-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathneuralNetwork.py
349 lines (283 loc) · 12.3 KB
/
neuralNetwork.py
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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
# -*- coding: utf-8 -*-
"""
Created on Mon Jan 9 12:33:43 2023
@author: uqalim8
"""
import torch.nn as nn
import torch, math
from functorch import make_functional
from hyperparameters import cCUDA, cTYPE
from regularizers import none_reg
class auto_Encoder_MNIST(nn.Module):
def __init__(self):
super(auto_Encoder_MNIST, self).__init__()
self.encoder = nn.Sequential(nn.Linear(28*28, 512),
nn.Tanh(),
nn.Linear(512, 256),
nn.Tanh(),
nn.Linear(256, 128),
nn.Tanh(),
nn.Linear(128, 64),
nn.Tanh(),
nn.Linear(64, 32),
nn.Tanh(),
nn.Linear(32, 16),
)
self.decoder = nn.Sequential(nn.Linear(16, 32),
nn.Tanh(),
nn.Linear(32, 64),
nn.Tanh(),
nn.Linear(64, 128),
nn.Tanh(),
nn.Linear(128, 256),
nn.Tanh(),
nn.Linear(256, 512),
nn.Tanh(),
nn.Linear(512, 28*28),
)
def forward(self, x):
x = self.encoder(x)
x = self.decoder(x)
return x
class FFN(nn.Module):
"""
Do not initialise the weights at zeros
"""
def __init__(self, input_dim, output_dim):
super().__init__()
self._nn = nn.Sequential(nn.Linear(input_dim, 256),
nn.Sigmoid(),
nn.Linear(256, 256),
nn.Tanh(),
nn.Linear(256, 128),
nn.Sigmoid(),
nn.Linear(128, 128),
nn.Tanh(),
nn.Linear(128, 64),
nn.Sigmoid(),
nn.Linear(64, 64),
nn.Tanh(),
nn.Linear(64, 32),
nn.Sigmoid(),
nn.Linear(32, 32),
nn.Tanh(),
nn.Linear(32, output_dim),
nn.Softmax(dim = 1))
def forward(self, x):
return self._nn(x)
class RNNet(nn.Module):
def __init__(self, input_dim, hidden_size, layers, depth, output):
super().__init__()
self._hidden_size = hidden_size
self._layers = layers
self._input_dim = input_dim
self._nn = nn.Sequential(nn.Linear(layers * hidden_size, depth[0]),
nn.Tanh(),
nn.Linear(depth[0], depth[1]),
nn.Tanh(),
nn.Linear(depth[1], output))
self._rnn = nn.RNN(input_dim, hidden_size, layers, batch_first = True)
def forward(self, x):
with torch.backends.cudnn.flags(enabled=False):
x = self._rnn(x)[1]
x = x.movedim(0, 1).reshape(-1, self._layers * self._hidden_size)
return self._nn(x)
class Wrapper:
def __init__(self):
self._funcs = {"0" : self.f, "1" : self.g, "2" : self.Hv,
"01" : self.fg, "02" : self.fHv, "12" : self.gHv,
"012": self.fgHv}
def f(self, w):
raise NotImplementedError
def g(self, w):
raise NotImplementedError
def Hv(self, w):
raise NotImplementedError
def fg(self, w):
raise NotImplementedError
def fgHv(self, w):
raise NotImplementedError
def gHv(self, w):
raise NotImplementedError
def fHv(self, w):
raise NotImplementedError
def __call__(self, x, order):
return self._funcs[order](x)
class funcWrapper(Wrapper):
def __init__(self, func):
super().__init__()
self.func = func
def _gradIt(self, w):
if w.requires_grad:
return w.detach().requires_grad_(True)
else:
return w.requires_grad_(True)
def f(self, w):
with torch.no_grad():
return self.func(w)
def g(self, w):
w = self._gradIt(w)
f = self.func(w)
g = torch.autograd.grad(f, w)[0]
f.detach()
return g.detach()
def fg(self, w):
w = self._gradIt(w)
f = self.func(w)
g = torch.autograd.grad(f, w)[0]
return f.detach(), g.detach()
def fgHv(self, w):
w = self._gradIt(w)
f = self.func(w)
g = torch.autograd.grad(f, w, create_graph = True)[0]
Hv = lambda v : torch.autograd.grad((g,), w, v, create_graph = False,
retain_graph = True)[0].detach()
return f.detach(), g.detach(), Hv
class nnWrapper(Wrapper):
def __init__(self, func, loss):
super().__init__()
self.func, self.loss = func, loss
def _toModule_toFunctional(self, w):
if w.requires_grad:
w = w.detach().requires_grad_(True)
else:
w = w.requires_grad_(True)
nn.utils.vector_to_parameters(w, self.func.parameters())
return make_functional(self.func, disable_autograd_tracking = False)
def f(self, x, X, Y):
device = x.device
functional, w = self._toModule_toFunctional(x)
with torch.no_grad():
return self.loss(functional(w, X.to(device)), Y.to(device))
def g(self, x, X, Y):
device = x.device
functional, w = self._toModule_toFunctional(x)
val = self.loss(functional(w, X.to(device)), Y.to(device))
g = torch.autograd.grad(val, w)
g = nn.utils.parameters_to_vector(g)
return g.detach()
def fg(self, x, X, Y):
device = x.device
functional, w = self._toModule_toFunctional(x)
val = self.loss(functional(w, X.to(device)), Y.to(device))
g = torch.autograd.grad(val, w)
g = nn.utils.parameters_to_vector(g)
return val.detach(), g.detach()
def fgHv(self, x, X, Y):
device = x.device
functional, x = self._toModule_toFunctional(x)
val = self.loss(functional(x, X.to(device)), Y.to(device))
g = torch.autograd.grad(val, x, create_graph = True)
g = nn.utils.parameters_to_vector(g)
Hv = lambda v : nn.utils.parameters_to_vector(
torch.autograd.grad(g, x, grad_outputs = v, create_graph = False, retain_graph = True)
).detach()
return val.detach(), g.detach(), Hv
def Hv(self, x, X, Y):
device = x.device
functional, x = self._toModule_toFunctional(x)
val = self.loss(functional(x, X.to(device)), Y.to(device))
g = torch.autograd.grad(val, x, create_graph = True)
g = nn.utils.parameters_to_vector(g)
return lambda v : nn.utils.parameters_to_vector(torch.autograd.grad(g, x, v, create_graph = False, retain_graph = True)
).detach()
def _HvSingle(self, v, x, X, Y):
device = x.device
functional, x = self._toModule_toFunctional(x)
val = self.loss(functional(x, X.to(device)), Y.to(device))
g = torch.autograd.grad(val, x, create_graph = True)
g = nn.utils.parameters_to_vector(g)
return nn.utils.parameters_to_vector(torch.autograd.grad(g, x, v)).detach()
def __call__(self, x, order, X, Y):
return self._funcs[order](x, X, Y)
class ObjFunc:
def __init__(self, func, loss, X, Y, reg, mini, Hsub):
self.X, self.Y = X.to(cCUDA), Y.to(cCUDA)
self.fun = nnWrapper(func, loss)
self.Hsub = Hsub
self.mini = mini
if reg is None:
self.reg = none_reg
else:
self.reg = funcWrapper(reg)
def minibatch(self, x, order):
n = self.X.shape[0]
m = math.ceil(n * self.mini)
perm = torch.randperm(n)
if order == "0" or order == "1":
return self.fun(x, order, self.X[perm[:m]], self.Y[perm[:m]]) + self.reg(x, order)
if order == "01":
f, g = self.fun(x, "01", self.X[perm[:m]], self.Y[perm[:m]])
f_reg, g_reg = self.reg(x, order)
return f + f_reg, g + g_reg
def subHessian(self, x, order):
n = self.X.shape[0]
m = math.ceil(n * self.Hsub)
perm = torch.randperm(n)
if order == "012":
f1, g1, Hv = self.fun(x, order, self.X[perm[:m]], self.Y[perm[:m]])
f2, g2 = self.fun(x, "01", self.X[perm[m:]], self.Y[perm[m:]])
f_reg, g_reg, Hv_reg = self.reg(x, order)
f = m * f1 / n + (n - m) * f2 / n
g = m * g1 / n + (n - m) * g2 / n
return f + f_reg, g + g_reg, lambda v : Hv(v) + Hv_reg(v)
if order == "2":
return lambda v : self.fun.Hv(x, self.X[perm[:m]], self.Y[perm[:m]])(v) + self.reg(x, order)(v)
if order == "02" or order == "12":
raise NotImplementedError
def __call__(self, x, order):
if order == "f":
return self.fun(x, "0", self.X, self.Y) + self.reg(x, "0")
if "2" in order and self.Hsub != 1:
return self.subHessian(x, order)
if self.mini != 1:
return self.minibatch(x, order)
if order == "0" or order == "1":
return self.fun(x, order, self.X, self.Y) + self.reg(x, order)
if order == "2":
return lambda w : self.fun(x, order, self.X, self.Y)(w) + self.reg(x, order)(w)
if order == "01":
f, g = self.fun(x, order, self.X, self.Y)
reg_f, reg_g = self.reg(x, order)
return f + reg_f, g + reg_g
if order == "012":
f, g, Hv = self.fun(x, order, self.X, self.Y)
reg_f, reg_g, reg_Hv = self.reg(x, order)
return f + reg_f, g + reg_g, lambda x : Hv(x) + reg_Hv(x)
if order == "12" or order == "02":
gof, Hv = self.fun(x, order, self.X, self.Y)
reg_gof, reg_Hv = self.reg(x, order)
return gof + reg_gof, lambda x : Hv(x) + reg_Hv(x)
if __name__ == "__main__":
from derivativeTest import derivativeTest
from funcs import logisticFun, logisticModel
#from regularizers import non_convex
trainX = torch.rand((2000, 28 * 28)).to(cTYPE)
trainY = torch.rand((2000, 10)).to(cTYPE)
ffn = FFN(28 * 28, 10).to(cTYPE)
loss = nn.MSELoss()
s = nn.utils.parameters_to_vector(ffn.parameters()).detach()
#fun = ObjFunc(ffn, loss, trainX, trainY, None, 1)
# Test to see if the accumulative derivatives are the same as
# non-accumulative derivatives and if the derivatives are correct
# cSPLIT = 20000
# f1, g1, Hv1 = derivativeTest(lambda x : fun(x, "012"), s)
# cSPLIT = 27
# f2, g2, Hv2 = derivativeTest(lambda x : fun(x, "012"), s)
# assert torch.all(torch.isclose(f1, f2))
# assert torch.all(torch.isclose(g1, g2))
# assert torch.all(torch.isclose(Hv1, Hv2))
# Check others
# for i in range(100):
# x = torch.rand(s.shape[0], dtype = cTYPE)
# cSPLIT = 10000
# f1, g1 = fun(x, "01")
# cSPLIT = 479
# f2, g2 = fun(x, "01")
# assert torch.isclose(torch.mean(g1), torch.mean(g2))
# assert torch.isclose(f1, f2)
# print("Passed!")
#tests.subHessian_test(lambda x, y, v : ObjFunc(ffn, loss, x, y, None, 1, v), s.shape[0], trainX, trainY)
#print("Passed!")
#tests.minibatch_test(lambda x, y, v : ObjFunc(ffn, loss, x, y, None, v, 1), s.shape[0], trainX, trainY)
#print("Passed!")