-
Notifications
You must be signed in to change notification settings - Fork 6
/
TorchSeq2PC.py
195 lines (151 loc) · 6.22 KB
/
TorchSeq2PC.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
import torch
print('Running TorchSeq2PC.py')
#NEWER
# Perform a forward pass on a Sequential model
# where X,Y are one batch of inputs,labels
# Returns activations for all layers (vhat), loss, and gradient of loss
# wrt last-layer activations (dLdy)
# vhat,Loss,dLdy=FwdPassPlus(model,LossFun,X,Y)
def FwdPassPlus(model,LossFun,X,Y):
# Number of layers, counting the input as layer 0
DepthPlusOne=len(model)+1
# Forward pass
vhat=[None]*DepthPlusOne
vhat[0]=X
for layer in range(1,DepthPlusOne):
f=model[layer-1]
vhat[layer]=f(vhat[layer-1])
Loss = LossFun(vhat[-1], Y)
# Compute gradient of loss with respect to output
dLdy=torch.autograd.grad(Loss, vhat[-1])[0]
return vhat,Loss,dLdy
# Compute prediction errors (epsilon) and beliefs (v)
# using predictive coding algorithm modified by
# the fixed prediction assumption
# see: Millidge, Tschantz, and Buckley. Predictive coding approximates backprop along arbitrary computation graphs.
# v,epsilon=FixedPredPCPredErrs(model,vhat,dLdy,eta=1,n=None)
def FixedPredPCPredErrs(model,vhat,dLdy,eta=1,n=None):
# Number of layers, counting the input as layer 0
DepthPlusOne=len(model)+1
if n==None:
n=len(model)
# Initialize epsilons
epsilon=[None]*DepthPlusOne
epsilon[-1]=dLdy
# Initialize v to a copy of vhat with no gradients needed
# (can this be moved up to the loop above?)
v=[None]*DepthPlusOne
for layer in range(DepthPlusOne):
v[layer]=vhat[layer].clone().detach()
# Iterative updates of v and epsilon using stored values of vhat
for i in range(n):
for layer in reversed(range(DepthPlusOne-1)):#range(DepthPlusOne-2,-1,-1):
epsilon[layer]=vhat[layer]-v[layer]
_,epsdfdv=torch.autograd.functional.vjp(model[layer],vhat[layer],epsilon[layer+1])
dv=epsilon[layer]-epsdfdv
v[layer]=v[layer]+eta*dv
# This helps free up memory
with torch.no_grad():
for layer in range(1,DepthPlusOne-1):
v[layer]=v[layer].clone()
epsilon[layer]=epsilon[layer].clone()
return v,epsilon
# Compute prediction errors (epsilon) and beliefs (v)
# using a strict interpretation of predictive coding
# without the fixed prediction assumption.
# v,epsilon=StrictPCPredErrs(model,vinit,LossFun,Y,eta,n)
def StrictPCPredErrs(model,vinit,LossFun,Y,eta,n):
with torch.no_grad():
# Number of layers, counting the input as layer 0
DepthPlusOne=len(model)+1
# Initialize epsilons
epsilon=[None]*DepthPlusOne
# Initialize v to a copy of vinit with no gradients needed
# (can this be moved up to the loop above?)
v=[None]*DepthPlusOne
for layer in range(DepthPlusOne):
v[layer]=vinit[layer].clone()
# Iterative updates of v and epsilon
for i in range(n):
model.zero_grad()
layer=DepthPlusOne-1
vtilde=model[layer-1](v[layer-1])
Loss=LossFun(vtilde,Y)
epsilon[layer]=torch.autograd.grad(Loss,vtilde,retain_graph=False)[0] # -2 ~ DepthPlusOne-2
for layer in reversed(range(1,DepthPlusOne-1)):
epsilon[layer]=model[layer-1](v[layer-1])-v[layer]
_,epsdfdv=torch.autograd.functional.vjp(model[layer],v[layer],epsilon[layer+1])
dv=epsilon[layer]-epsdfdv
v[layer]=v[layer]+eta*dv
# This helps free up memory
with torch.no_grad():
for layer in range(1,DepthPlusOne-1):
v[layer]=v[layer].clone()
epsilon[layer]=epsilon[layer].clone()
return v,epsilon
# Compute exact prediction errors (epsilon) and beliefs (v)
# epsilon is defined as the gradient of the loss wrt to
# the activations and v=vhat-epsilon where vhat are the
# activations from a forward pass.
# v,epsilon=ExactPredErrs(model,LossFun,X,Y,vhat=None)
def ExactPredErrs(model,LossFun,X,Y,vhat=None):
# Number of layers, counting the input as layer 0
DepthPlusOne=len(model)+1
# Forward pass if it wasn't passed in
if vhat==None:
vhat=[None]*DepthPlusOne
vhat[0]=X
for layer in range(1,DepthPlusOne):
f=model[layer-1]
vhat[layer]=f(vhat[layer-1])
Loss = LossFun(vhat[-1], Y)
epsilon=[None]*DepthPlusOne
v=[None]*DepthPlusOne
for layer in range(1,DepthPlusOne):
epsilon[layer]=torch.autograd.grad(Loss,vhat[layer],allow_unused=True,retain_graph=True)[0]
v[layer]=vhat[layer]-epsilon[layer]
return v,epsilon
# Set gradients of model params based on PC approximations
def SetPCGrads(model,epsilon,X,v=None):
# Number of layers, counting the input as layer 0
DepthPlusOne=len(model)+1
# Forward pass if v wasn't passed in
if v==None:
v=[None]*DepthPlusOne
v[0]=X
for layer in range(1,DepthPlusOne):
f=model[layer-1]
v[layer]=f(v[layer-1])
# Compute new parameter values
for layer in range(0,DepthPlusOne-1):
with torch.no_grad():
vtemp0=v[layer].clone()
vtemp0.requires_grad=True
vtemp1=model[layer](vtemp0)
for p in model[layer].parameters():
dtheta=torch.autograd.grad(vtemp1,p,grad_outputs=epsilon[layer+1],allow_unused=True,retain_graph=True)[0]
p.grad = dtheta
# Perform a whole PC inference step
# Returns activations (vhat), loss, gradient of the loss wrt output (dLdy),
# beliefs (v), and prediction errors (epsilon)
# vhat,Loss,dLdy,v,epsilon=PCInfer(model,LossFun,X,Y,ErrType="FixedPred",eta=.1,n=20,vinit=None)
def PCInfer(model,LossFun,X,Y,ErrType,eta=.1,n=20,vinit=None):
# Fwd pass (plus return vhat and dLdy)
vhat,Loss,dLdy=FwdPassPlus(model,LossFun,X,Y)
# Get beliefs and prediction errors
if ErrType=="FixedPred":
v,epsilon=FixedPredPCPredErrs(model,vhat,dLdy,eta,n)
SetPCGrads(model,epsilon,X,vhat)
elif ErrType=="Strict":
if vinit==None:
vinit=vhat
v,epsilon=StrictPCPredErrs(model,vhat,LossFun,Y,eta,n)
SetPCGrads(model,epsilon,X,v)
elif ErrType=="Exact":
v,epsilon=ExactPredErrs(model,LossFun,X,Y)
SetPCGrads(model,epsilon,X,vhat)
else:
raise ValueError('ErrType must be \"FixedPred\", \"Strict\", or \"Exact\"')
# Set gradients in model
#SetPCGrads(model,epsilon,X,vhat)
return vhat,Loss,dLdy,v,epsilon