b903f8ec84d21253a1eebff9b29519a9e934f254,examples/tox21/train_tox21.py,,main,#,40

Before Change

                                repeat=False, shuffle=False)

    accfun_mode = args.accfun_mode
    if accfun_mode == 0:
        accfun = F.binary_accuracy
    elif accfun_mode == 1:
        from sklearn import metrics

        def get_1d_numpy_array(v):
            if isinstance(v, chainer.Variable):
                v = v.data
            return cuda.to_cpu(v).ravel()

        def accfun(y, t):
            // -- calc & report ROC-AUC ---
            // note that this is dirty hack implementation.
            // roc auc is calculated per minibatch, and mean is taken to show
            // PrintReport. This calculation is not same as total batch roc auc
            // calculation.
            t_data = get_1d_numpy_array(t)
            y_data = get_1d_numpy_array(y)

            y_data = y_data[t_data != -1]
            t_data = t_data[t_data != -1]
            try:
                roc_auc = metrics.roc_auc_score(t_data, y_data)
                reporter.report({"roc_auc": roc_auc}, classifier)
            except ValueError as e:
                // When `t_data` only contains one label (ex. only 0), roc auc
                // cannot be calculated and ValueError is raised.
                // This implementation just ignores this minibatch for roc auc
                // calculation.
                pass
            // --- calc ROC-AUC end ---
            return F.binary_accuracy(y, t)
    else:
        raise ValueError("Invalid accfun_mode {}".format(accfun_mode))

    classifier = L.Classifier(predictor_,
                              lossfun=F.sigmoid_cross_entropy,
                              accfun=accfun)
    if args.gpu >= 0:

After Change

        trainer.extend(ROCAUCEvaluator(
            train_eval_iter, classifier, predictor=predictor_,
            device=args.gpu, converter=concat_mols, name="train"))
        trainer.extend(ROCAUCEvaluator(
            val_iter, classifier, predictor=predictor_,
            device=args.gpu, converter=concat_mols, name="val"))
        trainer.extend(E.PrintReport([
            "epoch", "main/loss", "main/accuracy", "train/roc_auc",
            "validation/main/loss", "validation/main/accuracy",
            "val/roc_auc", "elapsed_time"]))

Instances