Fix LightningCLI compatibility (#288)

Change AutoModel type hint to fix LightningCLI compatibility.

lightning_transformers/core/model.py

@@ -12,24 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from pathlib import Path
-from typing import IO, TYPE_CHECKING, Any, Callable, Dict, Optional, Tuple, Type, Union
+from typing import IO, Any, Callable, Dict, Optional, Tuple, Type, Union
 
 import pytorch_lightning as pl
 import torch
 import transformers
 from pytorch_lightning.utilities import rank_zero_warn
-from transformers import AutoConfig, PreTrainedTokenizerBase
+from transformers import AutoConfig, Pipeline, PreTrainedTokenizerBase
 from transformers import pipeline as hf_transformers_pipeline
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.utilities.deepspeed import enable_transformers_pretrained_deepspeed_sharding
 from lightning_transformers.utilities.imports import _ACCELERATE_AVAILABLE
 
 if _ACCELERATE_AVAILABLE:
     from accelerate import load_checkpoint_and_dispatch
 
-if TYPE_CHECKING:
-    from transformers import AutoModel, Pipeline
-
 
 class TaskTransformer(pl.LightningModule):
     """Base class for task specific transformers, wrapping pre-trained language models for downstream tasks. The
@@ -47,7 +45,7 @@ class TaskTransformer(pl.LightningModule):
 
     def __init__(
         self,
-        downstream_model_type: Type["AutoModel"],
+        downstream_model_type: Type[_BaseAutoModelClass],
         pretrained_model_name_or_path: Optional[str] = None,
         tokenizer: Optional[PreTrainedTokenizerBase] = None,
         pipeline_kwargs: Optional[dict] = None,
@@ -127,7 +125,7 @@ def configure_metrics(self, stage: str) -> Optional[Any]:
         pass
 
     @property
-    def tokenizer(self) -> Optional["PreTrainedTokenizerBase"]:
+    def tokenizer(self) -> Optional[PreTrainedTokenizerBase]:
         if (
             self._tokenizer is None
             and hasattr(self, "trainer")  # noqa: W503
@@ -138,7 +136,7 @@ def tokenizer(self) -> Optional["PreTrainedTokenizerBase"]:
         return self._tokenizer
 
     @tokenizer.setter
-    def tokenizer(self, tokenizer: "PreTrainedTokenizerBase") -> None:
+    def tokenizer(self, tokenizer: PreTrainedTokenizerBase) -> None:
         self._tokenizer = tokenizer
 
     @property
@@ -150,7 +148,7 @@ def hf_pipeline_task(self) -> Optional[str]:
         return None
 
     @property
-    def hf_pipeline(self) -> "Pipeline":
+    def hf_pipeline(self) -> Pipeline:
         if self._hf_pipeline is None:
             if self.hf_pipeline_task is not None:
                 self._hf_pipeline = hf_transformers_pipeline(
@@ -161,7 +159,7 @@ def hf_pipeline(self) -> "Pipeline":
         return self._hf_pipeline
 
     @hf_pipeline.setter
-    def hf_pipeline(self, pipeline: "Pipeline") -> None:
+    def hf_pipeline(self, pipeline: Pipeline) -> None:
         self._hf_pipeline = pipeline
 
     def hf_predict(self, *args, **kwargs) -> Any:

lightning_transformers/task/nlp/language_modeling/model.py

@@ -11,17 +11,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Any, Type
+from typing import Any, Type
 
 import torch
 import transformers
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class LanguageModelingTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Language Modeling Task.
@@ -33,7 +31,7 @@ class LanguageModelingTransformer(TaskTransformer):
     """
 
     def __init__(
-        self, *args, downstream_model_type: Type["AutoModel"] = transformers.AutoModelForCausalLM, **kwargs
+        self, *args, downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForCausalLM, **kwargs
     ) -> None:
         super().__init__(downstream_model_type, *args, **kwargs)
 

lightning_transformers/task/nlp/masked_language_modeling/model.py

@@ -11,15 +11,13 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Type
+from typing import Type
 
 import transformers
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class MaskedLanguageModelingTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Masked Language Modeling Task.
@@ -31,7 +29,7 @@ class MaskedLanguageModelingTransformer(TaskTransformer):
     """
 
     def __init__(
-        self, *args, downstream_model_type: Type["AutoModel"] = transformers.AutoModelForMaskedLM, **kwargs
+        self, *args, downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForMaskedLM, **kwargs
     ) -> None:
         super().__init__(downstream_model_type, *args, **kwargs)
 

lightning_transformers/task/nlp/multiple_choice/model.py

@@ -11,17 +11,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Type
+from typing import Type
 
 import torch
 import transformers
 from torchmetrics import Accuracy, Precision, Recall
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class MultipleChoiceTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Multiple Choice Task.
@@ -34,7 +32,10 @@ class MultipleChoiceTransformer(TaskTransformer):
     """
 
     def __init__(
-        self, *args, downstream_model_type: Type["AutoModel"] = transformers.AutoModelForMultipleChoice, **kwargs
+        self,
+        *args,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForMultipleChoice,
+        **kwargs,
     ) -> None:
         super().__init__(downstream_model_type, *args, **kwargs)
 

lightning_transformers/task/nlp/question_answering/model.py

@@ -12,18 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from functools import partial
-from typing import TYPE_CHECKING, Any, Type
+from typing import Any, Type
 
 import torch
 import transformers
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 from lightning_transformers.task.nlp.question_answering import QuestionAnsweringDataModule
 from lightning_transformers.task.nlp.question_answering.datasets.squad.metric import SquadMetric
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class QuestionAnsweringTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Question Answering Task.
@@ -36,7 +34,10 @@ class QuestionAnsweringTransformer(TaskTransformer):
     """
 
     def __init__(
-        self, *args, downstream_model_type: Type["AutoModel"] = transformers.AutoModelForQuestionAnswering, **kwargs
+        self,
+        *args,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForQuestionAnswering,
+        **kwargs,
     ) -> None:
         super().__init__(downstream_model_type, *args, **kwargs)
 

lightning_transformers/task/nlp/summarization/model.py

@@ -11,16 +11,14 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Type
+from typing import Type
 
 import transformers
 from torchmetrics.text.rouge import ROUGEScore
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core.seq2seq.model import Seq2SeqTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class SummarizationTransformer(Seq2SeqTransformer):
     """Defines ``LightningModule`` for the Summarization Task.
@@ -36,7 +34,7 @@ class SummarizationTransformer(Seq2SeqTransformer):
     def __init__(
         self,
         *args,
-        downstream_model_type: Type["AutoModel"] = transformers.AutoModelForSeq2SeqLM,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForSeq2SeqLM,
         use_stemmer: bool = True,
         **kwargs
     ) -> None:

lightning_transformers/task/nlp/text_classification/model.py

@@ -11,17 +11,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Any, Dict, Type
+from typing import Any, Dict, Type
 
 import torch
 import transformers
 from torchmetrics import Accuracy, Precision, Recall
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class TextClassificationTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Text Classification Task.
@@ -36,7 +34,7 @@ class TextClassificationTransformer(TaskTransformer):
     def __init__(
         self,
         *args,
-        downstream_model_type: Type["AutoModel"] = transformers.AutoModelForSequenceClassification,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForSequenceClassification,
         **kwargs,
     ) -> None:
         super().__init__(downstream_model_type, *args, **kwargs)

lightning_transformers/task/nlp/token_classification/model.py

@@ -11,17 +11,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Any, Dict, List, Type, Union
+from typing import Any, Dict, List, Type, Union
 
 import torch
 import transformers
 from torchmetrics import Accuracy, F1Score, Precision, Recall
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class TokenClassificationTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Text Classification Task.
@@ -37,7 +35,7 @@ def __init__(
         self,
         *args,
         labels: Union[int, List[str]],
-        downstream_model_type: Type["AutoModel"] = transformers.AutoModelForTokenClassification,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForTokenClassification,
         **kwargs,
     ) -> None:
         num_labels = labels if isinstance(labels, int) else len(labels)

lightning_transformers/task/nlp/translation/model.py

@@ -11,18 +11,16 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Type
+from typing import Type
 
 import transformers
 from torchmetrics.text.bleu import BLEUScore
 from transformers import MBartTokenizer
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core.seq2seq.model import Seq2SeqTransformer
 from lightning_transformers.task.nlp.translation import TranslationDataModule
 
-if TYPE_CHECKING:
-    from transformers import AutoModel
-
 
 class TranslationTransformer(Seq2SeqTransformer):
     """Defines ``LightningModule`` for the Translation Task.
@@ -39,7 +37,7 @@ class TranslationTransformer(Seq2SeqTransformer):
     def __init__(
         self,
         *args,
-        downstream_model_type: Type["AutoModel"] = transformers.AutoModelForSeq2SeqLM,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForSeq2SeqLM,
         n_gram: int = 4,
         smooth: bool = False,
         **kwargs,

lightning_transformers/task/vision/image_classification/model.py

@@ -11,17 +11,15 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-from typing import TYPE_CHECKING, Any, Dict, Type
+from typing import Any, Dict, Type
 
 import torch
 import transformers
 from torchmetrics import Accuracy, Precision, Recall
+from transformers.models.auto.auto_factory import _BaseAutoModelClass
 
 from lightning_transformers.core import TaskTransformer
 
-if TYPE_CHECKING:
-    from transformers import AutoModel, Pipeline
-
 
 class ImageClassificationTransformer(TaskTransformer):
     """Defines ``LightningModule`` for the Text Classification Task.
@@ -34,7 +32,10 @@ class ImageClassificationTransformer(TaskTransformer):
     """
 
     def __init__(
-        self, *args, downstream_model_type: Type["AutoModel"] = transformers.AutoModelForImageClassification, **kwargs
+        self,
+        *args,
+        downstream_model_type: Type[_BaseAutoModelClass] = transformers.AutoModelForImageClassification,
+        **kwargs,
     ) -> None:
         super().__init__(downstream_model_type, *args, **kwargs)
         self.metrics = {}
@@ -83,6 +84,6 @@ def hf_pipeline_task(self) -> str:
         return "image-classification"
 
     @property
-    def hf_pipeline(self) -> "Pipeline":
+    def hf_pipeline(self) -> transformers.Pipeline:
         self._hf_pipeline_kwargs["feature_extractor"] = self.tokenizer
         return super().hf_pipeline

tests/core/test_cli.py

@@ -0,0 +1,32 @@
+from importlib.util import find_spec
+from unittest import mock
+
+import pytest
+from pytorch_lightning.cli import LightningCLI
+
+from lightning_transformers.task.vision.image_classification import (
+    ImageClassificationDataModule,
+    ImageClassificationTransformer,
+)
+
+
+@pytest.mark.skipif(find_spec("jsonargparse") is None, reason="jsonargparse is required")
+def test_lightning_cli_image_classification():
+    config = {
+        "data": {
+            "dataset_name": "beans",  # Resolve from TransformerDataModule.__init__
+        },
+        "model": {
+            "pretrained_model_name_or_path": "nateraw/tiny-vit-random",  # Resolve from TaskTransformer.__init__
+        },
+    }
+    with mock.patch("sys.argv", ["any.py", f"--config={config}"]):
+        cli = LightningCLI(
+            ImageClassificationTransformer,
+            ImageClassificationDataModule,
+            run=False,
+        )
+    assert cli.config.data.dataset_name == "beans"
+    assert cli.config.model.pretrained_model_name_or_path == "nateraw/tiny-vit-random"
+    assert isinstance(cli.config_init.data, ImageClassificationDataModule)
+    assert isinstance(cli.config_init.model, ImageClassificationTransformer)

tests/task/nlp/test_token_classification.py

@@ -21,7 +21,7 @@ def test_smoke_train(hf_cache_path):
         dataset_name="conll2003",
         preprocessing_num_workers=1,
         label_all_tokens=False,
-        revision="master",
+        revision="main",
         limit_test_samples=64,
         limit_val_samples=64,
         limit_train_samples=64,