How to use the thinc.v2v.Maxout function in thinc

>> zero_init(Affine(nr_class, width, drop_factor=0.0))
                >> logistic
            )
            return model

        lower = HashEmbed(width, nr_vector, column=1)
        prefix = HashEmbed(width // 2, nr_vector, column=2)
        suffix = HashEmbed(width // 2, nr_vector, column=3)
        shape = HashEmbed(width // 2, nr_vector, column=4)

        trained_vectors = FeatureExtracter(
            [ORTH, LOWER, PREFIX, SUFFIX, SHAPE, ID]
        ) >> with_flatten(
            uniqued(
                (lower | prefix | suffix | shape)
                >> LN(Maxout(width, width + (width // 2) * 3)),
                column=0,
            )
        )

        if pretrained_dims:
            static_vectors = SpacyVectors >> with_flatten(
                Affine(width, pretrained_dims)
            )
            # TODO Make concatenate support lists
            vectors = concatenate_lists(trained_vectors, static_vectors)
            vectors_width = width * 2
        else:
            vectors = trained_vectors
            vectors_width = width
            static_vectors = None
        tok2vec = vectors >> with_flatten(

def build_model(nr_class, width, depth, conv_depth, vectors_name, **kwargs):
    with Model.define_operators({"|": concatenate, ">>": chain, "**": clone}):
        embed = (
            HashEmbed(width, 5000, column=1)
            | StaticVectors(vectors_name, width, column=5)
            | HashEmbed(width // 2, 750, column=2)
            | HashEmbed(width // 2, 750, column=3)
            | HashEmbed(width // 2, 750, column=4)
        ) >> LN(Maxout(width))

        sent2vec = (
            with_flatten(embed)
            >> Residual(
                prepare_self_attention(Affine(width*3, width), nM=width, nH=4)
                >> MultiHeadedAttention()
                >> with_flatten(Maxout(width, width, pieces=3))
            )
            >> flatten_add_lengths
            >> ParametricAttention(width, hard=False)
            >> Pooling(mean_pool)
            >> Residual(LN(Maxout(width)))
        )

        model = (
            foreach(sent2vec, drop_factor=2.0)
            >> Residual(
                prepare_self_attention(Affine(width*3, width), nM=width, nH=4)
                >> MultiHeadedAttention()
                >> with_flatten(LN(Affine(width, width)))
            )
            >> flatten_add_lengths
            >> ParametricAttention(width, hard=False)

cfg = dict(locals())
    print(cfg)
    train_data, check_data, nr_tag = ancora_pos_tags()
    train_data = list(train_data)
    check_data = list(check_data)

    extracter = FeatureExtracter("es", attrs=[LOWER, SHAPE, PREFIX, SUFFIX])
    with Model.define_operators({"**": clone, ">>": chain, "+": add, "|": concatenate}):
        lower_case = HashEmbed(width, 100, column=0)
        shape = HashEmbed(width // 2, 200, column=1)
        prefix = HashEmbed(width // 2, 100, column=2)
        suffix = HashEmbed(width // 2, 100, column=3)

        model = (
            with_flatten(
                (lower_case | shape | prefix | suffix) >> Maxout(width, pieces=3)
            )
            >> PyTorchBiLSTM(width, width, depth)
            >> with_flatten(Softmax(nr_tag))
        )

    train_X, train_y = preprocess(model.ops, extracter, train_data, nr_tag)
    dev_X, dev_y = preprocess(model.ops, extracter, check_data, nr_tag)

    n_train = float(sum(len(x) for x in train_X))
    global epoch_train_acc
    with model.begin_training(train_X[:10], train_y[:10], **cfg) as (
        trainer,
        optimizer,
    ):
        trainer.each_epoch.append(track_progress(**locals()))
        trainer.batch_size = min_batch_size

)
        )

        if pretrained_dims:
            static_vectors = SpacyVectors >> with_flatten(
                Affine(width, pretrained_dims)
            )
            # TODO Make concatenate support lists
            vectors = concatenate_lists(trained_vectors, static_vectors)
            vectors_width = width * 2
        else:
            vectors = trained_vectors
            vectors_width = width
            static_vectors = None
        tok2vec = vectors >> with_flatten(
            LN(Maxout(width, vectors_width))
            >> Residual((ExtractWindow(nW=1) >> LN(Maxout(width, width * 3)))) ** depth,
            pad=depth,
        )
        cnn_model = (
            tok2vec
            >> flatten_add_lengths
            >> ParametricAttention(width)
            >> Pooling(sum_pool)
            >> Residual(zero_init(Maxout(width, width)))
            >> zero_init(Affine(nr_class, width, drop_factor=0.0))
        )

        linear_model = build_bow_text_classifier(
            nr_class, ngram_size=cfg.get("ngram_size", 1), exclusive_classes=False
        )
        if cfg.get("exclusive_classes"):

# This is annoying, but the parser etc have the flatten step after
    # the tok2vec. To load the weights in cleanly, we need to match
    # the shape of the models' components exactly. So what we cann
    # "tok2vec" has to be the same set of processes as what the components do.
    with Model.define_operators({">>": chain, "|": concatenate}):

        l2r_model = (
            tok2vec.l2r
            >> flatten
            >> LN(Maxout(output_size, tok2vec.l2r.nO, pieces=3))
            >> zero_init(Affine(output_size, drop_factor=0.0))
        )
        r2l_model = (
            tok2vec.r2l
            >> flatten
            >> LN(Maxout(output_size, tok2vec.r2l.nO, pieces=3))
            >> zero_init(Affine(output_size, drop_factor=0.0))
        )

        model = tok2vec.embed >> (l2r_model | r2l_model)

    model.tok2vec = tok2vec
    model.begin_training([nlp.make_doc("Give it a doc to infer shapes")])
    tok2vec.begin_training([nlp.make_doc("Give it a doc to infer shapes")])
    tokvecs = tok2vec([nlp.make_doc('hello there'), nlp.make_doc(u'and hello')])
    print(tokvecs.shape)
    return model

def _encoder(in_width, hidden_with, end_width):
        conv_depth = 2
        cnn_maxout_pieces = 3

        with Model.define_operators({">>": chain, "**": clone}):
            convolution = Residual((ExtractWindow(nW=1) >>
                                    LN(Maxout(hidden_with, hidden_with * 3, pieces=cnn_maxout_pieces))))

            encoder = SpacyVectors \
                      >> with_flatten(LN(Maxout(hidden_with, in_width)) >> convolution ** conv_depth, pad=conv_depth) \
                      >> flatten_add_lengths \
                      >> ParametricAttention(hidden_with)\
                      >> Pooling(mean_pool) \
                      >> Residual(zero_init(Maxout(hidden_with, hidden_with))) \
                      >> zero_init(Affine(end_width, hidden_with, drop_factor=0.0))

            # TODO: ReLu or LN(Maxout)  ?
            # sum_pool or mean_pool ?

        return encoder

# (f ** 3)(x) -> f''(f'(f(x))), where f, f' and f'' have distinct weights.
    # * concatenate (|): Merge the outputs of two models into a single vector,
    # i.e. (f|g)(x) -> hstack(f(x), g(x))
    Model.lsuv = True
    # Model.ops = CupyOps()
    with Model.define_operators({">>": chain, "**": clone, "|": concatenate, "+": add}):
        mwe_encode = ExtractWindow(nW=1) >> LN(
            Maxout(width, drop_factor=0.0, pieces=pieces)
        )

        sent2vec = (
            flatten_add_lengths
            >> with_getitem(
                0,
                (HashEmbed(width, 3000) | StaticVectors("en_vectors_web_lg", width))
                >> LN(Maxout(width, width * 2))
                >> Residual(mwe_encode) ** depth,
            )  # : word_ids{T}
            >> Pooling(mean_pool, max_pool)
            >> Residual(LN(Maxout(width * 2, pieces=pieces), nO=width * 2)) ** 2
            >> logistic
        )
        model = Siamese(sent2vec, CauchySimilarity(width * 2))

    print("Read and parse data: %s" % dataset)
    if dataset == "quora":
        train, dev = datasets.quora_questions()
    elif dataset == "snli":
        train, dev = datasets.snli()
    elif dataset == "stackxc":
        train, dev = datasets.stack_exchange()
    elif dataset in ("quora+snli", "snli+quora"):

def create_pretraining_model(nlp, tok2vec):
    """Define a network for the pretraining. We simply add an output layer onto
    the tok2vec input model. The tok2vec input model needs to be a model that
    takes a batch of Doc objects (as a list), and returns a list of arrays.
    Each array in the output needs to have one row per token in the doc.
    """
    output_size = nlp.vocab.vectors.data.shape[1]
    output_layer = chain(
        LN(Maxout(300, pieces=3)), Affine(output_size, drop_factor=0.0)
    )
    # This is annoying, but the parser etc have the flatten step after
    # the tok2vec. To load the weights in cleanly, we need to match
    # the shape of the models' components exactly. So what we cann
    # "tok2vec" has to be the same set of processes as what the components do.
    tok2vec = chain(tok2vec, flatten)
    model = chain(tok2vec, output_layer)
    model = masked_language_model(nlp.vocab, model)
    model.tok2vec = tok2vec
    model.output_layer = output_layer
    model.begin_training([nlp.make_doc("Give it a doc to infer shapes")])
    return model

column=cols.index(ORTH),
            )
        elif char_embed:
            embed = concatenate_lists(
                CharacterEmbed(nM=64, nC=8),
                FeatureExtracter(cols) >> with_flatten(norm),
            )
            reduce_dimensions = LN(
                Maxout(width, 64 * 8 + width, pieces=cnn_maxout_pieces)
            )
        else:
            embed = norm

        convolution = Residual(
            ExtractWindow(nW=1)
            >> LN(Maxout(width, width * 3, pieces=cnn_maxout_pieces))
        )
        if char_embed:
            tok2vec = embed >> with_flatten(
                reduce_dimensions >> convolution ** conv_depth, pad=conv_depth
            )
        else:
            tok2vec = FeatureExtracter(cols) >> with_flatten(
                embed >> convolution ** conv_depth, pad=conv_depth
            )

        if bilstm_depth >= 1:
            tok2vec = tok2vec >> PyTorchBiLSTM(width, width, bilstm_depth)
        # Work around thinc API limitations :(. TODO: Revise in Thinc 7
        tok2vec.nO = width
        tok2vec.embed = embed
    return tok2vec