Fix discriminative metric

[liyiersan]

I was trying to use the discriminative metric in tsgm following evaluation.ipynb. I found in most cases, the results would be very close to 0.5. Besides, there would be a warning about using softmax in one class classification and the training loss was abnormal.
I have checked the source codes of classification models in the zoo, and I find that all models utilize softmax activation before the output layer.

 m_output = layers.Dense(self._output_dim, activation="softmax")(x)

When setting the output_dim to 1 following evaluation.ipynb for classification models, the output would be all ones, in which the classification model would not work.

# use LSTM classification model from TSGM zoo.
model = tsgm.models.zoo["clf_cl_n"](
    seq_len=Xr.shape[1], feat_dim=Xr.shape[2], output_dim=1).model # set output_dim to 1 would not work for softmax
model.compile(
    tf.keras.optimizers.Adam(),
    tf.keras.losses.CategoricalCrossentropy(from_logits=False)
)

If directly set output_dim to 2, there would be a size error for metric calculation. I think this is due to the implementation of DiscriminativeMetric:

 	y_pred = (model.predict(X_test) > 0.5).astype(int)
    if metric is None:
        return sklearn.metrics.accuracy_score(y_test, y_pred)
    else:
        return metric(y_test, y_pred)

This version is suitable for binary classification with sigmoid activation. Therefore, I recommend modifying it so that it also works with softmax.

        # check the shape, 1D array or N-D arrary
        if len(pred.shape) == 1: # binary classification with sigmoid activation
            y_pred = (pred > 0.5).astype(int)
        else: # multiple classification with softmax activation
            y_pred = np.argmax(pred, axis=-1).astype(int)
        if metric is None:
            return sklearn.metrics.accuracy_score(y_test, y_pred)
        else:
            return metric(y_test, y_pred)

Besides, the definition of the classification model should be changed as follows:

# use LSTM classification model from TSGM zoo.
model = tsgm.models.zoo["clf_cl_n"](
    # set output_dim to 2 so that softmax can work properly
    seq_len=Xr.shape[1], feat_dim=Xr.shape[2], output_dim=2).model 
model.compile(
    tf.keras.optimizers.Adam(),
    # SparseCategoricalCrossentropy for multiple classes
    tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
)

[liyiersan]

In addition, function test_discriminative_metric() in test_metrics.py should be updated:

def test_discriminative_metric():
    ts = np.sin(np.arange(10)[:, None, None] + np.arange(6)[None, :, None])  # sin_sequence, [10, 6, 3]
    D1 = tsgm.dataset.Dataset(ts, y=None)

    diff_ts = np.sin(np.arange(10)[:, None, None] + np.arange(6)[None, :, None]) + 1000  # sin_sequence, [10, 6, 3]
    D2 = tsgm.dataset.Dataset(diff_ts, y=None)
    # output_dim =2 for softmax
    model = tsgm.models.zoo["clf_cl_n"](seq_len=ts.shape[1], feat_dim=ts.shape[2], output_dim=2).model
    model.compile(
        tf.keras.optimizers.Adam(),
        # SparseCategoricalCrossentropy for multiple classes
        tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
    )
    discr_metric = tsgm.metrics.DiscriminativeMetric()
    # should be easy to be classified 
    assert discr_metric(d_hist=D1, d_syn=D2, model=model, test_size=0.2, random_seed=42, n_epochs=5) == 1.0
    assert discr_metric(d_hist=D1, d_syn=D2, model=model, metric=sklearn.metrics.precision_score, test_size=0.2, random_seed=42, n_epochs=5) == 1.0