-
Notifications
You must be signed in to change notification settings - Fork 2
/
dbltrain_options.txt
70 lines (67 loc) · 4.19 KB
/
dbltrain_options.txt
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
BiCVM Distributed Representation Learner: Copyright 2013-2014 Karl Moritz Hermann
Allowed options:
--type arg (=additive) type of model (additive, flattree)
--input1 arg l1 corpus. sentence aligned to l2.
--input2 arg l2 corpus. sentence aligned to l1.
--model1-in arg initial model 1
--model2-in arg initial model 2
--model1-out arg (=modelA) base filename of model 1 output files
--model2-out arg (=modelB) base filename of model 2 output files
--tree arg (=ccg) tree type (ccg, stanford)
-s [ --cv-split ] arg (=-1) cross-validation (split)
--word-width arg (=50) width of word representation vectors.
--iterations arg (=-1) (maximum) number of iterations (lbfgs
default: 0 / sgd 250)
--ftiterations arg (=1000) (maximum) number of finetune iterations
--dump-frequency arg (=10) frequency at which to dump the model
-n [ --num-sentences ] arg (=0) number of sentences to consider
--method arg (=lbfgs) training method (options:
lbfgs,sgd,fgc,adagrad)
--linesearch arg (=armijo) LBFGS linesearch (morethuente, wolfe,
armijo, strongwolfe)
--embeddings arg (=-1) use embeddings to initialize dictionary
(0=senna,1=turian,2=cldc)
--batches arg (=100) number batches (adagrad minibatch)
--ftcbatches arg (=100) number finetune batches (adagrad minibatch)
--initI arg (=1) initialize weight matrices to partial
identity?
--updateD1 arg (=1) learn D weights?
--updateF1 arg (=1) learn F weights?
--updateWd1 arg (=1) learn Wd weights?
--updateWl1 arg (=1) learn Wl weights?
--updateD2 arg (=1) learn D weights for B?
--updateF2 arg (=1) learn F weights for B?
--updateWd2 arg (=1) learn Wd weights for B?
--updateWl2 arg (=1) learn Wl weights for B?
--calc_rae_error1 arg (=0) consider the reconstruction error?
--calc_lbl_error1 arg (=0) consider the label error?
--calc_bi_error1 arg (=0) consider the bi error (matching root)?
--calc_thr_error1 arg (=0) consider the throughprop error?
--calc_uae_error1 arg (=0) consider the unfolding error?
--calc_rae_error2 arg (=0) consider the reconstruction error?
--calc_lbl_error2 arg (=0) consider the label error?
--calc_bi_error2 arg (=0) consider the bi error (matching root)?
--calc_thr_error2 arg (=0) consider the throughprop error?
--calc_uae_error2 arg (=0) consider the unfolding error?
--lambdaD arg (=1) L2 Regularization for Embeddings
--lambdaWd arg (=1) L2 Regularization for Tree Matrices
--lambdaBd arg (=1) L2 Regularization for Tree Biases
--lambdaWl arg (=1) L2 Regularization for Label Matrices
--lambdaBl arg (=1) L2 Regularization for Label Biases
--l1 arg (=0) L1 Regularization
--alpha arg (=0.200000003) autoencoder error vs label error
--gamma arg (=0.100000001) noisy error as percentage of normal bi-error
--epsilon arg (=9.99999997e-07) convergence parameter for LBFGS
--eta arg (=0.200000003) (initial) eta for SGD
--ftceta arg (=0.00999999978) (initial) eta for finetune AdaGrad
--norm arg (=0) normalization type (see train_update.cc)
-d [ --dynamic-mode ] arg (=0) type of sentence representation: 0
(root+avg), 1 (root), 2 (avg),
3(concat-all), 4 (complicated)
--noise arg (=2) number of noise samples per positive
training example
--hinge_loss_margin arg (=1) Hinge loss margin
Command line specific options:
-h [ --help ] print help message
-c [ --config ] arg config file specifying additional command
line options