I am trying to implement custom objective function for multiclassification following the guides in: https://catboost.ai/docs/concepts/python-usages-examples.html#custom-objective-function
I would like to know what is the expected format of the gradient and the Hessian matrix defined in calc_ders_multi().
To see how it works, I tried to reproduce the MultiClass loss function, but with defined gradient and Hessian matrix the program starts and gets stuck without throwing any errors. If I don't define a Hessian (return []), the program runs without problems.
import numpy as np
from catboost import Pool, CatBoostClassifier
class MultiClassObjective(object):
def calc_ders_multi(self, approxes, targets, weights):
# approxes, targets, weights are indexed containers of floats
# (containers with only __len__ and __getitem__ defined).
# weights parameter can be None.
n_classes = len(approxes)
exponents = np.exp(approxes)
softmax = exponents / np.sum(exponents)
targets = np.array([1 if x == targets else 0 for x in range(n_classes)])
der1 = targets - softmax
der2 = np.zeros((n_classes, n_classes))
for x in range(n_classes):
for y in range(n_classes):
if x == y:
der2[x, y] = -softmax[x] * (1 - softmax[y])
else:
der2[x, y] = softmax[x] * softmax[y]
if weights is not None:
der1 *= weights
der2 *= weights
return der1, der2
# return der1, []
train_data = [[1, 4, 5, 6],
[4, 5, 6, 7],
[11, 20, 30, 30],
[30, 40, 50, 60],
[100, 300, 200, 400]]
train_labels = [0, 0, 1, 1, 2]
train_data = Pool(data=train_data , label=train_labels )
# Initialize CatBoostClassifier with custom `loss_function`
model = CatBoostClassifier(loss_function=MultiClassObjective(),
eval_metric="MultiClass",
iterations=100,
random_seed=0,
verbose=True
)
# Fit model
model.fit(train_data)
