Calculate pandas categorical hash once.

[glogiotatidis]

Build the pandas_categorical index hash when loading the model, instead of doing it on every predict().

Doing it on predict() significantly slows down the prediction, especially for larger categorical feature lists and wastes CPU cycles since this can be just calculated once.

After change:

Number of categorical features: 3
Biggest category length: 11945
Predict for 10.000 times: 0.3876569999847561 seconds

Before change:

Number of categorical features: 3
Biggest category length: 11945
Predict for 10.000 times: 15.50964699999895 seconds

Benchmark code:

    time = Benchmark.measure do
      x_test = [[3.7, 1.2,  "cat9", "100.8.6448"], [7.5, 0.5,  "cat0", "100.0.4262",]]
      booster = LightGBM::Booster.new(model_file: "test/support/categorical.txt")
      pandas_categorical = booster.instance_variable_get(:@pandas_categorical)
      puts "Number of categorical features: #{pandas_categorical.length}"
      puts "Biggest category length: #{pandas_categorical.max_by { |category| category.length }.length}"
      10000.times do
        y_pred = booster.predict(x_test)
      end
    end
    puts "Benchmark cached: #{time.real} seconds"

Added a large categorical feature in the test model with almost 12k entries.

[nunosilva800]

LGTM!

[glogiotatidis]

@ankane hey! could you please review this and cut a release if you think it's good to go?

[ankane]

Thanks @glogiotatidis, nice find!

[glogiotatidis]

Thanks for merging @ankane. Do you plan to release a patch version soon with this?

[ankane]

It'll probably be a few weeks.