-
Notifications
You must be signed in to change notification settings - Fork 24
/
conf
226 lines (183 loc) · 6.69 KB
/
conf
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
##################### edu.cmu.lti.util.run.Tunner
num_processes=1
# number of working processes for tuning parameters in parallel (default=4)
tunner_mode=sweep
# the type of parameter tuning (default=sweep)
num_avarage=1
# we do experiments N times and average the scores (default=1)
split_summary=false
starting_id=0
print_std_dev=false
# also print the standard deviations of the evaluation measures
#cmd=java -Xms16G -Xmx16G -cp ~/code_java/ni/class edu.cmu.pra.LearnerPRA
cmd=perl ../../pra.20G.pl
# the command used to run an experiment during parameter tuning
##################### edu.cmu.lti.util.run.Learner
num_threads=1
task=train
# decide the type of task to perform (default=sCV)
train // train using trainFile
, test //test using testFile
, STT //train test by splitting trainFile
, TT //train test using trainFile, testFile
, CV //cross validation using trainFile
, sCV //self cross validation using trainFile
, sLOO //self leave-on-out cross validation using trainFile
, LOO //leave one out cross validation using trainFile
, predict // without knowing the ground truth
id=a
# add a postfix the result folder
num_CV_folds=3
# number of fold for cross validations (default=5)
train_rate=0.7
# the ratio when splitting data into train and test sets (default=0.8)
time_field=0
# the column id of query time in the data files
output_folder=./
#train_samples=../tmp/training_queries/<target_relation>
train_samples=./queriesR_train/<target_relation>
# data file used for training
test_samples=./queriesR_test/<target_relation>
# data file used for test
createQueries_max_relation_count=100000
# max number of edges for a relation during query creation
createQueries_min_relation_count=10
# min number of edges for a relation during query creation
##################### edu.cmu.pra.LearnerPRA
target_relation=worksfor
# (optional) what relation we are training for
# _competeswith athleteplaysforteam citylocatedinstate
# _specializationof agentcompeteswithagent
graph_folder=./graphs/NELL165/
# where is the graph NELL446svo NELL165
model=./models/<target_relation>
# a weight file can be loaded for prediction
model_folder=./models/
prediction_folder=./predictions/
top_prediction_results=5
#subgraph_store=
# subgraphs that can be loaded/unloaded
rank_mode=Path
# the loss function during training (default=P)
Path //PathRank: each relation path has a weight
Rel //RelationRank: each relation has a weight
loss_mode=log
# the loss function (default=log)
none //no training
,log // log loss
,exp //exp loss
,hinge //hinge loss
L1=0.001
#L1 regularization
L2=0.001
#L2 regularization
pairwise_loss=false
# optimizing partial ordering or relevance (default=false)
false //optimizing relevance
true //optimizing partial ordering
negative_mode=Sqr
# how to select negative entities
all //use all negative samples
,topK //take top K*rNega samples ranked by plain weighting
,expX // take a^x:
, Sqr // take x^2: 0,1,4,9,16,25,...th negative samples
,Tri // take x^3: 0,1,8,27,64,125,...th negative samples
,poly // take x^a:
,none // add none
negative_weight=20
# the weight shared by all negative samples
cache_data=true
inspect_data=true
output_training_data=true
output_training_data_folder=./training_data/
#max_num_exploration_queries=400
max_num_exploration_particles=1000000
################## edu.cmu.pra.model.PRAModel ########################################
max_steps=3
# max length of relation paths -0
bias=true
# add a bias term in the scoring function. can improve log-likelihood (default=true)
bias_value=1
# value of the bias feature. can control its regularization (default=0.1)
#blocked_paths=1:a(_Author)p(Cites)p
# some paths are not allowed in the path tree: idSeed,path;idSeed,path;...
# a(_Author)p(Cites)p _Author>Cites
#and_paths=1~_generalizations
# the targets have to be in the result of this path
#not_paths=1:a(_Author)p
# the targets cannot be in the result of this path
#a(_Author)p _Author
#rerank_and_paths=2:_generalizations
# the and-filter used at rerank stage
#no_cost_relations=athleteplaysinleague;_athleteplaysinleague;teammate
# a ; separated list of promoted relations which have no cost
# (regular relations each have cost 1)
# to avoid creating infite long paths, the path creation precedure
# will prevent going forward-backward with zero-cost relations
min_feature_hit=3
min_feature_accuracy=0.01
cache_RW=false
################## edu.cmu.pra.model.EntityRank
entity_rank=false
# whether to use entity rank or not
min_ER_steps=3
# max length of entity rank paths (default=3)
ER_time_gap=2
# only generates entity rank one for every N years, in order to save memory (default=1)
################## edu.cmu.pra.graph.Graph
timed_graph=false
# whether the graph contains time information
time_type=year
# the entity type which represent time (assuming which can be parsed as integers)
time_relation=Year
# the relation type that connect time entities to other entities
# OBSOLETED: moved to the task file
################## edu.cmu.pra.graph.Walker
#inferred_relation_file=./infered_relations
RW_mode=LVParticle
# how to keep the random walk distributions sparse (default=Q)
Truncate, //Truncation
Beem, //Beam
LVWalker, // Sampling (low variance)
Walker, // Sampling
Particle, //particle filtering
LVParticle, //particle filtering (low variance)
Exact, //none
truncate=0.003
# absolute truncation
beam_width=10000
#beam truncation
num_walkers=10000
#sampling
min_particle=0.0001
#particle filtering
blocked_field=0
# the field in query file which leads to the target node
RW_renormalize=false
damping_factor=1.0
##################### Optimization #####################
max_train_iteration=100
epsCvg=0.00001
# 0.002
maxLinfStep=10000
maxL1Step=20000
20 100
stepAdapt=0.1
num_stable=2
#################### Visualization#####################
feature_comments=../../mFeatureComments.read
# add explanations to selected relation paths
LBFGS_history=false
#################### edu.cmu.lti.util.run.Reporter #####################
#publish_folder=./www/test/
publish_folder=/afs/cs.cmu.edu/user/nlao/www/demo/wk/test/
# the folder where report is to be copied to
reporter_keys=agentcompeteswithagent
#reporter_keys=athleteplaysforteam,athleteplaysinleague,athleteplayssport,stadiumlocatedincity,teamhomestadium,teamplaysincity,teamplaysinleague,teamplayssport,competeswith,hasofficeincity,journalistwritesforpublication,professionistypeofprofession,teamwontrophy,worksfor
# the key words to search and group results
#result_filter=max_steps=3_0
# only consider a subset of runs for which paramter string contains this filter string
col1=target_relation
col2=nSampling
valueCol=mrr,MAP,tRW,tCV
# the metrics to be shown in the report