Source code for finetune.target_models.association

import itertools
from collections import Counter

import tensorflow as tf
import numpy as np

from finetune.base import BaseModel
from finetune.encoding.target_encoders import SequenceLabelingEncoder
from finetune.nn.target_blocks import association
from finetune.nn.crf import sequence_decode
from finetune.encoding.sequence_encoder import indico_to_finetune_sequence, finetune_to_indico_sequence
from finetune.input_pipeline import BasePipeline
from finetune.errors import FinetuneError
from finetune.base import LOGGER
from finetune.model import PredictMode


class AssociationPipeline(BasePipeline):
    def __init__(self, config, multi_label):
        super(AssociationPipeline, self).__init__(config)
        self.multi_label = multi_label
        self.association_encoder = SequenceLabelingEncoder()
        self.association_encoder.fit(config.association_types + [self.config.pad_token])
        self.association_pad_idx = self.association_encoder.transform([self.config.pad_token])

    def _post_data_initialization(self, Y):
        Y_ = list(itertools.chain.from_iterable([y[0] for y in Y]))
        super()._post_data_initialization(Y_)

    def text_to_tokens_mask(self, X, Y=None):
        pad_token = [self.config.pad_token] if self.multi_label else self.config.pad_token
        if Y is not None:
            Y = list(zip(*Y))
        out_gen = self._text_to_ids(X, Y=Y, pad_token=(pad_token, pad_token, -1, -2))
        class_list = self.association_encoder.classes_.tolist()
        assoc_pad_id = class_list.index(pad_token)
        for out in out_gen:
            feats = {"tokens": out.token_ids, "mask": out.mask}
            if Y is None:
                yield feats
            else:
                labels = []
                assoc_mat = [[assoc_pad_id for _ in range(len(out.labels))] for _ in range(len(out.labels))]
                for i, (l, _, _, idx) in enumerate(out.labels):
                    labels.append(l)
                    for j, (_, a_t, a_i, _) in enumerate(out.labels):
                        if a_t != pad_token and idx == a_i:
                            assoc_mat[i][j] = class_list.index(a_t)

                yield feats, {"labels": self.label_encoder.transform(labels),
                              "associations": np.array(assoc_mat, dtype=np.int32)}

    def _compute_class_counts(self, encoded_dataset):
        counter = Counter()
        for doc, target_arr in encoded_dataset:
            targets = target_arr['labels'][doc['mask'].astype(np.bool)]
            counter.update(
                self.label_encoder.inverse_transform(targets)
            )
        return counter

    def _format_for_encoding(self, X):
        return [X]

    def _format_for_inference(self, X):
        return [[x] for x in X]

    def feed_shape_type_def(self):
        TS = tf.TensorShape
        target_shape = (
            [self.config.max_length, self.label_encoder.target_dim]
            if self.multi_label else [self.config.max_length]
        )
        return (
            (
                {
                    "tokens": tf.int32,
                    "mask": tf.float32
                },
                {
                    "labels": tf.int32,
                    "associations": tf.int32
                }
            ),
            (
                {
                    "tokens": TS([self.config.max_length, 2]),
                    "mask": TS([self.config.max_length])
                },
                {
                    "labels": TS(target_shape),
                    "associations": TS([self.config.max_length, self.config.max_length])
                }
            )
        )

    def _target_encoder(self):
        return SequenceLabelingEncoder()


[docs]class Association(BaseModel): """ Labels each token in a sequence as belonging to 1 of N token classes and then builds a set of edges between the labeled edges. :param config: A :py:class:`finetune.config.Settings` object or None (for default config). :param \**kwargs: key-value pairs of config items to override. """ def __init__(self, **kwargs): """ For a full list of configuration options, see `finetune.config`. :param n_epochs: defaults to `5`. :param lr_warmup: defaults to `0.1`, :param low_memory_mode: defaults to `True`, :param chunk_long_sequences: defaults to `True` :param **kwargs: key-value pairs of config items to override. """ self.multi_label = False super().__init__(**kwargs) def _get_input_pipeline(self): return AssociationPipeline(config=self.config, multi_label=False) def _initialize(self): if self.config.multi_label_sequences: raise FinetuneError("Multi label association not supported") return super()._initialize()
[docs] def finetune(self, Xs, Y=None, batch_size=None): """ :param Xs: A list of strings. :param Y: A list of labels of the same format as sequence labeling but with an option al additional field of the form: ``` { ... "association":{ "index": a, "relationship": relationship_name } ... ``` where index is the index of the relationship target into the label list and relationship_name is the type of the relationship. """ if self.config.association_types is None: raise FinetuneError("Please set config.association_types before calling finetune.") Xs, Y_new, association_type, association_idx, idxs = indico_to_finetune_sequence( Xs, encoder=self.input_pipeline.text_encoder, labels=Y, multi_label=False, none_value=self.config.pad_token ) Y = list(zip(Y_new, association_type, association_idx, idxs)) if Y is not None else None return super().finetune(Xs, Y=Y, batch_size=batch_size)
def prune_probs(self, prob_matrix, labels): viable_edges = self.config.viable_edges association_types = list(self.input_pipeline.association_encoder.classes_) if viable_edges is None: return prob_matrix for i, l1 in enumerate(labels): if l1 not in viable_edges: prob_matrix[i, :, :] = 0.0 continue elif None not in viable_edges[l1]: prob_matrix[i, :, self.input_pipeline.association_pad_idx] = 0.0 for cls in association_types: for j, l2 in enumerate(labels): if l1 not in viable_edges or l2 not in [c_t[0] for c_t in viable_edges[l1] if c_t and c_t[1] == cls]: prob_matrix[i, j, association_types.index(cls)] = 0.0 # this edge doesnt fit the schema return prob_matrix
[docs] def predict(self, X): """ Produces a list of most likely class labels as determined by the fine-tuned model. :param X: A list / array of text, shape [batch] :returns: list of class labels. """ pad_token = [self.config.pad_token] if self.multi_label else self.config.pad_token if self.config.viable_edges is None: LOGGER.warning("config.viable_edges is not set, this is probably incorrect.") #TODO(Ben) combine this into the sequence labeling model?? chunk_size = self.config.max_length - 2 step_size = chunk_size // 3 arr_encoded = list(itertools.chain.from_iterable( self.input_pipeline._text_to_ids([x], pad_token=(pad_token, pad_token, -1, -2)) for x in X )) lens = [len(a.char_locs) for a in arr_encoded] labels, batch_probas, associations = [], [], [] predict_keys = [ PredictMode.SEQUENCE, PredictMode.SEQUENCE_PROBAS, PredictMode.ASSOCIATION, PredictMode.ASSOCIATION_PROBAS ] for l, pred in zip(lens, self._inference(X, predict_keys=predict_keys)): pred_labels = self.input_pipeline.label_encoder.inverse_transform(pred[PredictMode.SEQUENCE]) pred_labels = [label if i < l else self.config.pad_token for i, label in enumerate(pred_labels)] labels.append(pred_labels) batch_probas.append(pred[PredictMode.SEQUENCE_PROBAS]) pred["association_probs"] = self.prune_probs(pred[PredictMode.ASSOCIATION_PROBAS], pred_labels) most_likely_associations, most_likely_class_id = zip( *[np.unravel_index(np.argmax(a, axis=None), a.shape) for a in pred[PredictMode.ASSOCIATION_PROBAS]] ) associations.append(( most_likely_associations, self.input_pipeline.association_encoder.inverse_transform(most_likely_class_id), [ prob[idx, cls] for prob, idx, cls in zip( pred["association_probs"], most_likely_associations, most_likely_class_id ) ] )) all_subseqs = [] all_labels = [] all_probs = [] all_assocs = [] doc_idx = -1 for chunk_idx, (label_seq, proba_seq, association) in enumerate(zip(labels, batch_probas, associations)): association_idx, association_class, association_prob = association position_seq = arr_encoded[chunk_idx].char_locs start_of_doc = arr_encoded[chunk_idx].token_ids[0][0] == self.input_pipeline.text_encoder.start end_of_doc = ( chunk_idx + 1 >= len(arr_encoded) or arr_encoded[chunk_idx + 1].token_ids[0][0] == self.input_pipeline.text_encoder.start ) start, end = 0, None if start_of_doc: # if this is the first chunk in a document, start accumulating from scratch doc_subseqs = [] doc_labels = [] doc_probs = [] doc_assocs = [] doc_idx += 1 start_of_token = 0 if not end_of_doc: end = step_size * 2 else: if end_of_doc: # predict on the rest of sequence start = step_size else: # predict only on middle third start, end = step_size, step_size * 2 label_seq = label_seq[start:end] position_seq = position_seq[start:end] proba_seq = proba_seq[start:end] tok_idx_to_subseq = dict() for tok_idx, (label, position, proba) in enumerate(zip(label_seq, position_seq, proba_seq)): if position == -1: # indicates padding / special tokens continue # if there are no current subsequence # or the current subsequence has the wrong label if not doc_subseqs or label != doc_labels[-1]: # start new subsequence doc_subseqs.append(X[doc_idx][start_of_token:position]) doc_probs.append([proba]) doc_assocs.append( [ (tok_idx, association_idx[tok_idx], association_class[tok_idx], association_prob[tok_idx]) ] ) doc_labels.append(label) else: # continue appending to current subsequence doc_subseqs[-1] += X[doc_idx][start_of_token:position] doc_probs[-1].append(proba) doc_assocs[-1].append( (tok_idx, association_idx[tok_idx], association_class[tok_idx], association_prob[tok_idx]) ) tok_idx_to_subseq[tok_idx] = len(doc_labels) - 1 start_of_token = position if end_of_doc: # last chunk in a document prob_dicts = [] for prob_seq in doc_probs: # format probabilities as dictionary probs = np.mean(np.vstack(prob_seq), axis=0) prob_dicts.append(dict(zip(self.input_pipeline.label_encoder.classes_, probs))) if self.multi_label: del prob_dicts[-1][self.config.pad_token] doc_assocs_by_idx = [] for assoc in doc_assocs: doc_assocs_by_idx.append([]) for from_idx, to_idx, cls, prob in assoc: if from_idx in tok_idx_to_subseq and to_idx in tok_idx_to_subseq: doc_assocs_by_idx[-1].append((tok_idx_to_subseq[from_idx], tok_idx_to_subseq[to_idx], cls, prob)) all_subseqs.append(doc_subseqs) all_labels.append(doc_labels) all_probs.append(prob_dicts) all_assocs.append(doc_assocs_by_idx) _, doc_annotations = finetune_to_indico_sequence(raw_texts=X, subseqs=all_subseqs, labels=all_labels, probs=all_probs, none_value=self.config.pad_token, subtoken_predictions=self.config.subtoken_predictions, associations=all_assocs) return doc_annotations
[docs] def featurize(self, X): """ Embeds inputs in learned feature space. Can be called before or after calling :meth:`finetune`. :param Xs: An iterable of lists or array of text, shape [batch, n_inputs, tokens] :returns: np.array of features of shape (n_examples, embedding_size). """ return self._featurize(X)
[docs] def predict_proba(self, X): """ Produces a list of most likely class labels as determined by the fine-tuned model. :param X: A list / array of text, shape [batch] :returns: list of class labels. """ return self.predict(X)
@staticmethod def _target_model(config, featurizer_state, targets, n_outputs, train=False, reuse=None, **kwargs): return association( hidden=featurizer_state['sequence_features'], pool_idx=featurizer_state['pool_idx'], targets=targets, n_targets=n_outputs, config=config, train=train, reuse=reuse, **kwargs ) def _predict_op(self, logits, **kwargs): associations = logits["association"] logits = logits["sequence"] trans_mats = kwargs.get("transition_matrix") if self.multi_label: logits = tf.unstack(logits, axis=-1) label_idxs = [] label_probas = [] for logits_i, trans_mat_i in zip(logits, trans_mats): idx, prob = sequence_decode(logits_i, trans_mat_i) label_idxs.append(idx) label_probas.append(prob[:, :, 1:]) label_idxs = tf.stack(label_idxs, axis=-1) label_probas = tf.stack(label_probas, axis=-1) else: label_idxs, label_probas = sequence_decode(logits, trans_mats) association_prob = tf.nn.softmax(associations, axis=-1) association_pred = tf.argmax(associations, axis=-1) return ( { PredictMode.SEQUENCE: label_idxs, PredictMode.ASSOCIATION: association_pred }, { PredictMode.SEQUENCE_PROBAS: label_probas, PredictMode.ASSOCIATION_PROBAS: association_prob } ) def _predict_proba_op(self, logits, **kwargs): return tf.no_op()