Source code for transformers.models.vit.feature_extraction_vit

# coding=utf-8
# Copyright Google AI and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
"""Feature extractor class for ViT."""

from typing import List, Optional, Union

import numpy as np
from PIL import Image

from ...feature_extraction_utils import BatchFeature, FeatureExtractionMixin
from ...file_utils import TensorType
from ...image_utils import ImageFeatureExtractionMixin, is_torch_tensor
from ...utils import logging


logger = logging.get_logger(__name__)


class ViTFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
    r"""
    Constructs a ViT feature extractor.

    This feature extractor inherits from :class:`~transformers.FeatureExtractionMixin` which contains most of the main
    methods. Users should refer to this superclass for more information regarding those methods.

    Args:
        image_mean (:obj:`int`, defaults to :obj:`[0.5, 0.5, 0.5]`):
            The sequence of means for each channel, to be used when normalizing images.
        image_std (:obj:`int`, defaults to :obj:`[0.5, 0.5, 0.5]`):
            The sequence of standard deviations for each channel, to be used when normalizing images.
        do_normalize (:obj:`bool`, `optional`, defaults to :obj:`True`):
            Whether or not to normalize the input with mean and standard deviation.
        do_resize (:obj:`bool`, `optional`, defaults to :obj:`True`):
            Whether to resize the input to a certain :obj:`size`.
        size (:obj:`int`, `optional`, defaults to 224):
            Resize the input to the given size. Only has an effect if :obj:`do_resize` is set to :obj:`True`.
    """

    model_input_names = ["pixel_values"]

    def __init__(self, image_mean=None, image_std=None, do_normalize=True, do_resize=True, size=224, **kwargs):
        super().__init__(**kwargs)
        self.image_mean = [0.5, 0.5, 0.5]
        self.image_std = [0.5, 0.5, 0.5]
        self.do_normalize = do_normalize
        self.do_resize = do_resize
        self.size = size

    def __call__(
        self,
        images: Union[
            Image.Image, np.ndarray, "torch.Tensor", List[Image.Image], List[np.ndarray], List["torch.Tensor"]  # noqa
        ],
        return_tensors: Optional[Union[str, TensorType]] = None,
        **kwargs
    ) -> BatchFeature:
        """
        Main method to prepare for the model one or several image(s).

        .. warning::

           NumPy arrays and PyTorch tensors are converted to PIL images when resizing, so the most efficient is to pass
           PIL images.

        Args:
            images (:obj:`PIL.Image.Image`, :obj:`np.ndarray`, :obj:`torch.Tensor`, :obj:`List[PIL.Image.Image]`, :obj:`List[np.ndarray]`, :obj:`List[torch.Tensor]`):
                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
                number of channels, H and W are image height and width.

            return_tensors (:obj:`str` or :class:`~transformers.file_utils.TensorType`, `optional`):
                If set, will return tensors instead of list of python integers. Acceptable values are:

                * :obj:`'tf'`: Return TensorFlow :obj:`tf.constant` objects.
                * :obj:`'pt'`: Return PyTorch :obj:`torch.Tensor` objects.
                * :obj:`'np'`: Return Numpy :obj:`np.ndarray` objects.s
                * :obj:`'jax'`: Return JAX :obj:`jnp.ndarray` objects.

        Returns:
            :class:`~transformers.BatchFeature`: A :class:`~transformers.BatchFeature` with the following fields:

            - **pixel_values** -- Pixel values to be fed to a model.
        """
        # Input type checking for clearer error
        valid_images = False

        # Check that images has a valid type
        if isinstance(images, (Image.Image, np.ndarray)) or is_torch_tensor(images):
            valid_images = True
        elif isinstance(images, (list, tuple)):
            if len(images) == 0 or isinstance(images[0], (Image.Image, np.ndarray)) or is_torch_tensor(images[0]):
                valid_images = True

        if not valid_images:
            raise ValueError(
                "Images must of type `PIL.Image.Image`, `np.ndarray` or `torch.Tensor` (single example),"
                "`List[PIL.Image.Image]`, `List[np.ndarray]` or `List[torch.Tensor]` (batch of examples)."
            )

        is_batched = bool(
            isinstance(images, (list, tuple))
            and (isinstance(images[0], (Image.Image, np.ndarray)) or is_torch_tensor(images[0]))
        )

        if not is_batched:
            images = [images]

        # transformations (resizing + normalization)
        if self.do_resize and self.size is not None:
            images = [self.resize(image=image, size=self.size) for image in images]
        if self.do_normalize:
            images = [self.normalize(image=image, mean=self.image_mean, std=self.image_std) for image in images]

        # return as BatchFeature
        data = {"pixel_values": images}
        encoded_inputs = BatchFeature(data=data, tensor_type=return_tensors)

        return encoded_inputs