231n.github.io/neural-networks-3/#sgd)
if not state.m then
state.m = x.new(#x):zero()
state.tmp = x.new(#x)
end
state.tmp:copy(state.m)
state.m:mul(alpha):add(-lr, dx)
x:add(-alpha, state.tmp)
x:add(1+alpha, state.m)
end
function adagrad(x, dx, lr, epsilon, state)
if not state.m then
state.m = x.new(#x):zero()
state.tmp = x.new(#x)
end
-- calculate new mean squared values
state.m:addcmul(1.0, dx, dx)
-- perform update
state.tmp:sqrt(state.m):add(epsilon)
x:addcdiv(-lr, dx, state.tmp)
end
-- rmsprop implementation, simple as it should be
function rmsprop(x, dx, lr, alpha, epsilon, state)
if not state.m then
state.m = x.new(#x):zero()
state.tmp = x.new(#x)
end
-- calculate new (leaky) mean squared values
state.m:mul(alpha)
state.m:addcmul(1.0-alpha, dx, dx)
-- perform update
state.tmp:sqrt(state.m):add(epsilon)
x:addcdiv(-lr, dx, state.tmp)
end
function adam(x, dx, lr, beta1, beta2, epsilon, state)
local beta1 = beta1 or 0.9
local beta2 = beta2 or 0.999
local epsilon = epsilon or 1e-8
if not state.m then
-- Initialization
state.t = 0
-- Exponential moving average of gradient values
state.m = x.new(#dx):zero()
-- Exponential moving average of squared gradient values
state.v = x.new(#dx):zero()
-- A tmp tensor to hold the sqrt(v) + epsilon
state.tmp = x.new(#dx):zero()
end
-- Decay the first and second moment running average coefficient
state.m:mul(beta1):add(1-beta1, dx)
state.v:mul(beta2):addcmul(1-beta2, dx, dx)
state.tmp:copy(state.v):sqrt():add(epsilon)
state.t = state.t + 1
local biasCorrection1 = 1 - beta1^state.t
local biasCorrection2 = 1 - beta2^state.t
local stepSize = lr * math.sqrt(biasCorrection2)/biasCorrection1
-- perform update
x:addcdiv(-stepSize, state.m, state.tmp)
end
-------------------------
-- Performance Evaluators
-------------------------
function getMean(inputs,targets,model)
diff = torch.abs(model:forward(inputs) - targets)
mean = diff:mean()
return mean
end
function performanceEvaluator(trainset,model,fcnFlag,cudaFlag,items)
means = torch.Tensor(items:size(1))
iSize = fcnFlag and 32 or 1;
for i=1,items:size(1) do
if cudaFlag then
inputs = trainset.images[items[i]]:cuda()
targets = torch.Tensor(1,iSize,iSize):fill(trainset.labels[items[i]]*0.1):cuda()
else
inputs = trainset.images[items[i]]
targets = torch.Tensor(1,iSize,iSize):fill(trainset.labels[items[i]]*0.1)
end
means[i] = (getMean(inputs,targets,model)*100)/(trainset.labels[items[i]]*0.1)
print('Image '.. string.format('%5d',items[i])..' | Label '.. string.format('%2d',trainset.labels[items[i]]) .. ' | Sigma '..string.format('%2.1f', trainset.labels[items[i]]*0.1) ..' | Pred ' ..string.format('%1.8f',model:forward(inputs)[{1,1,1}]) .. ' | MSE ' .. string.format('%1.8f',criterion:forward(model:forward(inputs),targets)) .. ' | RMSE ' .. string.format('%1.8f',torch.sqrt(criterion:forward(model:forward(inputs),targets))) .. ' | Percent Error ' .. string.format('%3.2f',means[i]) .. '%')
end
print('Means of means: ' .. means:mean())
end
function classPerformanceEvaluator(trainset,model,fcnFlag,items)
acc = torch.Tensor(30):fill(0)
hist = torch.Tensor(30):fill(0)
iSize = fcnFlag and 32 or 1;
for i=1,items:size(1) do
inputs = trainset.images[items[i]]:cuda()
targets = torch.Tensor(1,iSize,iSize):fill(trainset.labels[items[i]]*0.1):cuda()
acc[trainset.labels[items[i]]] = acc[trainset.labels[items[i]]]+ (getMean(inputs,targets,model)*100)/(trainset.labels[items[i]]*0.1)
hist[trainset.labels[items[i]]] = hist[trainset.labels[items[i]]] + 1
end
print(torch.cdiv(acc,hist))
print(torch.cdiv(acc,hist):mean())
end
function randomEvaluator(dataSize,evalSize)
local list = torch.randperm(dataSize)[{{1,evalSize}}]
performanceEvaluator(trainset,model,true,true,list)
end
function testClass(dataset,model,cudaFlag,bSize,size)
print('<trainer> on testing Set:')
testLogger = testLogger or require('optim').Logger('./test.log')
local confusion = require('optim').ConfusionMatrix(30)
for t = 1,size,bSize do
xlua.progress(t, dataset:size())
local inputs = dataset.images[{{t, math.min(t+bSize-1,size)}}]
if cudaFlag then
inputs = inputs:cuda()
end
local targets = dataset.labels[{{t, math.min(t+bSize-1,size)}}]
local outputs = model:forward(inputs)
confusion:batchAdd(outputs, targets)
end
print(confusion)
print('\27[31mTest: ' .. confusion.totalValid * 100)
testLogger:add{['% mean class accuracy (test set)'] = confusion.totalValid * 100}
confusion:zero()
end
-------------------------
-- Trainers
-------------------------
function trainerSingle(trainset,model,lr,item,cudaFlag)
params,grad_params = model:getParameters();
iSize = fcnFlag and 32 or 1;
inputs = trainset.images[item]
targets = torch.Tensor(1,iSize,iSize):fill(trainset.labels[item]*0.1)
if cudaFlag then
inputs = inputs:cuda()
targets = targets:cuda()
end
grad_params:zero();
outputs = model:forward(inputs);
currentError = currentError + criterion:forward(outputs, targets);
df_do = criterion:backward(outputs, targets);
model:backward(inputs, df_do);
-- sgd(params,grad_params,lr)
-- rmsprop(params,grad_params,lr,0.99,1e-8,config)
-- adam(params,grad_params,lr,0.9,0.999,1e-7,config)
adagrad(params,grad_params,lr,1e-7,config)
end
function trainerBatch(dataset, model, lr, bSize, size, cudaFlag, classFlag, fcnFlag)
print('Training with batch size ' .. bSize .. ' and learning rate ' .. lr .. ' and size ' .. size)
local params,grad_params = model:getParameters();
local iSize = fcnFlag and 32 or 1;
local confusion = require('optim').ConfusionMatrix(30)
trainLogger = trainLogger or optim.Logger('./train.log')
local p = torch.randperm(size):long();
for t = 1,size,bSize do
grad_params:zero();
local inputs = dataset.images:index(1,p[{{t, math.min(t+bSize-1,size)}}])
if classFlag then
targets = dataset.labels:index(1,p[{{t, math.min(t+bSize-1,size)}}])
else
targets = torch.Tensor(math.min(t+bSize-1,size)-t+1,1,iSize,iSize)
for i=t,math.min(t+bSize-1,size) do
targets[i-t+1] = targets[i-t+1]:fill(0.1*trainset.labels[i-t+1])
end
end
if cudaFlag then
inputs = inputs:cuda();
targets = classFlag and targets or targets:cuda();
criterion = criterion:cuda()
end
local outputs = model:forward(inputs);
if classFlag then
confusion:batchAdd(outputs, targets)
end
local f = criterion:forward(outputs, targets);
local df_do = criterion:backward(outputs, targets);
model:backward(inputs, df_do);
adagrad(params,grad_params,lr,1e-8,config)
-- sgd(params,grad_params,lr)
currentError = currentError + f
xlua.progress(t,size)
end
if classFlag then
print(confusion)
print('\27[32mTrain: ' .. confusion.totalValid * 100)
trainLogger:add{['% mean class accuracy (train set)'] = confusion.totalValid * 100}
confusion:zero()
end
end