How to implement record_evaluation callback in LightGBM python?

Question

I am using the model = lgb.train function. When trying to plot the evaluation metric against epochs of a LightGBM model (i.e., lgb.plot_metric(model)) I get the following error:

TypeError: booster must be dict or LGBMModel. To use plot_metric with Booster type, first record the metrics using record_evaluation callback then pass that to plot_metric as argument booster

But I cannot find any info in the documentation about parameters to set up the mentioned callback. Is there any way to implement this without resorting to the scikit-learn version of LightGBM?

Answer 1

The following should help to plot the metrics. I guess the documentation isn't really clear on usage but here is an example notebook. The evals dictionary contains an OrderedDict and can be plotted using the plot_metric method.

train_dt = lgb.Dataset(data=train,label=train_y)
valid_dt = lgb.Dataset(data = valid, 
                       label=valid_y, 
                       reference=train_dt)

params = {
        'objective': 'regression',
        'metric': 'root_mean_squared_error',
        'num_leaves': 41,
}
evals={}
mod = lgb.train(params=params, 
                train_set = train_dt, 
                valid_sets=[train_dt, valid_dt],
                callbacks = [lgb.record_evaluation(evals)])

lgb.plot_metric(evals)