pyabsa.tasks.AspectPolarityClassification.instructor.ensembler

Classes

APCEnsembler

Thin wrapper to build datasets/tokenizers and ensemble APC models.

Functions

model_pool_check(models)

Ensure all models in the ensemble belong to the same APC family.

Module Contents

pyabsa.tasks.AspectPolarityClassification.instructor.ensembler.model_pool_check(models)

Ensure all models in the ensemble belong to the same APC family.

Mixing LCF-, BERT-baseline-, and GloVe-based models is not supported because their input conventions and output heads differ.

class pyabsa.tasks.AspectPolarityClassification.instructor.ensembler.APCEnsembler(config, load_dataset=True, **kwargs)

Bases: torch.nn.Module

Thin wrapper to build datasets/tokenizers and ensemble APC models.

• Merges inputs requirements across models into config.inputs_cols

• Lazily builds tokenizer/encoder and datasets as needed

• Supports concatenation or mean-averaging of logits for ensembling

config

inputs_cols

models

tokenizer = None

bert = None

embedding_matrix = None

train_set = None

test_set = None

valid_set = None

test_dataloader = None

valid_dataloader = None

dense

forward(inputs)