The exception raised by any Adaseq object when it's misconfigured (e.g. missing properties, invalid properties, unknown properties).

Word-level Dropout module During training, randomly zeroes some of the elements of the input tensor at word level with probability `dropout_r

Conditional random field. This module implements a conditional random field [LMP01]_. The forward computation of this class computes the log l

Turn token embedding sequenece to a single vector.

Filters out everything that is at the ERROR level or higher. This is meant to be used with a stdout handler when a stderr handler is also con

evaluate macro and micro set p/r/f1 scores

Weighed BCE loss, multiply the loss of positive examples with a scaler

Biaffine Attention

Biaffine scorer.

Conditional random field. This module implements a conditional random field [LMP01]_. The forward computation of this class computes the log l

represent a data batch, support `tensorize` function.

MLP on label embeddings to parameterize potentials

TODO

The module proposed in the EMNLP2022 Paper to model label correlations, require label embeddings and use mean-field variational inference.

Partial/Fuzzy Conditional random field.

TODO

Computes a parameterised scalar mixture of N tensors, `mixture = gamma * sum(s_k * tensor_k)` where `s = softmax(w)`, with `w` and `gamma` sc

evaluate macro and micro set p/r/f1 scores

2-layer feed forward

Sin-Cos Embedding. ref: https://spaces.ac.cn/archives/8265

A custom subclass of 'logging.Logger' that keeps a set of messages to implement {debug,info,etc.}_once() methods.

Logger hook in text, Output log to both console and local json file. Args: by_epoch (bool, optional): Whether EpochBasedtrainer is used.

Named Entity Recognition as Dependency Parsing (Yu et al., ACL 2020) ref: https://aclanthology.org/2020.acl-main.577/

Evaluate metrics for typing tasks, with threshold tuning

Base class for custumized dataset builder.

A `DataCollator` support padding some fields to same length. Support padding encoder related fields: input_ids, token_type_ids, mask, and

Names for different data_collators

class to dump model predictions

Names for different dataset dumpers

An `DatasetManager` is used by trainers to load a dataset from `modelscope`, local files by built-in `DatasetBuilder`s, or custom huggingface

The decoder base class for downstream tasks

Names for different decoders

Default trainer class for AdaSeq. This trainer inherits from EpochBasedTrainer with some modifications. It implements some common data proces

The embedder base class for encoding input_ids to hidden-states

Names for different embedders

A simple lookup table for word embedding. This module implements a simple lookup table that stores embeddings of a fixed dictionary and size.

The encoder base class for encoding embeddings to features.

Names for different encoders

Builder for entity typing datasets. features: id: string, data record id. tokens: list[str] input tokens. spans: List[Dic

BuilderConfig for entity typing datasets

The base class of exporter inheriting from TorchModelExporter. This class provides the default implementations for exporting onnx and torch scrip

GlobalPointer model. ref: https://arxiv.org/abs/2208.03054 ref: https://github.com/xhw205/Efficient-GlobalPointer-torch

Registered as a `Encoder` with name "gru".

Names for different hooks

Names for different lr_schedulers

Registered as a `Encoder` with name "lstm".

Builder for entity typing datasets. features: id: string, data record id. tokens: list[str] input tokens. spans: List[Dic

BuilderConfig for entity typing datasets

Names for different metrics

The model base class

Names for different models

Padding method for multilabel span concat typing dataset.

Preprocessor for multilabel (aka multi-type) span concat typing task. cand_size: number of candidates, -1 when use all labels candidates

Concat based Single Mention MultiLabel Entity Typing model This model is used for single mention multilabel entity typing tasks. Input format

Concat based Single Mention MultiLabel Entity Typing model MCCE This model is used for single mention multilabel entity typing tasks. Input f

Preprocessor for multilabel (aka multi-type) span concat typing task.

Padding method for multilabel span typing dataset.

Span based MultiLabel Entity Typing model This model is used for multilabel entity typing tasks. Args: num_labels (int): number of l

Preprocessor for multilabel (aka multi-type) span typing task. span targets are processed into mention_boundary, type_ids. examples: s

Some common pre-process operations for NLP tasks. Args: model_dir (str): pre-trained model name or path. is_word2vec (bool):

Builder for entity typing datasets. features: id: string, data record id. tokens: list[str] input tokens. spans: List[Dic

BuilderConfig for Named Entity Recognition datasets

Named Entity Recognition dumper.

TODO

Names for different optimizers

PU Learning Loss. https://aclanthology.org/P19-1231.pdf https://aclanthology.org/2020.findings-emnlp.60.pdf

PartialBCELoss

use `model` and `preprocessor` to create a pipeline for prediction Args: model (str or Model): A model instance or a model local dir or a

Names for different pipelines

Names for different preprocessors

Padding method for pretraining dataset.

Evaluate metrics for typing tasks and identification tasks

Pretraining model Args: num_labels (int): number of labels embedder (Union[embedder, str], `optional`): embedder used in the mode

Preprocessor for pretraining. span targets are processed into mention_boundary, type_ids, ident_ids. examples: span: {'start':1, 'end'

Dataset Builder for Relation Extraction

Builder Config for Relation Extraction

The metric computation class for relation extraction tasks.

Relation extraction model This model is used for relation extraction tasks.

Relation Extraction data preprocessor

Registered as a `Encoder` with name "rnn".

Collator for the sequence labeling task

The metric computation class for sequence-labeling tasks. This metric class uses seqeval to calculate scores. Args: return_macro_f1

Sequence labeling model This model is used for sequence labeling tasks. Various decoders are supported, including argmax, crf, partial-crf, e

An exporter for sequence labeling model

A pipeline for sequence labeling models Model outputs should contain: predicts: [[0, 0, 1, 2, 2, ...]] Pipeline outputs will be l

Preprocessor for Sequence Labeling

A NER pipeline for span based models, e.g. biaffine model and global pointer model. Model outputs should contain: predicts: [[(start,

Preprocessor of span-based model. span targets are processed into `span_labels`

Padding method for span extraction dataset.

The metric computation class for span-extraction tasks.

Abstract class for subcommands

Names for different tasks

usage: adaseq test [-h] -w WORK_DIR [-d DEVICE] [-ckpt CHECKPOINT_PATH] optional arguments: -h, --help show this help messa

usage: adaseq train [-h] -c CONFIG_PATH [-w WORK_DIR] [-n RUN_NAME] [-f FORCE] [-ckpt CHECKPOINT_PATH] [--seed SEED] [-d

Names for different trainers

Uses a pretrained model from `transformers` as a `Embedder`. # Parameters model_name_or_path : `str` The name of the `transform

Two stage ner model This model is used for two-stage few-shot ner model with pre-trained NER model (see pretraining_model module). Args:

Preprocessor for twostage-ner. span targets are processed into mention_boundary, type_ids, ident_ids. examples: span: {'start':1, 'end

Padding method for two stage ner dataset.