Clone wiki

sequor / Options

Sequor implements the following options for training models.

--rate=NUM Learning rate. A constant scaling the update to the weight vector. Use smaller learning weight if accuracy on heldout data oscillates.
--beam=INT Beam size. Sequor uses beam decoding instead of Viterbi decoding. This lets you trade off some accuracy against speed of training.
--iter=INT Number of iterations. Training will stop after that many iterations of weight updates.
--min-count=INT
 Minimum feature frequency for label dictionary. A label dictionary associates a features marked with Index (see https://bitbucket.org/gchrupala/sequor/wiki/Templates) with a list of all the labels it co-occurs with in the training data. For prediction, only the labels in this list are considered for an example with this feature. An index feature should typically be a word form or equivalent. This option lets you specify the minimum frequency an index feature should occur in order to be included in the label dictionary. During prediction at a particular position, if no index feature is found, or if the index feature is not included in the dictionary, the full set of label is considered. Do not mark any features with Index in order to disable the use of label dictionary.
--heldout=FILE Path to heldout data. This file should contain additional example which will not be used for training. Accuracy on this data will be calculated and printed during training.
--hash Use hashing instead of feature dictionary. By default sequor creates a feature dictionary where each feature is assigned an index, which can be used to look up the feature's weight in the weight vector. As an alternative sequor can use a hash function to create feature indices. This saves memory, but introduces potential hashing conflicts, which may in turn impact performance. Use this option to reduce memory usage.
--hash-sample=INT
 Sample size to estimate number of features when hashing. This option controls how many examples to use in order to estimate the number of features in the training data and decide on the size of the weight vector. This option is only used with feature hashing.
--hash-max-size=INT
 Maximum size of parameter vector when hashing. You can set the size of the weight (parameter) vector explicitly rather than based on an estimate of the number of features. This option is only used with feature hashing.

Updated