Skip to content

Latest commit

 

History

History
13 lines (10 loc) · 1.96 KB

EXAMPLE.MD

File metadata and controls

13 lines (10 loc) · 1.96 KB
Raw

Example that uses manual index

This example explores the newest features of StackNet (version 1.04). Namely it looks at input_index, inlcude_target and includes models from libfm, libffm and vowpal wabbit. The input_index expects a file (index_file.txt) that specifies the cv fold of each row. This file needs to have the same row size as the training file. It excludes headers. The inlcude_target=true command just appends the target variable in the holdout prediction files if the output_name command is used.

To run follow the next steps:

  1. Ensure you have train.csv, test.csv, params.txt* and index_file.txt in the same folder with StackNet.jar and the lib/ folder. The .csv files have dense format, they include headers and the target variable in the beginning
  2. You need to follow the instructions to make certain that the executables for the algorithms are working. Specifically, you need to check libfm, vowpal wabbit, libffm. Please don't skip this step. Most errors are due to the fact that the binaries are not working properly.
  3. Make certain you have Java higher than 1.6 installed and that Java is in your PATH. Have a look at [this] (https://www.java.com/en/download/help/path.xml) if you encounter trouble.
  4. Run StackNet with java -jar StackNet.jar train task=regression train_file=train.csv test_file=test.csv params=params.txt pred_file=pred.csv output_name=models input_index=index_file.txt bin=2 include_target=true test_target=true has_head=true verbose=true Threads=1 folds=4 metric=rmse . Note there folds=4, however in index_file.txt there are 5 folds. 5 Folds will be used, the folds=4 is ignored.
  5. The holdout predictions (in models1.csv) inlcude an extra column with the target in the beginning. 7 columns in total.