def build_cloze_characters_multi_task_model(
vocab: "Vocab", tok2vec: Model, maxout_pieces: int, hidden_size: int, nr_char: int
) -> Model:
output_layer = chain(
list2array(),
Maxout(nO=hidden_size, nP=maxout_pieces),
LayerNorm(nI=hidden_size),
MultiSoftmax([256] * nr_char, nI=hidden_size),
)
model = build_masked_language_model(vocab, chain(tok2vec, output_layer))
model.set_ref("tok2vec", tok2vec)
model.set_ref("output_layer", output_layer)
return model
def build_cloze_multi_task_model(
vocab: "Vocab", tok2vec: Model, maxout_pieces: int, hidden_size: int
) -> Model:
nO = vocab.vectors.data.shape[1]
output_layer = chain(
list2array(),
Maxout(
nO=hidden_size,
nI=tok2vec.get_dim("nO"),
nP=maxout_pieces,
normalize=True,
dropout=0.0,
),
Linear(nO=nO, nI=hidden_size, init_W=zero_init),
)
model = chain(tok2vec, output_layer)
model = build_masked_language_model(vocab, model)
model.set_ref("tok2vec", tok2vec)
model.set_ref("output_layer", output_layer)
return model
def build_tb_parser_model(
tok2vec: Model[List[Doc], List[Floats2d]],
state_type: Literal["parser", "ner"],
extra_state_tokens: bool,
hidden_width: int,
maxout_pieces: int,
use_upper: bool,
nO: Optional[int] = None,
) -> Model:
"""
Build a transition-based parser model. Can apply to NER or dependency-parsing.
Transition-based parsing is an approach to structured prediction where the
task of predicting the structure is mapped to a series of state transitions.
You might find this tutorial helpful as background:
https://explosion.ai/blog/parsing-english-in-python
The neural network state prediction model consists of either two or three
subnetworks:
* tok2vec: Map each token into a vector representations. This subnetwork
is run once for each batch.
* lower: Construct a feature-specific vector for each (token, feature) pair.
This is also run once for each batch. Constructing the state
representation is then simply a matter of summing the component features
and applying the non-linearity.
* upper (optional): A feed-forward network that predicts scores from the
state representation. If not present, the output from the lower model is
used as action scores directly.
tok2vec (Model[List[Doc], List[Floats2d]]):
Subnetwork to map tokens into vector representations.
state_type (str):
String value denoting the type of parser model: "parser" or "ner"
extra_state_tokens (bool): Whether or not to use additional tokens in the context
to construct the state vector. Defaults to `False`, which means 3 and 8
for the NER and parser respectively. When set to `True`, this would become 6
feature sets (for the NER) or 13 (for the parser).
hidden_width (int): The width of the hidden layer.
maxout_pieces (int): How many pieces to use in the state prediction layer.
Recommended values are 1, 2 or 3. If 1, the maxout non-linearity
is replaced with a ReLu non-linearity if use_upper=True, and no
non-linearity if use_upper=False.
use_upper (bool): Whether to use an additional hidden layer after the state
vector in order to predict the action scores. It is recommended to set
this to False for large pretrained models such as transformers, and True
for smaller networks. The upper layer is computed on CPU, which becomes
a bottleneck on larger GPU-based models, where it's also less necessary.
nO (int or None): The number of actions the model will predict between.
Usually inferred from data at the beginning of training, or loaded from
disk.
"""
if state_type == "parser":
nr_feature_tokens = 13 if extra_state_tokens else 8
elif state_type == "ner":
nr_feature_tokens = 6 if extra_state_tokens else 3
else:
raise ValueError(Errors.E917.format(value=state_type))
t2v_width = tok2vec.get_dim("nO") if tok2vec.has_dim("nO") else None
tok2vec = chain(
tok2vec,
list2array(),
Linear(hidden_width, t2v_width),
)
tok2vec.set_dim("nO", hidden_width)
lower = _define_lower(
nO=hidden_width if use_upper else nO,
nF=nr_feature_tokens,
nI=tok2vec.get_dim("nO"),
nP=maxout_pieces,
)
upper = None
if use_upper:
with use_ops("cpu"):
# Initialize weights at zero, as it's a classification layer.
upper = _define_upper(nO=nO, nI=None)
return TransitionModel(tok2vec, lower, upper, resize_output)