add taskmodules and models from pie-modules (#498)

This PR implements #459, i.e., it adds the models and taskmodules
implemented originally in
[pie-modules](https://github.com/ArneBinder/pie-modules) (except for QA
and span-pair based RE, see potential follow-ups below).

- added models:
  - `SequenceClassificationModelWithPooler`
  - `SequencePairSimilarityModelWithPooler`
  - `SimpleTokenClassificationModel`
  - `SimpleGenerativeModel`
  - `SimpleSequenceClassificationModel`
  - `TokenClassificationModelWithSeq2SeqEncoderAndCrf`
- added taskmodules:
  - `RETextClassificationWithIndicesTaskModule`
  - `TextToTextTaskModule`
  - `LabeledSpanExtractionByTokenClassificationTaskModule`
  - `PointerNetworkTaskModuleForEnd2EndRE`
  - `CrossTextBinaryCorefTaskModule`
 
**IMPORTANT: This restricts the version of transformers to
`>=4.35.0,<4.37.0`! So, this is breaking.**

requires: 
- #482
- #499

Additional changes:
 - add `tabulate`, and `pytorch-crf` to dev dependencies
- set dependence `torchmetrics[text] >=1.5, <2` to solve conflicts with
`nltk` (`text` loads the required additional dependencies and `>=1.5`
ensures that no deprecated nltk models are loaded. Note that we already
use the modern nltk models in
[`pie_documents.document.processing.NltkSentenceSplitter`](https://github.com/ArneBinder/pie-documents/blob/main/src/pie_documents/document/processing/sentence_splitter.py))
- add `SpanNotAlignedWithTokenException` and `get_aligned_token_span` to
`utils.document`
- add `RequiresMaxInputLength` and `RequiresTaskmoduleConfig` to
`models.interface`

potential follow-ups:
- [ ] add remaining models (SimpleExtractiveQuestionAnsweringModel and
SpanTupleClassificationModel)
- [ ] add remaining taskmodules (ExtractiveQuestionAnsweringTaskModule,
and RESpanPairClassificationTaskModule)

---------

Co-authored-by: Danylo Mysak <[email protected]>

Author: ArneBinder

codecov.yml

@@ -2,8 +2,8 @@ coverage:
     status:
         project:
             default:
-                target: 68% # TODO: switch back to auto
+                target: 80% # TODO: switch back to auto
                 threshold: 1% # the leniency in hitting the target
         patch:
             default:
-                target: 75% # TODO: switch back to 100%
+                target: 85% # TODO: switch back to 100%

poetry.lock

@@ -21,8 +21,11 @@ dependencies = [
     "pie-documents >=0.1.0, <0.2.0",
     "torch >=1.10",
     "pytorch-lightning >=2, <3",
-    "torchmetrics >1, <2",
-    "transformers >=4.18, <5",
+    "torchmetrics[text] >=1.5, <2",
+    # >=4.35 because of BartModelWithDecoderPositionIds,
+    # <4.37 because of generation config created from model config in BartAsPointerNetwork
+    # TODO: check the upper bound, since this should be already fixed (https://github.com/ArneBinder/pie-modules/pull/205)
+    "transformers >=4.35.0,<4.37.0",
 ]
 
 [project.urls]
@@ -51,14 +54,22 @@ classifiers = [
 optional = true
 
 [tool.poetry.group.dev.dependencies]
-# testing (with coverage and run-time type checking)
+
+# lazily imported packges
+# for taskmodule tests with collect statistics
+tabulate = "^0.9"
+# for TokenClassificationModelWithSeq2SeqEncoderAndCrf
+pytorch-crf = ">=0.7.2"
+
+# testing utilities (with coverage and run-time type checking)
 pytest = ">=6.2.5"
 pytest-xdist = "^3.8.0"
 pytest-cov = "^6.2.1"
 typeguard = ">=2.13.3"
 sh = "^2"
 types-requests = "^2.27.7"
 python-dotenv = "^0.20.0"
+
 # code quality and static type checking (via pre-commit)
 pre-commit = ">=2.16.0"
 pre-commit-hooks = ">=4.1.0"

pyproject.toml

@@ -1,3 +1,13 @@
+from pytorch_ie.models.sequence_classification_with_pooler import (
+    SequenceClassificationModelWithPooler,
+    SequencePairSimilarityModelWithPooler,
+)
+from pytorch_ie.models.simple_generative import SimpleGenerativeModel
+from pytorch_ie.models.simple_sequence_classification import SimpleSequenceClassificationModel
+from pytorch_ie.models.simple_token_classification import SimpleTokenClassificationModel
+from pytorch_ie.models.token_classification_with_seq2seq_encoder_and_crf import (
+    TokenClassificationModelWithSeq2SeqEncoderAndCrf,
+)
 from pytorch_ie.models.transformer_seq2seq import TransformerSeq2SeqModel
 from pytorch_ie.models.transformer_span_classification import TransformerSpanClassificationModel
 from pytorch_ie.models.transformer_text_classification import TransformerTextClassificationModel
@@ -8,4 +18,10 @@
     "TransformerSpanClassificationModel",
     "TransformerTextClassificationModel",
     "TransformerTokenClassificationModel",
+    "SequenceClassificationModelWithPooler",
+    "SequencePairSimilarityModelWithPooler",
+    "SimpleTokenClassificationModel",
+    "SimpleGenerativeModel",
+    "SimpleSequenceClassificationModel",
+    "TokenClassificationModelWithSeq2SeqEncoderAndCrf",
 ]

src/pytorch_ie/models/__init__.py

@@ -0,0 +1,2 @@
+from .bart_as_pointer_network import BartAsPointerNetwork
+from .bart_with_decoder_position_ids import BartModelWithDecoderPositionIds