oumi.core.processors

oumi.core.processors#

Submodules#

oumi.core.processors.base_image_processor module#

class oumi.core.processors.base_image_processor.BaseImageProcessor[source]#

Bases: ABC

Base class for oumi image processors.

abstractmethod __call__(images: list[Image], *, return_tensors: str | None = 'pt') → BatchFeature[source]#

Extracts image features.

Parameters:

  • images – A list of input images.

  • return_tensors – The format of returned tensors.

Returns:

The model-specific input features.

Return type:

transformers.BatchFeature

oumi.core.processors.base_processor module#

class oumi.core.processors.base_processor.BaseProcessor[source]#

Bases: ABC

Base class for oumi processors.

The high-level purpose of a processor is to generate model-specific input features from input data such as text, images, conversations, etc.

abstractmethod __call__(*, text: list[str], padding: bool, images: list[Image] | None = None, return_tensors: str | None = 'pt') → BatchEncoding[source]#

Invokes the processor to extract features.

Parameters:

  • text – A list of text prompts.

  • padding – Whether to pad sequences to common length.

  • images – A list of input images.

  • return_tensors – The format of returned tensors.

Returns:

The model-specific input features.

Return type:

transformers.BatchEncoding

abstractmethod apply_chat_template(conversation: list[Message], add_generation_prompt: bool = False) → str[source]#

Applies a chat template.

Parameters:

  • conversation – A list of messages (conversation “turns”).

  • add_generation_prompt – Whether to append generation prompt to the output.

Returns:

A text prompt, which includes all input messages formatted into a string.

abstract property chat_template: str#

Returns a chat template.

abstract property ignore_features: list[str]#

Returns a list of keys of features to ignore from feeding the model.

abstract property image_processor: BaseImageProcessor | None#

Returns an image processor.

abstract property image_token: str | None#

Returns an image token.

abstract property image_token_id: int | None#

Returns an image token id.

abstract property label_ignore_index: int | None#

Returns a label ignore index.

abstract property processor_name: str#

Returns a processor name.

abstract property raw_processor: Callable#

Returns the underlying raw processor.

The use of this method is generally discouraged. Only use it if you know what you are doing e.g., direct access to the underlying processor is required by some third-party library.

abstractmethod save_config(output_dir: Path | str) → None[source]#

Saves processor config to the directory.

abstract property tokenizer: PreTrainedTokenizerBase#

Returns a tokenizer associated with this processor.

abstractmethod truncate_text(text: str, *, max_tokens: int, truncation_side: str = 'right') → tuple[str, int][source]#

Truncates text to max_length in tokens.

Parameters:

  • text – A text prompt.

  • max_tokens – Maximum number of tokens to keep.

  • truncation_side – The side to truncate the tokens (“right” or “left”).

Returns:

A tuple containing truncated text prompt and the number of tokens.

oumi.core.processors.default_image_processor module#

class oumi.core.processors.default_image_processor.DefaultImageProcessor(worker_processor: Any)[source]#

Bases: BaseImageProcessor

Default implementation of image processor that wraps a callable object.

__call__(images: list[Image], *, return_tensors: str | None = 'pt') → BatchFeature[source]#

Extracts image features.

Parameters:

  • images – A list of input images.

  • return_tensors – The format of returned tensors.

Returns:

The model-specific input features.

Return type:

transformers.BatchFeature

oumi.core.processors.default_processor module#

class oumi.core.processors.default_processor.DefaultProcessor(processor_name: str, worker_processor: Any, tokenizer: PreTrainedTokenizerBase, *, label_ignore_index: int | None, ignore_features: list[str] | None = None)[source]#

Bases: BaseProcessor

Default implementation of processor that wraps a worker processor.

Validates that worker conforms to basic required invariants.

__call__(*, text: list[str], padding: bool, images: list[Image] | None = None, return_tensors: str | None = 'pt') → BatchEncoding[source]#

Invokes the processor to extract features.

Parameters:

  • text – A list of text prompts.

  • padding – Whether to pad sequences to common length.

  • images – A list of input images.

  • return_tensors – The format of returned tensors.

Returns:

The model-specific input features.

Return type:

transformers.BatchEncoding

apply_chat_template(conversation: list[Message], add_generation_prompt: bool = False) → str[source]#

Applies a chat template.

Parameters:

  • conversation – A list of messages (conversation “turns”).

  • add_generation_prompt – Whether to append generation prompt to the output.

Returns:

A text prompt, which includes all input messages formatted into a string.

property chat_template: str#

Returns a chat template.

property ignore_features: list[str]#

Returns a list of keys of features to ignore from feeding the model.

property image_processor: BaseImageProcessor | None#

Returns an image processor.

property image_token: str | None#

Returns an image token.

property image_token_id: int | None#

Returns an image token id.

property label_ignore_index: int | None#

Returns a label ignore index.

property processor_name: str#

Returns a processor name.

property raw_processor: Callable#

Returns the underlying raw processor.

save_config(output_dir: Path | str) → None[source]#

Saves processor config to the directory.

property tokenizer: PreTrainedTokenizerBase#

Returns a tokenizer associated with this processor.

truncate_text(text: str, *, max_tokens: int, truncation_side: str = 'right') → tuple[str, int][source]#

Truncates text to max_length in tokens.

Parameters:

  • text – A text prompt.

  • max_tokens – Maximum number of tokens to keep.

  • truncation_side – The side to truncate the tokens (“right” or “left”).

Returns:

A tuple containing truncated text prompt and the number of tokens.

Contents