Skip to content

Latest commit

 

History

History
149 lines (92 loc) · 8.61 KB

README.md

File metadata and controls

149 lines (92 loc) · 8.61 KB
Raw

vwoptimize

Drop-in wrapper for Vowpal Wabbit that adds hyper-parameter tuning, more performance metrics, text preprocessing, reading from csv/tsv, feature extraction and k-fold cross-validation.

Hyper-parameter tuning

vwoptimize.py can automatically select the best hyper-parameters, by doing grid search for discrete and Nelder-Mead search for continuous parameters or by optimizing all parameters at once using hyperopt.

Using grid-search

In order to enable grid search, list possible values separated by a slash and append a question mark. For example,

$ python vwoptimize.py -d rcv1.train.vw  -b 22/24/26? --ngram 1/2? -f my.model

will try 6 configurations, select the one that gives the lowest progressive validation loss (reported by VW as average loss) and save the best model in "my.model" file.

Using Nelder-Mead

If there is no slash but there is a question mark, the parameter is treated as a float and fine-tuned using Nelder-Mead algorithm from scipy:

$ vwoptimize.py -d rcv1.train.vw -b 24 --ngram /2? --learning_rate 0.500? --l1 1e-07?

The number of digits after comma controls the precision of the tuner (if "0.500?" is specified then "0.500" and "0.501" might be tried but not "0.5005"). If the number is written in scientific notation ("1e-07?") then the search is done in log-space.

Using hyperopt

Adding "--hyperopt N" enables optimization using hyperopt for N rounds. For that to work, one need to provide boundaries for each parameter. For example,

$ vwoptimize.py -d rcv1.train.vw -b 24 --ngram 1..3? --learning_rate 0.100..5.000? --l1 1e-11..1e-2?

In order to select optimization algorithm, use --hyperopt_alg ALG where ALG can be "tpe" or "rand" or "package_name.module_name.function_name" for custom implementation of hyperopt's "suggest" method.

Specifying metric to optimize

By default, vwoptimize.py reads the loss reported by Vowpal Wabbit and uses that as an optimization objective. It is also possible to specify custom metrics. For example, this will try different loss functions and select the one that gives the best accuracy:

$ vwoptimize.py -d data.vw -b 28 --loss_function squared/hinge/logistic? --metric acc

In this case, vwoptimize.py will ask vowpal wabbit to store predictions in a temporary file and then use them together with true values extracted from data.vw to calculate the required metrics.

Some other metrics that can be used there: f1, brier, auc, logloss.

If the value of --metric is a comma-separated list, then only the first one is used as optimization objective and others are just printed for information. For example,

$ vwoptimize.py -d data.vw -b 28 --loss_function squared/hinge/logistic? --metric acc,precision,recall,vw_average_loss

will perform the same optimization as previous but also report extra metrics for each run.

Cross-validation

When doing multiple passes over data, the metrics reported by --metric are no longer suitable for tuning (vwoptimize.py does not automatically switches to using holdout set like VW itself does and thus ends up using predictions over already seen examples). K-fold cross validation avoids that by explicitly separating training and testing sets:

$ vwoptimize.py -d data.vw -b 28 --loss_function squared/hinge/logistic? --kfold 5 -c -k --passes 1/2/5? --metric acc

This will calculate mean test-set accuracy over 5 different runs, each time using 80% of data for training and 20% for testing.

The --kfold option will not shuffle the dataset and will always use the same split.

Using vwoptimize.py for model evaluation

The --metric option can be used without the optimizer, in a regular run:

$ vwoptimize.py -d data.vw -i model -t --metric precision,recall,acc,auc

Preprocessing the input

Handling CSV/TSV inputs

vwoptimize.py can work with CSV/TSV files. One need to specify --columnspec that describes how to interpret each column:

$ vwoptimize.py -d data.csv --columnspec y,weight,text_s,vw_w --ignoreheader --tovw /dev/stdout --lowercase --strip_punct
1 10 |s hello |w YEAR:2017

$ cat data.csv
label,weight,message
1,10,Hello!,YEAR:2017

In this example,

  • --ignoreheader tells vwoptimize.py to skip the first line of the input
  • --columnspec y,weight,text_s,vw_w describes how to interpret the 4 columns found in the csv file:
    • y means first column is used as a label
    • weight means second column is used as example weight
    • text_s means the third column is interpreted as text and put into namespace "s".
    • vw_w means the fourth column is interpreted as raw vowpal wabbit format and put into namespace "w". The difference between vw an text is that text has : and | removed from it as well as preprocessor options applied (--lowercase --strip_punct in this case), while vw is copied as is.
  • Finally, --tovw /dev/stdout tells to the preprocessed file to stdout and exit (rather than start training or tuning on the preprocessed input).

vwoptimize.py can recognize .csv, .tsv, .csv.gz, .tsv.gz and load them accordingly.

Other useful --columnspec values:

  • drop or empty string will ignore the field
  • info means put this message into tag section of .vw format
  • weight_train is like --weight but it is not taken into account when calculating the metrics with --metric (only affects training)
  • weight_metric is like --weight but it does not affect training (only affects --metric)

Setting class weights

$ vwoptimize.py -d data.vw --weight 1:0.5,2:0.9

This will multiply the weights of the examples that have label "1" by 0.5 and the examples with label "2" by 0.9. When used like this, it also affects the weights used when calculating the metrics.

The --weight_train option only affects the weights passed to VW but does not affect how the metrics are calculated. The --weight_metric does the opposite: it has not effect on training but it is used as sample weight when calculating metrics.

UPDATE: In recent VW, --classweight is equivalent to --weight_train.

Text processing

The following text processing options are available:

  • --lowercase Lowercase the text.
  • --strip_punct Strip punctuation.
  • --stem Stem each word. Requires NLTK and pycld2.
  • --split_chars Insert spaces between all characters.
  • --max_word_size=MAX_WORD_SIZE Limit the word size to MAX_WORD_SIZE characters
  • --max_words=MAX_WORDS Only keep the first MAX_WORDS words (applies individually on each column or namespaces)
  • --max_length=MAX_LENGTH Only keep the first MAX_LENGTH characters
  • --max_length_offset=MAX_LENGTH_OFFSET Ignore the first MAX_LENGTH_OFFSET characters
  • --htmlunescape Decode HTML-encoded characters (& -> &)
  • --NFKC Normalize unicode characters
  • --chinese_simplify Convert Traditional Chinese characters into Simplified Chinese
  • --split_ideographs / --split_hangul / --split_hiragana / --split_katakana Insert spaces between the characters of the corresponding unicode script
  • --split_combined Combines the four split options above

Saving & loading configuration

$ vwoptimize.py -d data.csv --columnspec y,weight,text_s,vw_w --ignoreheader --lowercase --strip_punct --writeconfig my.config -f my.model

This will save the relevant preprocessing options (--columnspec y,weight,text_s,vw_w --lowercase --strip_punct) into my.config file, so that they don't have to be specified when using the model to get the predictions:

$ vwoptimize.py -d test_data.csv --readconfig my.config -t -p predictions.txt --metric precision,recall

It is especially important if the best preprocessing options are selected automatically. For example:

$ vwoptimize.py -d data.csv --columnspec y,weight,text_s,vw_w --ignoreheader --lowercase? --strip_punct? --writeconfig my.config -f my.model

will grid search for best configuration of --lowercase & --strip_punct, use that to train the final model and store those settings in my.config so that the same transformations can be applied during prediction time.