ArneBinder

ZeroShotClassificationExplainer does not correctly explain ZeroShotClassificationPipeline results (single label)

2

In the case of a single label, the logic to calculate the classification probability with the `ZeroShotClassificationExplainer` (see [here](https://github.com/cdpierse/transformers-interpret/blob/v0.6.0/transformers_interpret/explainers/sequence_classification.py#L193-L194)) is different than the logic in the Huggingface `ZeroShotClassificationPipeline` (see [here](https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/pipelines/zero_shot_classification.py#L249-L251)):...

Semantic Web pointer

1

With respect to ideas regarding SPARQL and RDF mentioned in the readme: DBpedia provides some interesting resources, especially they published a dbpedia version in **[NIF-RDF](https://persistence.uni-leipzig.org/nlp2rdf/)**: https://wiki.dbpedia.org/downloads-2016-10 and see [here](https://github.com/NLP2RDF/DBpediaOpenDBpediaTextExtractionChallenge/blob/master/BO.ttl) for...

[Rendered](https://github.com/ChristophAlt/pytorch-ie/tree/docs_pie_concepts#-concepts--architecture)

`encode_inputs` should not do anything depending on the state of `is_training`, respective code can live in `encode_targets`. This will ease separation of concerns and testing. To implement this it may...

[WIP] fix training seq2seq

2

This was broken because pytorch-lightning tries to move the output of `TransformerSeq2SeqTaskModule.collate` to a device via `pytorch_lightning.core.datamodule.LightningDataModule.transfer_batch_to_device` that internally uses [`pytorch_lightning.utilities.apply_func.apply_to_collection`](https://pytorch-lightning.readthedocs.io/en/stable/api/pytorch_lightning.utilities.apply_func.html#pytorch_lightning.utilities.apply_func.apply_to_collection). This method fails if any part of the input...

Until #183 is implemented, we need at least descriptions for the most relevant parts of PyTorch-IE in the readme to make it usable by the public. This may require the...

[WIP] is_prepared attribute for taskmodule

1

Approach: The base taskmodule now has also a `_prepare()` method which should be overwritten in derived classes instead of `prepare()`. `prepare()` does now the following: 1. it checks, if the...

[WIP] use dataset.map in pipeline

2

If `documents` of type `Dataset` is passed to the pipeline, use `documents.map` to add the predictions. In this case, a `Dataset` is returned instead of `Sequence[Document]`. Note: Builds on top...

[WIP] Integrate predictions

4

This implements `AnnotationList.integrate_predictions(overwrite: bool = False)` which allows to move predictions to the main annotations.