kaiko-eva 0.3.3__py3-none-any.whl → 0.4.1__py3-none-any.whl

eva/language/data/datasets/text.py

@@ -0,0 +1,92 @@
+"""Base classes for text-image datasets."""
+
+import abc
+from typing import Any, Dict, Generic
+
+from typing_extensions import override
+
+from eva.language.data.datasets.base import LanguageDataset
+from eva.language.data.datasets.schemas import TransformsSchema
+from eva.language.data.datasets.typings import TargetType, TextSample
+from eva.language.data.messages import MessageSeries
+
+
+class TextDataset(LanguageDataset[TextSample[TargetType]], abc.ABC, Generic[TargetType]):
+    """Base dataset class for text-based tasks."""
+
+    def __init__(self, *args, transforms: TransformsSchema | None = None, **kwargs) -> None:
+        """Initializes the dataset.
+
+        Args:
+            *args: Positional arguments for the base class.
+            transforms: The transforms to apply to the text and target when
+                loading the samples.
+            **kwargs: Keyword arguments for the base class.
+        """
+        super().__init__(*args, **kwargs)
+
+        self.transforms = transforms
+
+    def load_metadata(self, index: int) -> Dict[str, Any] | None:
+        """Returns the dataset metadata.
+
+        Args:
+            index: The index of the data sample.
+
+        Returns:
+            The sample metadata.
+        """
+
+    @abc.abstractmethod
+    def load_text(self, index: int) -> MessageSeries:
+        """Returns the text content.
+
+        Args:
+            index: The index of the data sample.
+
+        Returns:
+            The text content.
+        """
+        raise NotImplementedError
+
+    @abc.abstractmethod
+    def load_target(self, index: int) -> TargetType:
+        """Returns the target label.
+
+        Args:
+            index: The index of the data sample.
+
+        Returns:
+            The target label.
+        """
+        raise NotImplementedError
+
+    @override
+    def __getitem__(self, index: int) -> TextSample[TargetType]:
+        item = TextSample(
+            text=self.load_text(index),
+            target=self.load_target(index),
+            metadata=self.load_metadata(index) or {},
+        )
+        return self._apply_transforms(item)
+
+    def _apply_transforms(self, sample: TextSample[TargetType]) -> TextSample[TargetType]:
+        """Applies the dataset transforms to the text and target.
+
+        Args:
+            sample: The text sample..
+
+        Returns:
+            The transformed sample.
+        """
+        if self.transforms:
+            text = self.transforms.text(sample.text) if self.transforms.text else sample.text
+            target = (
+                self.transforms.target(sample.target) if self.transforms.target else sample.target
+            )
+            return TextSample(
+                text=text,
+                target=target,
+                metadata=sample.metadata,
+            )
+        return sample

eva/language/data/datasets/typings.py

@@ -0,0 +1,39 @@
+"""Typings for multimodal datasets."""
+
+from typing import Any, Generic, TypeVar
+
+from typing_extensions import NamedTuple
+
+from eva.language.data.messages import MessageSeries
+
+TargetType = TypeVar("TargetType")
+"""The target data type."""
+
+
+class TextSample(NamedTuple, Generic[TargetType]):
+    """Text sample with target and metadata."""
+
+    text: MessageSeries
+    """One or multiple conversation messages."""
+
+    target: TargetType | None
+    """Target data."""
+
+    metadata: dict[str, Any] | None
+    """Additional metadata."""
+
+
+class PredictionSample(NamedTuple, Generic[TargetType]):
+    """Text sample with target and metadata."""
+
+    prediction: TargetType
+    """Prediction data."""
+
+    target: TargetType
+    """Target data."""
+
+    text: MessageSeries | None
+    """Conversation messages that were used as input."""
+
+    metadata: dict[str, Any] | None
+    """Additional metadata."""

eva/language/data/messages.py

@@ -0,0 +1,60 @@
+"""Types and classes for conversation messages in a multimodal context."""
+
+import dataclasses
+import enum
+from typing import Any, Dict, List
+
+
+class Role(str, enum.Enum):
+    """Roles for messages in a conversation."""
+
+    USER = "user"
+    ASSISTANT = "assistant"
+    SYSTEM = "system"
+
+
+@dataclasses.dataclass
+class Message:
+    """Base class for a message in a conversation."""
+
+    content: str
+    role: str
+
+    def to_dict(self) -> Dict[str, Any]:
+        """Convert the message to a dictionary."""
+        return dataclasses.asdict(self)
+
+
+@dataclasses.dataclass
+class UserMessage(Message):
+    """User message in a conversation."""
+
+    role: str = Role.USER
+
+
+@dataclasses.dataclass
+class AssistantMessage(Message):
+    """Assistant message in a conversation."""
+
+    role: str = Role.ASSISTANT
+
+
+@dataclasses.dataclass
+class SystemMessage(Message):
+    """System message in a conversation."""
+
+    role: str = Role.SYSTEM
+
+
+@dataclasses.dataclass
+class ModelSystemMessage(SystemMessage):
+    """System message for model-specific instructions."""
+
+
+@dataclasses.dataclass
+class TaskSystemMessage(SystemMessage):
+    """System message for task-specific instructions."""
+
+
+MessageSeries = List[Message]
+"""A series of conversation messages, can contain a mix of system, user, and AI messages."""

eva/language/models/__init__.py

@@ -1,25 +1,29 @@
 """Language Models API."""
 
-from eva.language.models import modules, wrappers
-from eva.language.models.modules import TextModule
-from eva.language.models.wrappers import HuggingFaceTextModel, LiteLLMTextModel
+from eva.language.models import modules, networks, wrappers
+from eva.language.models.modules import LanguageModule, OfflineLanguageModule
+from eva.language.models.wrappers import HuggingFaceModel, LiteLLMModel
 
 try:
-    from eva.language.models.wrappers import VLLMTextModel
+    from eva.language.models.wrappers import VllmModel
 
     __all__ = [
         "modules",
         "wrappers",
-        "TextModule",
-        "HuggingFaceTextModel",
-        "LiteLLMTextModel",
-        "VLLMTextModel",
+        "networks",
+        "HuggingFaceModel",
+        "LiteLLMModel",
+        "VllmModel",
+        "LanguageModule",
+        "OfflineLanguageModule",
     ]
 except ImportError:
     __all__ = [
         "modules",
         "wrappers",
-        "TextModule",
-        "HuggingFaceTextModel",
-        "LiteLLMTextModel",
+        "networks",
+        "HuggingFaceModel",
+        "LiteLLMModel",
+        "LanguageModule",
+        "OfflineLanguageModule",
     ]

eva/language/models/modules/__init__.py

@@ -1,5 +1,5 @@
 """Language Networks API."""
 
-from eva.language.models.modules.text import TextModule
+from eva.language.models.modules.language import LanguageModule, OfflineLanguageModule
 
-__all__ = ["TextModule"]
+__all__ = ["LanguageModule", "OfflineLanguageModule"]

eva/language/models/modules/language.py

@@ -0,0 +1,94 @@
+"""Model module for language models."""
+
+from typing import Any
+
+from lightning.pytorch.utilities.types import STEP_OUTPUT
+from torch import nn
+from typing_extensions import override
+
+from eva.core.metrics import structs as metrics_lib
+from eva.core.models.modules import module
+from eva.core.models.modules.utils import batch_postprocess
+from eva.language.models.typings import ModelOutput, PredictionBatch, TextBatch
+
+
+class LanguageModule(module.ModelModule):
+    """Model module for language tasks."""
+
+    def __init__(
+        self,
+        model: nn.Module,
+        metrics: metrics_lib.MetricsSchema | None = None,
+        postprocess: batch_postprocess.BatchPostProcess | None = None,
+    ) -> None:
+        """Initializes the text inference module.
+
+        Args:
+            model: Model instance to use for forward pass.
+            metrics: Metrics schema for evaluation.
+            postprocess: A helper function to post-process model outputs before evaluation.
+        """
+        super().__init__(metrics=metrics, postprocess=postprocess)
+
+        self.model = model
+
+    @override
+    def forward(self, batch: TextBatch, *args: Any, **kwargs: Any) -> ModelOutput:
+        return self.model(batch)
+
+    @override
+    def validation_step(self, batch: TextBatch, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
+        return self._batch_step(batch)
+
+    @override
+    def test_step(self, batch: TextBatch, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
+        return self._batch_step(batch)
+
+    def _batch_step(self, batch: TextBatch) -> STEP_OUTPUT:
+        text, targets, metadata = TextBatch(*batch)
+        output = self.forward(batch)
+
+        return {
+            "inputs": text,
+            "predictions": output.pop("generated_text"),  # type: ignore
+            "targets": targets,
+            "metadata": metadata,
+        } | output
+
+
+class OfflineLanguageModule(module.ModelModule):
+    """Model module for offline language tasks."""
+
+    def __init__(
+        self,
+        metrics: metrics_lib.MetricsSchema | None = None,
+        postprocess: batch_postprocess.BatchPostProcess | None = None,
+    ) -> None:
+        """Initializes the text inference module.
+
+        Args:
+            metrics: Metrics schema for evaluation.
+            postprocess: A helper function to post-process model outputs before evaluation.
+        """
+        super().__init__(metrics=metrics, postprocess=postprocess)
+
+    @override
+    def forward(self, batch: PredictionBatch, *args: Any, **kwargs: Any) -> PredictionBatch:
+        return batch
+
+    @override
+    def validation_step(self, batch: PredictionBatch, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
+        return self._batch_step(batch)
+
+    @override
+    def test_step(self, batch: PredictionBatch, *args: Any, **kwargs: Any) -> STEP_OUTPUT:
+        return self._batch_step(batch)
+
+    def _batch_step(self, batch: PredictionBatch) -> STEP_OUTPUT:
+        predictions, targets, text, metadata = PredictionBatch(*batch)
+        return {
+            "inputs": text,
+            "predictions": predictions,
+            "targets": targets,
+            "metadata": metadata,
+        }

eva/language/models/networks/__init__.py

@@ -0,0 +1,12 @@
+"""Language networks API."""
+
+from eva.language.models.networks.alibaba import Qwen205BInstruct
+from eva.language.models.networks.api import Claude35Sonnet20240620, Claude37Sonnet20250219
+from eva.language.models.networks.registry import model_registry
+
+__all__ = [
+    "Claude35Sonnet20240620",
+    "Claude37Sonnet20250219",
+    "Qwen205BInstruct",
+    "model_registry",
+]

eva/language/models/networks/alibaba.py

@@ -0,0 +1,26 @@
+"""Models from Alibaba."""
+
+import torch
+
+from eva.language.models import wrappers
+from eva.language.models.networks.registry import model_registry
+
+
+@model_registry.register("alibaba/qwen2-0-5b-instruct")
+class Qwen205BInstruct(wrappers.HuggingFaceModel):
+    """Qwen2 0.5B Instruct model."""
+
+    def __init__(self, system_prompt: str | None = None, cache_dir: str | None = None):
+        """Initialize the model."""
+        super().__init__(
+            model_name_or_path="Qwen/Qwen2-0.5B-Instruct",
+            model_kwargs={
+                "torch_dtype": torch.bfloat16,
+                "cache_dir": cache_dir,
+            },
+            generation_kwargs={
+                "max_new_tokens": 512,
+            },
+            system_prompt=system_prompt,
+            chat_mode=True,
+        )

eva/language/models/networks/api/__init__.py

@@ -0,0 +1,11 @@
+"""Multimodal API networks."""
+
+from eva.language.models.networks.api.anthropic import (
+    Claude35Sonnet20240620,
+    Claude37Sonnet20250219,
+)
+
+__all__ = [
+    "Claude35Sonnet20240620",
+    "Claude37Sonnet20250219",
+]

eva/language/models/networks/api/anthropic.py

@@ -0,0 +1,34 @@
+"""Models from Anthropic."""
+
+import os
+
+from eva.language.models import wrappers
+from eva.language.models.networks.registry import model_registry
+
+
+class _Claude(wrappers.LiteLLMModel):
+    """Base class for Claude models."""
+
+    def __init__(self, model_name: str, system_prompt: str | None = None):
+        if not os.getenv("ANTHROPIC_API_KEY"):
+            raise ValueError("ANTHROPIC_API_KEY env variable must be set.")
+
+        super().__init__(model_name=model_name, system_prompt=system_prompt)
+
+
+@model_registry.register("anthropic/claude-3-5-sonnet-20240620")
+class Claude35Sonnet20240620(_Claude):
+    """Claude 3.5 Sonnet (June 2024) model."""
+
+    def __init__(self, system_prompt: str | None = None):
+        """Initialize the model."""
+        super().__init__(model_name="claude-3-5-sonnet-20240620", system_prompt=system_prompt)
+
+
+@model_registry.register("anthropic/claude-3-7-sonnet-20250219")
+class Claude37Sonnet20250219(_Claude):
+    """Claude 3.7 Sonnet (February 2025) model."""
+
+    def __init__(self, system_prompt: str | None = None):
+        """Initialize the model."""
+        super().__init__(model_name="claude-3-7-sonnet-20250219", system_prompt=system_prompt)

eva/language/models/networks/registry.py

@@ -0,0 +1,5 @@
+"""Language Model Registry."""
+
+from eva.core.utils.registry import Registry
+
+model_registry = Registry()

eva/language/models/typings.py

@@ -0,0 +1,56 @@
+"""Type definitions for language models."""
+
+from typing import Any, Dict, Generic, List, TypedDict, TypeVar
+
+import torch
+from typing_extensions import NamedTuple, NotRequired
+
+from eva.language.data.messages import MessageSeries
+
+TargetType = TypeVar("TargetType")
+"""The target data type."""
+
+
+class TextBatch(NamedTuple, Generic[TargetType]):
+    """Text sample with target and metadata."""
+
+    text: List[MessageSeries]
+    """Text content."""
+
+    target: TargetType | None
+    """Target data."""
+
+    metadata: Dict[str, Any] | None
+    """Additional metadata."""
+
+
+class PredictionBatch(NamedTuple, Generic[TargetType]):
+    """Text sample with target and metadata."""
+
+    prediction: TargetType
+    """Prediction data."""
+
+    target: TargetType
+    """Target data."""
+
+    text: List[MessageSeries] | None
+    """Conversation messages that were used as input."""
+
+    metadata: Dict[str, Any] | None
+    """Additional metadata."""
+
+
+class ModelOutput(TypedDict):
+    """The output batch produced by the model forward pass."""
+
+    generated_text: List[str]
+    """The text generated by the model."""
+
+    input_ids: NotRequired[torch.Tensor | None]
+    """The token ids of the input text."""
+
+    output_ids: NotRequired[torch.Tensor | None]
+    """The token ids of the model output (usually containing both input and prediction)."""
+
+    attention_mask: NotRequired[torch.Tensor | None]
+    """The attention mask for the input tokens."""

eva/language/models/wrappers/__init__.py

@@ -1,11 +1,19 @@
 """Language Model Wrappers API."""
 
-from eva.language.models.wrappers.huggingface import HuggingFaceTextModel
-from eva.language.models.wrappers.litellm import LiteLLMTextModel
+from eva.language.models.wrappers.base import LanguageModel
+from eva.language.models.wrappers.from_registry import ModelFromRegistry
+from eva.language.models.wrappers.huggingface import HuggingFaceModel
+from eva.language.models.wrappers.litellm import LiteLLMModel
 
 try:
-    from eva.language.models.wrappers.vllm import VLLMTextModel
+    from eva.language.models.wrappers.vllm import VllmModel
 
-    __all__ = ["HuggingFaceTextModel", "LiteLLMTextModel", "VLLMTextModel"]
+    __all__ = [
+        "LanguageModel",
+        "HuggingFaceModel",
+        "LiteLLMModel",
+        "VllmModel",
+        "ModelFromRegistry",
+    ]
 except ImportError:
-    __all__ = ["HuggingFaceTextModel", "LiteLLMTextModel"]
+    __all__ = ["LanguageModel", "HuggingFaceModel", "LiteLLMModel", "ModelFromRegistry"]

eva/language/models/wrappers/base.py

@@ -0,0 +1,47 @@
+"""Base class for language model wrappers."""
+
+import abc
+from typing import Any, Callable
+
+from typing_extensions import override
+
+from eva.core.models.wrappers import base
+from eva.language.data.messages import ModelSystemMessage
+from eva.language.models.typings import ModelOutput, TextBatch
+
+
+class LanguageModel(base.BaseModel[TextBatch, ModelOutput]):
+    """Base class for language models.
+
+    Classes that inherit from this should implement the following methods:
+    - `load_model`: Loads & instantiates the model.
+    - `model_forward`: Implements the forward pass of the model. For API models,
+        this can be an API call.
+    - `format_inputs`: Preprocesses and converts the input batch into the format
+        expected by the `model_forward` method.
+    """
+
+    def __init__(
+        self, system_prompt: str | None, output_transforms: Callable | None = None
+    ) -> None:
+        """Creates a new model instance.
+
+        Args:
+            system_prompt: The system prompt to use for the model (optional).
+            output_transforms: Optional transforms to apply to the output of
+                the model's forward pass.
+        """
+        super().__init__(transforms=output_transforms)
+
+        self.system_message = ModelSystemMessage(content=system_prompt) if system_prompt else None
+
+    @override
+    def forward(self, batch: TextBatch) -> ModelOutput:
+        """Forward pass of the model."""
+        inputs = self.format_inputs(batch)
+        return super().forward(inputs)
+
+    @abc.abstractmethod
+    def format_inputs(self, batch: TextBatch) -> Any:
+        """Converts the inputs into the format expected by the model."""
+        raise NotImplementedError

eva/language/models/wrappers/from_registry.py

@@ -0,0 +1,54 @@
+"""Vision backbone helper class."""
+
+from typing import Any, Callable, Dict, List
+
+from torch import nn
+from typing_extensions import override
+
+from eva.core.models.wrappers import base
+from eva.core.utils import factory
+from eva.language.models.networks.registry import model_registry
+from eva.language.models.typings import TextBatch
+
+
+class ModelFromRegistry(base.BaseModel[TextBatch, List[str]]):
+    """Wrapper class for vision backbone models.
+
+    This class can be used by load backbones available in eva's
+    model registry by name. New backbones can be registered by using
+    the `@backbone_registry.register(model_name)` decorator.
+    """
+
+    def __init__(
+        self,
+        model_name: str,
+        model_kwargs: Dict[str, Any] | None = None,
+        model_extra_kwargs: Dict[str, Any] | None = None,
+        transforms: Callable | None = None,
+    ) -> None:
+        """Initializes the model.
+
+        Args:
+            model_name: The name of the model to load.
+            model_kwargs: The arguments used for instantiating the model.
+            model_extra_kwargs: Extra arguments used for instantiating the model.
+            transforms: The transforms to apply to the output tensor
+                produced by the model.
+        """
+        super().__init__(transforms=transforms)
+
+        self._model_name = model_name
+        self._model_kwargs = model_kwargs or {}
+        self._model_extra_kwargs = model_extra_kwargs or {}
+
+        self.model = self.load_model()
+
+    @override
+    def load_model(self) -> nn.Module:
+        ModelFromRegistry.__name__ = self._model_name
+
+        return factory.ModuleFactory(
+            registry=model_registry,
+            name=self._model_name,
+            init_args=self._model_kwargs | self._model_extra_kwargs,
+        )