0

I have trained a classification model calling CatBoostClassifier.fit(), also providing an eval_set.

Now, how can I fetch the best value of the evaluation metric, and the number of iteration when it was achieved during training? I can plot the information by setting plot=True in the call to fit(), but how can I assign it to a variable?

I can do it when I train the model calling cv(), as cv() returns the wanted information. But CatBoostClassifier.fit() doesn't return anything, accordingly to the documentation.

Here the snippet of code I am using to fit the model:

model = CatBoostClassifier(
                           random_seed=42,
                           logging_level='Silent',
                           eval_metric='Accuracy'
                          )

model.fit(X_train,
          y_train,
          cat_features=cat_features_idxs,
          eval_set=(X_val, y_val),
          plot=True
         )

Here how I manage to fetch the wanted information, if I use cv() instead:

cv_data = cv(Pool(X, y, cat_features = cat_features_idxs),
             model.get_params(),
             fold_count = 5,
             plot=True)

print('Validation accuracy (best average among cross-validation folds) is {} obtained at step {}.'.format(np.max(cv_data['test-Accuracy-mean']), np.argmax(cv_data['test-Accuracy-mean'])))
Fanta
  • 2,825
  • 5
  • 26
  • 35

1 Answers1

1

1) Just compute the score on the training data:

https://stackoverflow.com/a/17954831

model = CatBoostClassifier(
                       random_seed=42,
                       logging_level='Silent',
                       eval_metric='Accuracy'
                      )

model.fit(X_train,
          y_train,
          cat_features=cat_features_idxs,
          eval_set=(X_val, y_val),
          plot=True
         )

train_score = model.score(X_train, y_train) # train (learn) score

val_score = model.score(X_val, y_val) # val (test) score

Another way would be accessing the output files:

model = CatBoostClassifier(
                       random_seed=42,
                       logging_level='Silent',
                       eval_metric='Accuracy',
                       allow_writing_files=True
                      )

model.fit(X_train,
      y_train,
      cat_features=cat_features_idxs,
      eval_set=(X_val, y_val),
      plot=True
     )

import pandas as pd
test_error = pd.read_csv('catboost_info/test_error.tsv', sep='\t')
val_score = test_error.loc[test_error['Accuracy'] == test_error['Accuracy'].max()]['Accuracy'].values[0]
best_iter = int(test_error.loc[test_error['Accuracy'] == test_error['Accuracy'].min()]['iter'].values[0])
train_score = learn_error.loc[learn_error['iter'] == best_iter]['Accuracy'].values[0]

2) If you have pandas installed add as_pandas=True as a parameter of cv, then you can access cv_data as a Dataframe. e.g. cv_data['test-Accuracy-mean'].max().

https://tech.yandex.com/catboost/doc/dg/concepts/python-reference_cv-docpage/

You could also access the output files as above, in this case there will be a pair of folders for each fold.

Hope this helps!

grezzo
  • 26
  • 2