14a9d386baf93b8266fa21277b40d669a4e397d2,pyannote/audio/embedding/models.py,TristouNet,__call__,#TristouNet#Any#,83

Before Change

                if self.bidirectional:
                    // first backward LSTM needs to be given the input shape
                    // AND to be told to process the sequence backward
                    backward = LSTM(name="backward_{i:d}".format(i=i),
                                    go_backwards=True,
                                    input_shape=input_shape,
                                    output_dim=output_dim,
                                    return_sequences=return_sequences,
                                    activation="tanh",
                                    dropout_W=0.0,
                                    dropout_U=0.0)(x)

        if self.pooling == "average":
            forward = GlobalAveragePooling1D()(forward)
            if self.bidirectional:
                backward = GlobalAveragePooling1D()(backward)

        // concatenate forward and backward
        if self.bidirectional:
            x = merge([forward, backward], mode="concat", concat_axis=1)
        else:
            x = forward

        // stack dense layers

After Change

            x = mlp(x)

        // stack L2 normalization layer
        normalize = Lambda(lambda x: K.l2_normalize(x, axis=-1),
                           name="normalize")
        embeddings = normalize(x)

        return Model(input=inputs, output=embeddings)

Instances