-
Notifications
You must be signed in to change notification settings - Fork 0
/
params_def.go
222 lines (213 loc) · 9.03 KB
/
params_def.go
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
// Copyright (c) 2020, The CCNLab Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package main
import "github.com/emer/emergent/params"
// ParamSets is the default set of parameters -- Base is always applied, and others can be optionally
// selected to apply on top of that
var ParamSets = params.Sets{
{Name: "Base", Desc: "these are the best params", Sheets: params.Sheets{
"Network": ¶ms.Sheet{
// layer classes, specifics
{Sel: "Layer", Desc: "needs some special inhibition and learning params",
Params: params.Params{
"Layer.Inhib.FBAct.Tau": "30", // 30 > 20 >> 1 definitively
"Layer.Act.Dt.IntTau": "40", // 40 > 20
"Layer.Inhib.Layer.Gi": "1.1", // general default
"Layer.Inhib.Pool.Gi": "1.1", // general default
"Layer.Inhib.ActAvg.LoTol": "1.1", // no low adapt
"Layer.Inhib.ActAvg.AdaptRate": "0.2", // 0.5 default
"Layer.Act.Gbar.L": "0.2", // 0.2 now best
"Layer.Act.Decay.Act": "0.0", // 0 best
"Layer.Act.Decay.Glong": "0.0", // 0.5 > 0.2
"Layer.Act.KNa.Fast.Max": "0.1", // fm both .2 worse
"Layer.Act.KNa.Med.Max": "0.2", // 0.2 > 0.1 def
"Layer.Act.KNa.Slow.Max": "0.2", // 0.2 > higher
"Layer.Act.Noise.Dist": "Gaussian",
"Layer.Act.Noise.Mean": "0.0", // .05 max for blowup
"Layer.Act.Noise.Var": "0.01", // .01 a bit worse
"Layer.Act.Noise.Type": "NoNoise", // off for now
"Layer.Act.GTarg.GeMax": "1.2", // 1.2 > 1 > .8 -- rescaling not very useful.
"Layer.Act.Dt.LongAvgTau": "20", // 20 > 50 > 100
"Layer.Learn.TrgAvgAct.ErrLrate": "0.01", // 0.01 orig > 0.005
"Layer.Learn.TrgAvgAct.SynScaleRate": "0.005", // 0.005 orig > 0.01
"Layer.Learn.TrgAvgAct.TrgRange.Min": "0.5", // .5 > .2 overall
"Layer.Learn.TrgAvgAct.TrgRange.Max": "2.0", // objrec 2 > 1.8
}},
{Sel: ".CT", Desc: "CT gain factor is key",
Params: params.Params{
"Layer.CtxtGeGain": "0.15", // .1 > .05
"Layer.Inhib.Layer.Gi": "1.1",
"Layer.Act.KNa.On": "true",
"Layer.Act.NMDA.Gbar": "0.03",
"Layer.Act.GABAB.Gbar": "0.2",
"Layer.Act.Decay.Act": "0.0", // 0 better
"Layer.Act.Decay.Glong": "0.0",
}},
{Sel: "TRCLayer", Desc: "avg mix param",
Params: params.Params{
"Layer.TRC.NoTopo": "false", //
"Layer.TRC.AvgMix": "0.5", //
"Layer.TRC.DriveScale": "0.2", // .2 > .15 > .1, .05
"Layer.Act.NMDA.Gbar": "0.03",
"Layer.Act.GABAB.Gbar": "0.2", //
"Layer.Act.Decay.Act": "0.5",
"Layer.Act.Decay.Glong": "1", // clear long
}},
{Sel: "SuperLayer", Desc: "burst params don't really matter",
Params: params.Params{
"Layer.Burst.ThrRel": "0.1", // not big diffs
"Layer.Burst.ThrAbs": "0.1",
}},
{Sel: ".V1", Desc: "pool inhib, initial activity",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.1",
"Layer.Inhib.Pool.On": "false",
"Layer.Inhib.ActAvg.Init": "0.03",
"Layer.Inhib.ActAvg.Targ": "0.03",
}},
{Sel: ".LIP", Desc: "high, pool inhib",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.1",
"Layer.Inhib.Pool.On": "false", // false > true
"Layer.Inhib.ActAvg.Init": "0.05",
"Layer.Inhib.ActAvg.Targ": "0.05",
}},
{Sel: "#LIPP", Desc: "high, pool inhib",
Params: params.Params{
"Layer.Inhib.Layer.Gi": "1.1",
}},
{Sel: ".PopIn", Desc: "pop-code input",
Params: params.Params{
"Layer.Inhib.ActAvg.Init": "0.06",
"Layer.Inhib.ActAvg.Targ": "0.06",
}},
{Sel: "#EyePos", Desc: "eyeposition input",
Params: params.Params{
"Layer.Inhib.ActAvg.Init": "0.02", // .02 > .06 -- needs to be stronger
"Layer.Inhib.ActAvg.Targ": "0.02",
}},
// prjn classes, specifics
{Sel: "Prjn", Desc: "yes extra learning factors",
Params: params.Params{
"Prjn.Learn.Lrate.Base": "0.02", // .02 > .04 here & lvis
// "Prjn.SWt.Init.Sym": "false", // experimenting with asymmetry
"Prjn.PrjnScale.ScaleLrate": "2", // 2 = fast response, effective
"Prjn.PrjnScale.LoTol": "0.8", // good now...
"Prjn.PrjnScale.Init": "1",
"Prjn.PrjnScale.AvgTau": "500", // slower default
"Prjn.PrjnScale.Adapt": "false", // adapt bad maybe? put GeMax at 1.2, adjust to avoid
"Prjn.SWt.Adapt.On": "true", // true > false, esp in cosdiff
"Prjn.SWt.Adapt.Lrate": "0.01", //
"Prjn.SWt.Adapt.SigGain": "6",
"Prjn.SWt.Adapt.DreamVar": "0.0", // 0.02 good in lvis
"Prjn.SWt.Init.SPct": "1", // 1 > lower
"Prjn.SWt.Init.Mean": "0.5", // .5 > .4 -- key, except v2?
"Prjn.SWt.Limit.Min": "0.2", // .2-.8 == .1-.9; .3-.7 not better
"Prjn.SWt.Limit.Max": "0.8", //
}},
{Sel: "CTCtxtPrjn", Desc: "defaults for CT Ctxt prjns",
Params: params.Params{
"Prjn.PrjnScale.Rel": "1",
}},
{Sel: ".Fixed", Desc: "fixed weights",
Params: params.Params{
"Prjn.Learn.Learn": "false",
"Prjn.PrjnScale.Adapt": "false", // key to not adapt!
"Prjn.SWt.Init.Mean": "0.8", // 0.8 better
"Prjn.SWt.Init.Var": "0",
"Prjn.SWt.Init.Sym": "false",
}},
{Sel: ".Forward", Desc: "std feedforward",
Params: params.Params{
// "Prjn.PrjnScale.Init": "0.8", // weaker?
}},
{Sel: ".Back", Desc: "top-down back-projections MUST have lower relative weight scale, otherwise network hallucinates -- smaller as network gets bigger",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.2",
}},
{Sel: ".Inhib", Desc: "inhibitory projection",
Params: params.Params{
"Prjn.Learn.Learn": "true", // learned decorrel is good
"Prjn.Learn.Lrate.Base": "0.0001", // .0001 > .001 -- slower better!
"Prjn.SWt.Init.Var": "0.0",
"Prjn.SWt.Init.Mean": "0.1",
"Prjn.SWt.Adapt.On": "false",
"Prjn.PrjnScale.Init": "0.1", // .1 = .2, slower blowup
"Prjn.PrjnScale.Adapt": "false",
"Prjn.IncGain": "1", // .5 def
}},
{Sel: ".FwdWeak", Desc: "weak feedforward",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1", // .1 orig -- had a bug tho!! also trying .05
}},
{Sel: ".FmLIP", Desc: "no random weights here",
Params: params.Params{
"Prjn.SWt.Init.Var": "0.05", // was 0 -- trying .05
"Prjn.SWt.Init.Sym": "false",
}},
{Sel: ".BackStrong", Desc: "stronger",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.2", // .1 > orig .2 > .05 -- not sep fm BackMax -- .1 = better TE_V1Sim, V2P cosdiff
}},
{Sel: ".BackMax", Desc: "strongest",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.2", // .1 > .2, orig .5 -- see BackStrong
}},
{Sel: ".CTToPulv", Desc: "CT to pulvinar needs to be weaker in general, like most prjns",
Params: params.Params{
"Prjn.PrjnScale.Init": "0.8",
"Prjn.PrjnScale.Rel": "1.25",
}},
{Sel: ".BackToPulv", Desc: "top-down to pulvinar directly",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1",
}},
{Sel: ".FwdToPulv", Desc: "feedforward to pulvinar directly",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.1",
}},
{Sel: ".FmPulv", Desc: "default for pulvinar",
Params: params.Params{
"Prjn.PrjnScale.Rel": "0.2", // .2 > .1 > .05 still true
}},
{Sel: ".Lateral", Desc: "default for lateral",
Params: params.Params{
"Prjn.SWt.Init.Sym": "false",
"Prjn.PrjnScale.Rel": "0.02", // .02 > .05 == .01 > .1 -- very minor diffs on TE cat
"Prjn.SWt.Init.Mean": "0.5",
"Prjn.SWt.Init.Var": "0",
}},
{Sel: ".CTFmSuper", Desc: "CT from main super -- fixed one2one",
Params: params.Params{
"Prjn.SWt.Init.Mean": "0.8", // 0.8 > 0.5 with lower S -> CT rel (2 instead of 4)
"Prjn.PrjnScale.Rel": "1", // def 2
}},
{Sel: ".CTFmSuperLower", Desc: "CT from main super -- for lower layers",
Params: params.Params{
"Prjn.SWt.Init.Mean": "0.8", // 0.8 makes a diff for lower too, more V1 divergence at .5
"Prjn.PrjnScale.Rel": "1", // 1 maybe better
}},
{Sel: ".CTSelfLIP", Desc: "CT to CT for LIP",
Params: params.Params{
"Prjn.PrjnScale.Rel": "2", // 2 = 3 > 1
}},
{Sel: ".CTBack", Desc: "CT to CT back (top-down)",
Params: params.Params{
"Prjn.PrjnScale.Rel": ".2", // .2 > .1 in std
}},
{Sel: ".SToCT", Desc: "higher Super to CT back (top-down), leaks current state to prediction..",
Params: params.Params{
"Prjn.PrjnScale.Rel": ".2", // .2 > .1 in std
}},
{Sel: ".CTBackMax", Desc: "CT to CT back (top-down), max",
Params: params.Params{
"Prjn.PrjnScale.Rel": ".5",
}},
{Sel: ".SToCTMax", Desc: "higher Super to CT back (top-down), leaks current state to prediction..",
Params: params.Params{
"Prjn.PrjnScale.Rel": ".5",
}},
},
}},
}