Scikit-learn SVM 只有一个 class 异常

Question

I'm trying ensembling SVMs with Scikit-learn, specifically optimizing hyperparameters. I'm quite randomly getting the following error:

  File "C:\Users\jakub\anaconda3\envs\SVM_ensembles\lib\site-packages\sklearn\svm\_base.py", line 250, in _dense_fit
    self.probB_, self.fit_status_ = libsvm.fit(
  File "sklearn\svm\_libsvm.pyx", line 191, in sklearn.svm._libsvm.fit
ValueError: Invalid input - all samples with positive weights have the same label.

From what I understand, this is from file https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/svm/src/libsvm/svm.cpp and has something to do with examples from 1 class only going into SVM. I'm using stratified K-fold cross validation and have quite balanced dataset (45% one class, 55% other), so this should not happen anyway.

What can I do?

Optimizing code that throws error:

def get_best_ensemble_params(X_train, y_train, X_test, y_test, n_tries=5):
    search_spaces = {
        "max_samples": Real(0.1, 1, "uniform"),
        "max_features": Real(0.1, 1, "uniform"),

        "kernel": Categorical(["linear", "poly", "rbf", "sigmoid"]),
        "C": Real(1e-6, 1e+6, "log-uniform"),
        "gamma": Real(1e-6, 1e+1, "log-uniform")
    }

    best_accuracy = 0
    best_model = None
    for i in range(n_tries):
        done = False
        while not done:
            try:
                optimizer = BayesSearchCV(SVMEnsemble(), search_spaces, cv=3, n_iter=10, n_jobs=-1, n_points=10,
                                          verbose=1)
                optimizer.fit(X_train, y_train)  # <- ERROR HERE
                accuracy = accuracy_score(y_test, optimizer)
                if accuracy > best_accuracy:
                    best_accuracy = accuracy
                    best_model = optimizer
                done = True
                print(i, "job done")
            except:
                pass

    return best_model.best_params_


if __name__ == "__main__":
    dataset_name = "acute_inflammations"

    loading_functions = {
        "acute_inflammations": load_acute_inflammations,
        "breast_cancer_coimbra": load_breast_cancer_coimbra,
        "breast_cancer_wisconsin": load_breast_cancer_wisconsin
    }

    X, y = loading_functions[dataset_name]()
    X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=0)

    scaler = StandardScaler()
    X_train = scaler.fit_transform(X_train)
    X_test = scaler.fit_transform(X_test)

    params = get_best_ensemble_params(X_train, y_train, X_test, y_test)
    params["n_jobs"] = -1
    params["random_state"] = 0
    model = SVMEnsemble(n_estimators=20, **params)
    model.fit(X_train, y_train)
    y_pred = model.predict(X_test)

    print("Accuracy:", accuracy_score(y_test, y_pred))

My custom SVMEnsemble is just BaggingClassifier with hard-coded SVC :

import inspect
import numpy as np
from sklearn.ensemble import BaggingClassifier
from sklearn.svm import SVC
from skopt import BayesSearchCV


svm_possible_args = {"C", "kernel", "degree", "gamma", "coef0", "shrinking", "probability", "tol", "cache_size",
                     "class_weight", "max_iter", "decision_function_shape", "break_ties"}

bagging_possible_args = {"n_estimators", "max_samples", "max_features", "bootstrap", "bootstrap_features",
                         "oob_score", "warm_start", "n_jobs"}

common_possible_args = {"random_state", "verbose"}


class SVMEnsemble(BaggingClassifier):
    def __init__(self, voting_method="hard", n_jobs=-1,
                 n_estimators=10, max_samples=1.0, max_features=1.0,
                 C=1.0, kernel="linear", gamma="scale",
                 **kwargs):
        if voting_method not in {"hard", "soft"}:
            raise ValueError(f"voting_method {voting_method} is not recognized.")

        self._voting_method = voting_method
        self._C = C
        self._gamma = gamma
        self._kernel = kernel

        passed_args = {
            "n_jobs": n_jobs,
            "n_estimators": n_estimators,
            "max_samples": max_samples,
            "max_features": max_features,
            "C": C,
            "gamma": gamma,
            "cache_size": 1024,
        }


        kwargs.update(passed_args)

        svm_args = {
            "probability": True if voting_method == "soft" else False,
            "kernel": kernel
        }

        bagging_args = dict()

        for arg_name, arg_val in kwargs.items():
            if arg_name in svm_possible_args:
                svm_args[arg_name] = arg_val
            elif arg_name in bagging_possible_args:
                bagging_args[arg_name] = arg_val
            elif arg_name in common_possible_args:
                svm_args[arg_name] = arg_val
                bagging_args[arg_name] = arg_val
            else:
                raise ValueError(f"argument {voting_method} is not recognized.")

        self.svm_args = svm_args
        self.bagging_args = bagging_args

        base_estimator = SVC(**svm_args)
        super().__init__(base_estimator=base_estimator, **bagging_args)

    @property
    def voting_method(self):
        return self._voting_method

    @voting_method.setter
    def voting_method(self, new_voting_method):
        if new_voting_method == "soft":
            self._voting_method = new_voting_method
            self.svm_args["probability"] = True
            base_estimator = SVC(**self.svm_args)
            super().__init__(base_estimator=base_estimator, **self.bagging_args)
        elif self._voting_method == "soft":
            self._voting_method = new_voting_method
            self.svm_args["probability"] = False
            base_estimator = SVC(**self.svm_args)
            super().__init__(base_estimator=base_estimator, **self.bagging_args)
        else:
            self._voting_method = new_voting_method

    @property
    def C(self):
        return self._C

    @C.setter
    def C(self, new_C):
        self._C = new_C
        self.svm_args["C"] = new_C
        base_estimator = SVC(**self.svm_args)
        super().__init__(base_estimator=base_estimator, **self.bagging_args)

    @property
    def gamma(self):
        return self._gamma

    @gamma.setter
    def gamma(self, new_gamma):
        self._gamma = new_gamma
        self.svm_args["gamma"] = new_gamma
        base_estimator = SVC(**self.svm_args)
        super().__init__(base_estimator=base_estimator, **self.bagging_args)

    @property
    def kernel(self):
        return self._kernel

    @kernel.setter
    def kernel(self, new_kernel):
        self._kernel = new_kernel
        self.svm_args["kernel"] = new_kernel
        base_estimator = SVC(**self.svm_args)
        super().__init__(base_estimator=base_estimator, **self.bagging_args)

    def predict(self, X):
        if self._voting_method == "hard":
            return super().predict(X)
        elif self._voting_method == "soft":
            probabilities = np.zeros((X.shape[0], self.classes_.shape[0]))
            for estimator in self.estimators_:
                estimator_probabilities = estimator.predict_proba(X)
                probabilities += estimator_probabilities
            return self.classes_[probabilities.argmax(axis=1)]
        else:
            raise ValueError(f"voting_method {self._voting_method} is not recognized.")

Answer 1

From the way you describe your problem (that you are getting it "quite randomly") and the description of your data and the code I'm almost positive that the problem is with bagging classifier occasionally randomly selecting sub-sample of training examples with only one class. K-fold stratified split won't help you here because it only will control the original split(s) of your data into training/test, but not how BaggingClassifier picks random subsample of max_samples from training set. If you look at the code of how BaggingClassifier picks a subsample you'll notice there is no protection against such issue.

One very easy way to tell for sure is to replace the "max_samples": Real(0.1, 1, "uniform") with some smaller numbers eg "max_samples": Real(0.02, 0.03, "uniform") (or set to some fixed smaller value) and check that you start getting the error much more frequently.

I'm not sure if you really use it with n_tries=5 and n_iter=10 (seems a bit small for all the hyperparameters you have) or with larger values and/or maybe you run the whole script multiple times with different random seed, but in any case let's just compute the probability of getting such a problem with max_samples=0.1 and having a dataset with 120 examples with 55%/45% split. Let's say you got 96 examples for your training set with 45/55 split, eg 53+43 examples. Now with bootstrap enabled each time you train a bagging classifier it will randomly pick, say 10 examples out of 96 (with replacement since bootstrap is enabled by default). Chances of picking all of them from larger class are (53/96)^10, ie approximately 0.26%. That means that if you train 50 classifiers in a row like this chances of one of them hitting the issue are now 12.5%. And if you continuously run some searches like that you're pretty much inevitably running into this problem (I did use the fact that max_samples=0.1 here for simplicity which is not correct, but you'll likely to get close to that value frequently enough).

The last question is what to do with the issue. There are a few possible answers:

1. Ignore it - you get it randomly during the random search and no problem with all other attempts which didn't run into such issue. Additionally you can catch the ValueError exception and if the error message is coming from the SVM complaining on only class present for training - skip such search iteration.
2. Increase the minimum value for max_samples in your searches or make it dependent on number of examples.

There are also other possibilities too - eg after you split your data in train/test you can artificially inflate your training data by replacing each sample with eg N identical samples (where N is eg 2 or 10) to reduce the chance of having bagging classifier randomly picking subsample with only one class.