This project uses machine learning to predict what a user will type in a text editor.
-
Pull dependencies and compile:
./pull-dependencies # Downloads all libraries and datasets to lib (don't need to do frequently) make # Compiles everything -
Add the following line to your
.vimrcfile:source <absolute path of this directory>/plugin/smartautocomplete.vim -
Start the server:
./run -inPaths <paths>The server will scan all the files in the given paths, and enable autocompletion for files in these paths.
Now when you start vim on any file, and press Ctrl+P in insert mode, vim will call the server for the autocomplete suggestions rather than vim's own. What you end up selecting will be also sent to the server as additional training data so that autocomplete will get better over time.
Formally, the problem is to perform online learning to obtain a distribution over possible output completion y given the input context x.
-
Input context x contains
- Names and contents of all files in the project.
- The current file name.
- The current position in the current file.
-
Output completion y: a sequence of characters to be inserted at the current position in the current file.
To run a benchmark on the d3 dataset, execute the following command:
./run @mode=tune @data=d3
Results of the experiment will be put into a directory called state in the current directory. The location of the state directory can be modified by adding a @state= option to ./run.
Once you have run an experiment, you can visualize the results by starting a web server:
java -cp classes:lib/* smartAutocomplete.httpServer.StandaloneHttpServer state
Then go to localhost:8040/exec/<id> in a browser to visualize the experiment in directory <id>.exec. This doesn't work on every file, so if you don't see any colors when you click on a file, try another file. You can click on tokens to see information on how the classifier performed, and click on feature names to see details of how the feature made its prediction.
