Add BLIP (#20716)

* add new model like

* add v1

* v1

* v1

* vision encoder logits match

* v2

* fix

* add docstring

* CI tests pass

* fix tests

* make fixup

* add to `toctree`

* fix processors

* fix processors

* fix doc

* fill title

* add content doc

* remove from tokenization auto

* fix config

* change order

* add `# Copied from`

* few fixes

- add correct license on modeling text
- remove dummy argument

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* replace name

* refactor a bit

* more refactor

* remove unused arg

* make fixup + remove some `# Adapted from ...`

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* more `# Copied from`

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* now `generate` supports no prefix

* remove `FeatureExtractor`

* fix path

* correct dependency

* fix tests

* few fixes

* add integration tests

* add correct conversion script

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* add `blip` to tokenization auto

* fix docstrings

* fix test + add image

* remove processor from uncorrect place

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* clean up a bit

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* clean pixel mask

* clean pixel mask

* fix `F`

* Update src/transformers/models/blip/modeling_blip.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fix output

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fix pad token id

* remove `token_type_ids`

* make fixup

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* make fixup

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* add comments

* Update src/transformers/models/blip/modeling_blip.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* remove `token_type_ids`

* make fixup

* better name

* replace with `image_attention_mask`

* refactor

* make fixup

* better docstring

* replace `answer_xx`

* remove ununsed args

* add `labels`

* add `labels`

* fix processing tests

* make fixup

* make fixup

* put correct repo

* remove `pad`

* remove `crop` and `center_crop`

* Update src/transformers/models/blip/image_processing_blip.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fix

* remove `size_divisor`

* fix weights `init`

* remove unneeded functions

* add suggestions

* minor changes

- change slow test output for PT 1.13
- docstring order

* replace `feature_extractor` by `image_processor`

* fix doctests

* fix weight init order + add fp16 slow test

* add `blip` to doctest

* add correct repo name and fix test

* Update src/transformers/models/blip/processing_blip.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fix tests

* use `convert_to_rgb` from `image_transforms`

* make fixup

* fix large loading issue
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/models/blip/image_processing_blip.py

+# coding=utf-8
+# Copyright 2022 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.
+"""Image processor class for BLIP."""
+
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+
+from transformers.utils import is_vision_available
+from transformers.utils.generic import TensorType
+
+from ...image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from ...image_transforms import convert_to_rgb, normalize, rescale, resize, to_channel_dimension_format
+from ...image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    is_batched,
+    to_numpy_array,
+    valid_images,
+)
+from ...utils import logging
+
+
+if is_vision_available():
+    import PIL
+
+
+logger = logging.get_logger(__name__)
+
+
+class BlipImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a BLIP image processor.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by the
+            `do_resize` parameter in the `preprocess` method.
+        size (`dict`, *optional*, defaults to `{"height": 384, "width": 384}`):
+            Size of the output image after resizing. Can be overridden by the `size` parameter in the `preprocess`
+            method.
+        resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Only has an effect if `do_resize` is set to `True`. Can be
+            overridden by the `resample` parameter in the `preprocess` method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Wwhether to rescale the image by the specified scale `rescale_factor`. Can be overridden by the
+            `do_rescale` parameter in the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Only has an effect if `do_rescale` is set to `True`. Can be
+            overridden by the `rescale_factor` parameter in the `preprocess` method.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image. Can be overridden by the `do_normalize` parameter in the `preprocess`
+            method. Can be overridden by the `do_normalize` parameter in the `preprocess` method.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_MEAN`):
+            Mean to use if normalizing the image. This is a float or list of floats the length of the number of
+            channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method. Can be
+            overridden by the `image_mean` parameter in the `preprocess` method.
+        image_std (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_STD`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+            Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        **kwargs
+    ) -> None:
+
+        super().__init__(**kwargs)
+        size = size if size is not None else {"height": 384, "width": 384}
+        size = get_size_dict(size, default_to_square=True)
+
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else IMAGENET_STANDARD_MEAN
+        self.image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
+        self.do_convert_rgb = do_convert_rgb
+
+    def resize(
+        self,
+        image: np.ndarray,
+        size: Dict[str, int],
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Resize an image.
+
+        Resizes the shorter side of the image to `size["shortest_edge"]` while preserving the aspect ratio. If the
+        longer side is larger than the max size `(int(`size["shortest_edge"]` * 1333 / 800))`, the longer side is then
+        resized to the max size while preserving the aspect ratio.
+
+        Args:
+            image (`np.ndarray`):
+                Image to resize.
+            size (`Dict[str, int]`):
+                Controls the size of the output image. Should be of the form `{"shortest_edge": int}`.
+            resample (`PILImageResampling` filter, *optional*, defaults to `PILImageResampling.BICUBIC`):
+                Resampling filter to use when resiizing the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        size = get_size_dict(size, default_to_square=True)
+        output_size = (size["width"], size["height"])
+        return resize(image, size=output_size, resample=resample, data_format=data_format, **kwargs)
+
+    def rescale(
+        self,
+        image: np.ndarray,
+        scale: Union[int, float],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ):
+        """
+        Rescale an image by a scale factor. image = image * scale.
+
+        Args:
+            image (`np.ndarray`):
+                Image to rescale.
+            scale (`int` or `float`):
+                Scale to apply to the image.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        return rescale(image, scale=scale, data_format=data_format, **kwargs)
+
+    def normalize(
+        self,
+        image: np.ndarray,
+        mean: Union[float, List[float]],
+        std: Union[float, List[float]],
+        data_format: Optional[Union[str, ChannelDimension]] = None,
+        **kwargs
+    ) -> np.ndarray:
+        """
+        Normalize an image. image = (image - image_mean) / image_std.
+
+        Args:
+            image (`np.ndarray`):
+                Image to normalize.
+            mean (`float` or `List[float]`):
+                Image mean.
+            std (`float` or `List[float]`):
+                Image standard deviation.
+            data_format (`str` or `ChannelDimension`, *optional*):
+                The channel dimension format of the image. If not provided, it will be the same as the input image.
+        """
+        return normalize(image, mean=mean, std=std, data_format=data_format, **kwargs)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: Optional[bool] = None,
+        size: Optional[Dict[str, int]] = None,
+        resample: PILImageResampling = None,
+        do_rescale: Optional[bool] = None,
+        rescale_factor: Optional[float] = None,
+        do_normalize: Optional[bool] = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        do_convert_rgb: bool = None,
+        data_format: ChannelDimension = ChannelDimension.FIRST,
+        **kwargs,
+    ) -> PIL.Image.Image:
+        """
+        Preprocess an image or batch of images.
+
+        Args:
+            images (`ImageInput`):
+                Image to preprocess.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Controls the size of the image after `resize`. The shortest edge of the image is resized to
+                `size["shortest_edge"]` whilst preserving the aspect ratio. If the longest edge of this resized image
+                is > `int(size["shortest_edge"] * (1333 / 800))`, then the image is resized again to make the longest
+                edge equal to `int(size["shortest_edge"] * (1333 / 800))`.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. Only has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image values between [0 - 1].
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to normalize the image by if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to normalize the image by if `do_normalize` is set to `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                    - Unset: Return a list of `np.ndarray`.
+                    - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                    - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                    - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                    - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                    - `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                    - `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        resample = resample if resample is not None else self.resample
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        size = size if size is not None else self.size
+        size = get_size_dict(size, default_to_square=False)
+
+        if not is_batched(images):
+            images = [images]
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+
+        if do_resize and size is None or resample is None:
+            raise ValueError("Size and resample must be specified if do_resize is True.")
+
+        if do_rescale and rescale_factor is None:
+            raise ValueError("Rescale factor must be specified if do_rescale is True.")
+
+        if do_normalize and (image_mean is None or image_std is None):
+            raise ValueError("Image mean and std must be specified if do_normalize is True.")
+
+        # PIL RGBA images are converted to RGB
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_resize:
+            images = [self.resize(image=image, size=size, resample=resample) for image in images]
+
+        if do_rescale:
+            images = [self.rescale(image=image, scale=rescale_factor) for image in images]
+
+        if do_normalize:
+            images = [self.normalize(image=image, mean=image_mean, std=image_std) for image in images]
+
+        images = [to_channel_dimension_format(image, data_format) for image in images]
+
+        encoded_outputs = BatchFeature(data={"pixel_values": images}, tensor_type=return_tensors)
+
+        return encoded_outputs

src/transformers/models/blip/modeling_blip.py

+# coding=utf-8
+# Copyright 2022 The Salesforce Team Authors and The HuggingFace 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.
+""" PyTorch BLIP model."""
+
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn.functional import normalize
+
+from ...activations import ACT2FN
+from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
+from ...modeling_utils import PreTrainedModel
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_blip import BlipConfig, BlipTextConfig, BlipVisionConfig
+from .modeling_blip_text import BlipTextLMHeadModel, BlipTextModel
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "Salesforce/blip-vqa-base"
+
+BLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "Salesforce/blip-vqa-base",
+    "Salesforce/blip-vqa-capfit-large",
+    "Salesforce/blip-image-captioning-base",
+    "Salesforce/blip-image-captioning-large",
+    "Salesforce/blip-itm-base-coco",
+    "Salesforce/blip-itm-large-coco",
+    "Salesforce/blip-itm-base-flikr",
+    "Salesforce/blip-itm-large-flikr",
+    # See all BLIP models at https://huggingface.co/models?filter=blip
+]
+
+
+# Copied from transformers.models.clip.modeling_clip.contrastive_loss
+def contrastive_loss(logits: torch.Tensor) -> torch.Tensor:
+    return nn.functional.cross_entropy(logits, torch.arange(len(logits), device=logits.device))
+
+
+# Copied from transformers.models.clip.modeling_clip.clip_loss with clip->blip
+def blip_loss(similarity: torch.Tensor) -> torch.Tensor:
+    caption_loss = contrastive_loss(similarity)
+    image_loss = contrastive_loss(similarity.t())
+    return (caption_loss + image_loss) / 2.0
+
+
+@dataclass
+class BlipForConditionalGenerationModelOutput(ModelOutput):
+    """
+    Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the
+    last hidden states. This class also adds the loss term from the text decoder.
+
+    Args:
+        loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
+            Languge modeling loss from the text decoder.
+        decoder_logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`, *optional*):
+            Prediction scores of the language modeling head of the text decoder model.
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)`, *optional*):
+            The image embeddings obtained after applying the Vision Transformer model to the input image.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[Tuple[torch.FloatTensor]] = None
+    decoder_logits: Optional[Tuple[torch.FloatTensor]] = None
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BlipTextVisionModelOutput(ModelOutput):
+    """
+    Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the
+    last hidden states. This class also adds the loss term from the text decoder.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Languge modeling loss from the text decoder.
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BlipImageTextMatchingModelOutput(ModelOutput):
+    """
+    Adapted from the base class for vision model's outputs that also contains image embeddings of the pooling of the
+    last hidden states. This class also adds the loss term from the text decoder as well as the image-text similarity
+    scores.
+
+    Args:
+        itm_score (`torch.FloatTensor`):
+            The image-text similarity scores.
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+            Languge modeling loss from the text decoder.
+        image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+            The image embeddings obtained by applying the projection layer to the pooler_output.
+        last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+            one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        vision_pooler_output (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*):
+            Last layer hidden-state of the vision of the vision-only branch of the model.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        question_embeds (`torch.FloatTensor`):
+            The question embeddings obtained by the text projection layer.
+    """
+
+    itm_score: Optional[torch.FloatTensor] = None
+    loss: Optional[torch.FloatTensor] = None
+    image_embeds: Optional[torch.FloatTensor] = None
+    last_hidden_state: torch.FloatTensor = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    vision_pooler_output: Optional[torch.FloatTensor] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    question_embeds: Optional[Tuple[torch.FloatTensor]] = None
+
+
+@dataclass
+class BlipOutput(ModelOutput):
+    """
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `return_loss` is `True`):
+            Contrastive loss for image-text similarity.
+        logits_per_image:(`torch.FloatTensor` of shape `(image_batch_size, text_batch_size)`):
+            The scaled dot product scores between `image_embeds` and `text_embeds`. This represents the image-text
+            similarity scores.
+        logits_per_text:(`torch.FloatTensor` of shape `(text_batch_size, image_batch_size)`):
+            The scaled dot product scores between `text_embeds` and `image_embeds`. This represents the text-image
+            similarity scores.
+        text_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The text embeddings obtained by applying the projection layer to the pooled output of [`BlipTextModel`].
+        image_embeds(`torch.FloatTensor` of shape `(batch_size, output_dim`):
+            The image embeddings obtained by applying the projection layer to the pooled output of [`BlipVisionModel`].
+        text_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`BlipTextModel`].
+        vision_model_output(`BaseModelOutputWithPooling`):
+            The output of the [`BlipVisionModel`].
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    logits_per_image: torch.FloatTensor = None
+    logits_per_text: torch.FloatTensor = None
+    text_embeds: torch.FloatTensor = None
+    image_embeds: torch.FloatTensor = None
+    text_model_output: BaseModelOutputWithPooling = None
+    vision_model_output: BaseModelOutputWithPooling = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k] if k not in ["text_model_output", "vision_model_output"] else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+class BlipVisionEmbeddings(nn.Module):
+    def __init__(self, config: BlipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(
+            torch.randn(1, 1, self.embed_dim),
+        )
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        embeddings = embeddings + self.position_embedding[:, : embeddings.size(1), :].to(target_dtype)
+        return embeddings
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPTextEmbeddings with CLIP->Blip
+class BlipTextEmbeddings(nn.Module):
+    def __init__(self, config: BlipTextConfig):
+        super().__init__()
+        embed_dim = config.hidden_size
+
+        self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
+        self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+    ) -> torch.Tensor:
+        seq_length = input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if inputs_embeds is None:
+            inputs_embeds = self.token_embedding(input_ids)
+
+        position_embeddings = self.position_embedding(position_ids)
+        embeddings = inputs_embeds + position_embeddings
+
+        return embeddings
+
+
+class BlipAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = nn.Dropout(config.attention_dropout)
+
+        self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim)
+
+        self.projection = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        mixed_qkv = self.qkv(hidden_states)
+        mixed_qkv = (
+            self.qkv(hidden_states)
+            .reshape(bsz, tgt_len, 3, self.num_heads, embed_dim // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        query_states, key_states, value_states = (
+            mixed_qkv[0],
+            mixed_qkv[1],
+            mixed_qkv[2],
+        )
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2))
+
+        attention_scores = attention_scores * self.scale
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_states).permute(0, 2, 1, 3)
+
+        new_context_layer_shape = context_layer.size()[:-2] + (self.embed_dim,)
+        context_layer = context_layer.reshape(new_context_layer_shape)
+
+        output = self.projection(context_layer)
+
+        outputs = (output, attention_probs) if output_attentions else (output, None)
+
+        return outputs
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Blip
+class BlipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class BlipEncoderLayer(nn.Module):
+    def __init__(self, config: BlipConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = BlipAttention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim)
+        self.mlp = BlipMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            head_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = hidden_states + residual
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+
+        hidden_states = hidden_states + residual
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class BlipPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BlipConfig
+    base_model_prefix = "blip"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_range
+        if isinstance(module, nn.Conv2d) or isinstance(module, nn.Embedding) or isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=factor)
+            if hasattr(module, "bias") and module.bias is not None:
+                module.bias.data.zero_()
+
+        if isinstance(module, BlipVisionEmbeddings):
+            if hasattr(self.config, "vision_config"):
+                factor = self.config.vision_config.initializer_range
+            nn.init.trunc_normal_(
+                module.position_embedding,
+                mean=0.0,
+                std=factor,
+            )
+
+            nn.init.trunc_normal_(
+                module.class_embedding,
+                mean=0.0,
+                std=factor,
+            )
+
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, BlipEncoder):
+            module.gradient_checkpointing = value
+
+
+BLIP_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`BlipConfig`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+BLIP_TEXT_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`BlipProcessor`]. See [`BlipProcessor.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+BLIP_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`BlipImageProcessor`]. See [`BlipImageProcessor.__call__`] for details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+BLIP_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`BlipProcessor`]. See [`BlipProcessor.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
+            [`BlipImageProcessor`]. See [`BlipImageProcessor.__call__`] for details.
+        return_loss (`bool`, *optional*):
+            Whether or not to return the contrastive loss.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+class BlipEncoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`BlipEncoderLayer`].
+
+    Args:
+        config (`BlipConfig`):
+            The corresponding vision configuration for the `BlipEncoder`.
+    """
+
+    def __init__(self, config: BlipConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([BlipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(encoder_layer),
+                    hidden_states,
+                    attention_mask,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+class BlipVisionModel(BlipPreTrainedModel):
+    main_input_name = "pixel_values"
+    config_class = BlipVisionConfig
+
+    def __init__(self, config: BlipVisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = BlipVisionEmbeddings(config)
+        self.encoder = BlipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim)
+
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=BlipVisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        hidden_states = self.embeddings(pixel_values)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+
+@add_start_docstrings(BLIP_START_DOCSTRING)
+class BlipModel(BlipPreTrainedModel):
+    config_class = BlipConfig
+
+    def __init__(self, config: BlipConfig):
+        super().__init__(config)
+
+        if not isinstance(config.text_config, BlipTextConfig):
+            raise ValueError(
+                "config.text_config is expected to be of type BlipTextConfig but is of type"
+                f" {type(config.text_config)}."
+            )
+
+        if not isinstance(config.vision_config, BlipVisionConfig):
+            raise ValueError(
+                "config.vision_config is expected to be of type BlipVisionConfig but is of type"
+                f" {type(config.vision_config)}."
+            )
+
+        text_config = config.text_config
+        vision_config = config.vision_config
+
+        self.projection_dim = config.projection_dim
+        self.text_embed_dim = text_config.hidden_size
+        self.vision_embed_dim = vision_config.hidden_size
+
+        self.text_model = BlipTextModel(text_config)
+        self.vision_model = BlipVisionModel(vision_config)
+
+        self.visual_projection = nn.Linear(self.vision_embed_dim, self.projection_dim, bias=False)
+        self.text_projection = nn.Linear(self.text_embed_dim, self.projection_dim, bias=False)
+        self.logit_scale = nn.Parameter(torch.ones([]) * self.config.logit_scale_init_value)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BLIP_TEXT_INPUTS_DOCSTRING)
+    def get_text_features(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            text_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The text embeddings obtained by
+            applying the projection layer to the pooled output of [`BlipTextModel`].
+
+        Examples:
+
+        ```python
+        >>> from transformers import BlipProcessor, BlipModel
+
+        >>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+
+        >>> inputs = processor(text=["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
+        >>> text_features = model.get_text_features(**inputs)
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            return_dict=return_dict,
+        )
+
+        pooled_output = text_outputs[1]
+        text_features = self.text_projection(pooled_output)
+
+        return text_features
+
+    @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
+    def get_image_features(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> torch.FloatTensor:
+        r"""
+        Returns:
+            image_features (`torch.FloatTensor` of shape `(batch_size, output_dim`): The image embeddings obtained by
+            applying the projection layer to the pooled output of [`BlipVisionModel`].
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipModel
+
+        >>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> image_features = model.get_image_features(**inputs)
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            return_dict=return_dict,
+        )
+
+        pooled_output = vision_outputs[1]  # pooled_output
+        image_features = self.visual_projection(pooled_output)
+
+        return image_features
+
+    @add_start_docstrings_to_model_forward(BLIP_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BlipOutput, config_class=BlipConfig)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        return_loss: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BlipOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipModel
+
+        >>> model = BlipModel.from_pretrained("Salesforce/blip-image-captioning-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(
+        ...     text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True
+        ... )
+
+        >>> outputs = model(**inputs)
+        >>> logits_per_image = outputs.logits_per_image  # this is the image-text similarity score
+        >>> probs = logits_per_image.softmax(dim=1)  # we can take the softmax to get the label probabilities
+        ```"""
+        # Use BLIP model's config for some fields (if specified) instead of those of vision & text components.
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        text_outputs = self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[1]
+        image_embeds = self.visual_projection(image_embeds)
+
+        text_embeds = text_outputs[1]
+        text_embeds = self.text_projection(text_embeds)
+
+        # normalized features
+        image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
+        text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
+
+        # cosine similarity as logits
+        logit_scale = self.logit_scale.exp()
+        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
+        logits_per_image = logits_per_text.t()
+
+        loss = None
+        if return_loss:
+            loss = blip_loss(logits_per_text)
+
+        if not return_dict:
+            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_outputs, vision_outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return BlipOutput(
+            loss=loss,
+            logits_per_image=logits_per_image,
+            logits_per_text=logits_per_text,
+            text_embeds=text_embeds,
+            image_embeds=image_embeds,
+            text_model_output=text_outputs,
+            vision_model_output=vision_outputs,
+        )
+
+
+@add_start_docstrings(
+    """
+    BLIP Model for image captioning. The model consists of a vision encoder and a text decoder. One can optionally pass
+    `input_ids` to the model, which serve as a text prompt, to make the text decoder continue the prompt. Otherwise,
+    the decoder starts generating text from the [BOS] (beginning-of-sequence) token. will start generating the caption
+    from the text input. If no text input is provided, the decoder will start with the [BOS] token only.
+    """,
+    BLIP_START_DOCSTRING,
+)
+class BlipForConditionalGeneration(BlipPreTrainedModel):
+    config_class = BlipConfig
+    _keys_to_ignore_on_load_missing = [r"text_decoder.cls.predictions.decoder.bias"]
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: BlipConfig):
+        super().__init__(config)
+
+        self.vision_model = BlipVisionModel(config.vision_config)
+
+        self.text_decoder = BlipTextLMHeadModel(config.text_config)
+
+        self.decoder_input_ids = config.text_config.bos_token_id
+        self.decoder_pad_token_id = config.text_config.pad_token_id
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BlipForConditionalGenerationModelOutput, config_class=BlipVisionConfig)
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BlipForConditionalGenerationModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipForConditionalGeneration
+
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+        >>> model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        ```"""
+        batch_size = pixel_values.shape[0]
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[0]
+
+        if input_ids is None:
+            input_ids = torch.LongTensor([[self.decoder_input_ids] * batch_size]).to(image_embeds.device)
+
+        if labels is None:
+            labels = input_ids.masked_fill(input_ids == self.decoder_pad_token_id, -100)
+
+        outputs = self.text_decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=image_embeds,
+            labels=labels,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            outputs = (outputs[0], outputs[1], image_embeds, vision_outputs[0]) + vision_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return BlipForConditionalGenerationModelOutput(
+            loss=outputs.loss,
+            decoder_logits=outputs.logits,
+            image_embeds=image_embeds,
+            last_hidden_state=vision_outputs.last_hidden_state,
+            hidden_states=vision_outputs.hidden_states,
+            attentions=vision_outputs.attentions,
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        **generate_kwargs
+    ) -> torch.LongTensor:
+        r"""
+        Overrides *generate* function to be able to use the model as a conditional generator
+
+        Parameters:
+            pixel_values (*torch.FloatTensor* of shape *(batch_size, image_width, image_height)*:
+                Input image to be processed
+            input_ids (*torch.LongTensor* of shape *(batch_size, sequence_length)*, *optional*):
+                The sequence used as a prompt for the generation.
+            attention_mask (*torch.LongTensor* of shape *(batch_size, sequence_length)*, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+
+        Examples:
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipForConditionalGeneration
+
+        >>> model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt")
+
+        >>> outputs = model.generate(**inputs)
+        >>> print(processor.decode(outputs[0], skip_special_tokens=True))
+        two cats are laying on a couch
+        ```
+        """
+
+        batch_size = pixel_values.shape[0]
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+        )
+
+        image_embeds = vision_outputs[0]
+
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image_embeds.device)
+
+        if isinstance(input_ids, list):
+            input_ids = torch.LongTensor(input_ids)
+        elif input_ids is None:
+            input_ids = (
+                torch.LongTensor([[self.decoder_input_ids, self.config.text_config.eos_token_id]])
+                .repeat(batch_size, 1)
+                .to(image_embeds.device)
+            )
+
+        input_ids[:, 0] = self.config.text_config.bos_token_id
+        attention_mask = attention_mask[:, :-1] if attention_mask is not None else None
+
+        outputs = self.text_decoder.generate(
+            input_ids=input_ids[:, :-1],
+            eos_token_id=self.config.text_config.sep_token_id,
+            pad_token_id=self.config.text_config.pad_token_id,
+            attention_mask=attention_mask,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            **generate_kwargs,
+        )
+
+        return outputs
+
+
+@add_start_docstrings(
+    """
+    BLIP Model for visual question answering. The model consists of a vision encoder, a text encoder as well as a text
+    decoder. The vision encoder will encode the input image, the text encoder will encode the input question together
+    with the encoding of the image, and the text decoder will output the answer to the question.
+    """,
+    BLIP_START_DOCSTRING,
+)
+class BlipForQuestionAnswering(BlipPreTrainedModel):
+    config_class = BlipConfig
+    _keys_to_ignore_on_load_missing = [r"text_decoder.cls.predictions.decoder.bias"]
+
+    def __init__(self, config: BlipConfig):
+        super().__init__(config)
+
+        self.vision_model = BlipVisionModel(config.vision_config)
+
+        self.text_encoder = BlipTextModel(config.text_config, add_pooling_layer=False)
+
+        self.text_decoder = BlipTextLMHeadModel(config.text_config)
+
+        self.decoder_pad_token_id = config.text_config.pad_token_id
+        self.decoder_bos_token_id = config.text_config.bos_token_id
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BlipTextVisionModelOutput, config_class=BlipVisionConfig)
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        pixel_values: torch.FloatTensor,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BlipTextVisionModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipForQuestionAnswering
+
+        >>> model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> text = "How many cats are in the picture?"
+
+        >>> inputs = processor(images=image, text=text, return_tensors="pt")
+
+        >>> outputs = model(**inputs)
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        batch_size = input_ids.shape[0]
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[0]
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long)
+
+        question_embeds = self.text_encoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            return_dict=return_dict,
+        )
+
+        question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
+
+        if decoder_input_ids is None:
+            decoder_input_ids = torch.LongTensor([self.decoder_bos_token_id]).repeat((batch_size, 1))
+
+        if labels is None:
+            labels = decoder_input_ids.masked_fill(decoder_input_ids == self.decoder_pad_token_id, -100)
+
+        answer_output = self.text_decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=question_embeds,
+            encoder_attention_mask=attention_mask,
+            labels=labels,
+            return_dict=return_dict,
+            reduction="none",
+        )
+
+        decoder_loss = answer_output.loss.mean() if return_dict else answer_output[0].mean()
+
+        if not return_dict:
+            outputs = (decoder_loss, image_embeds, vision_outputs[0]) + vision_outputs[2:]
+            return tuple(output for output in outputs if output is not None)
+
+        return BlipTextVisionModelOutput(
+            loss=decoder_loss,
+            image_embeds=image_embeds,
+            last_hidden_state=vision_outputs.last_hidden_state,
+            hidden_states=vision_outputs.hidden_states,
+            attentions=vision_outputs.attentions,
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        input_ids: torch.LongTensor,
+        pixel_values: torch.FloatTensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        **generate_kwargs
+    ) -> torch.LongTensor:
+        r"""
+        Overrides *generate* function to be able to use the model as a conditional generator
+
+        Parameters:
+            input_ids (*torch.LongTensor* of shape *(batch_size, sequence_length)*):
+                The sequence used as a prompt for the generation.
+            pixel_values (*torch.FloatTensor* of shape *(batch_size, image_width, image_height)*:
+                Input image to be processed
+            attention_mask (*torch.LongTensor* of shape *(batch_size, sequence_length)*, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`. `1` for
+                tokens that are NOT MASKED, `0` for MASKED tokens.
+            **generate_kwargs:
+                Additional arguments passed to the *generate* function of the decoder
+
+
+        Examples:
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipForQuestionAnswering
+
+        >>> model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> text = "How many cats are in the picture?"
+
+        >>> inputs = processor(images=image, text=text, return_tensors="pt")
+
+        >>> outputs = model.generate(**inputs)
+        >>> print(processor.decode(outputs[0], skip_special_tokens=True))
+        2
+        ```
+        """
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+        )
+
+        image_embeds = vision_outputs[0]
+
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(image_embeds.device)
+
+        if isinstance(input_ids, list):
+            input_ids = torch.LongTensor(input_ids)
+
+        question_outputs = self.text_encoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            return_dict=False,
+        )
+
+        question_embeds = question_outputs[0]
+
+        question_attention_mask = torch.ones(question_embeds.size()[:-1], dtype=torch.long).to(question_embeds.device)
+
+        bos_ids = torch.full(
+            (question_embeds.size(0), 1), fill_value=self.decoder_bos_token_id, device=question_embeds.device
+        )
+
+        outputs = self.text_decoder.generate(
+            input_ids=bos_ids,
+            eos_token_id=self.config.text_config.sep_token_id,
+            pad_token_id=self.config.text_config.pad_token_id,
+            encoder_hidden_states=question_embeds,
+            encoder_attention_mask=question_attention_mask,
+            **generate_kwargs,
+        )
+
+        return outputs
+
+
+@add_start_docstrings(
+    """
+    BLIP Model with a vision and text projector, and a classification head on top. The model is used in the context of
+    image-text retrieval. Given an image and a text, the model returns the probability of the text being relevant to
+    the image.
+    """,
+    BLIP_START_DOCSTRING,
+)
+class BlipForImageTextRetrieval(BlipPreTrainedModel):
+    config_class = BlipConfig
+
+    def __init__(self, config: BlipConfig):
+        super().__init__(config)
+
+        self.vision_model = BlipVisionModel(config.vision_config)
+
+        self.text_encoder = BlipTextModel(config.text_config, add_pooling_layer=False)
+
+        # vision projection layer
+        self.vision_proj = nn.Linear(config.vision_config.hidden_size, config.image_text_hidden_size)
+
+        # text projection layer
+        self.text_proj = nn.Linear(config.text_config.hidden_size, config.image_text_hidden_size)
+
+        # image text matching head
+        self.itm_head = nn.Linear(config.text_config.hidden_size, 2)
+
+        self.decoder_pad_token_id = config.text_config.pad_token_id
+        self.decoder_bos_token_id = config.text_config.bos_token_id
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BlipTextVisionModelOutput, config_class=BlipVisionConfig)
+    def forward(
+        self,
+        input_ids: torch.LongTensor,
+        pixel_values: torch.FloatTensor,
+        use_itm_head: Optional[bool] = True,
+        attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BlipTextVisionModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import BlipProcessor, BlipForImageTextRetrieval
+
+        >>> model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-base")
+        >>> processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-base")
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> text = "an image of a cat"
+
+        >>> inputs = processor(images=image, text=text, return_tensors="pt")
+        >>> outputs = model(**inputs)
+        ```
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        image_embeds = vision_outputs[0]
+        image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long)
+
+        if use_itm_head:
+            question_embeds = self.text_encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                encoder_hidden_states=image_embeds,
+                encoder_attention_mask=image_atts,
+                return_dict=return_dict,
+            )
+            question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
+
+            output = self.itm_head(question_embeds[:, 0, :])
+        else:
+            question_embeds = self.text_encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                return_dict=return_dict,
+            )
+            question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
+
+            image_feat = normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)
+            text_feat = normalize(self.text_proj(question_embeds[:, 0, :]), dim=-1)
+
+            output = image_feat @ text_feat.t()
+
+        if not return_dict:
+            outputs = (output, vision_outputs[0]) + vision_outputs[2:] + (question_embeds,)
+            return tuple(output for output in outputs if output is not None)
+
+        return BlipImageTextMatchingModelOutput(
+            itm_score=output,
+            last_hidden_state=vision_outputs.last_hidden_state,
+            hidden_states=vision_outputs.hidden_states,
+            attentions=vision_outputs.attentions,
+            question_embeds=question_embeds,
+        )

src/transformers/models/blip/modeling_blip_text.py

+# coding=utf-8
+# Copyright 2022 The Salesforce Team Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the BSD-3-clause license (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://opensource.org/licenses/BSD-3-Clause
+#
+# 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.
+
+
+import math
+from typing import Optional, Tuple
+
+import torch
+import torch.utils.checkpoint
+from torch import Tensor, device, nn
+from torch.nn import CrossEntropyLoss
+
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+)
+from transformers.modeling_utils import (
+    PreTrainedModel,
+    apply_chunking_to_forward,
+    find_pruneable_heads_and_indices,
+    prune_linear_layer,
+)
+from transformers.utils import logging
+
+from .configuration_blip import BlipTextConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L52
+class BlipTextEmbeddings(nn.Module):
+    """Construct the embeddings from word and position embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+
+        self.config = config
+
+    def forward(self, input_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
+
+        if inputs_embeds is None:
+            input_ids = input_ids.to(self.word_embeddings.weight.device)
+            inputs_embeds = self.word_embeddings(input_ids)
+
+        embeddings = inputs_embeds
+
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L97
+class BlipTextSelfAttention(nn.Module):
+    def __init__(self, config, is_cross_attention):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention heads (%d)"
+                % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.encoder_hidden_size, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BlipTextModel forward() function)
+            attention_scores = attention_scores + attention_mask.to(attention_scores.device)
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert -> BlipText
+class BlipTextSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#242
+class BlipTextAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.self = BlipTextSelfAttention(config, is_cross_attention)
+        self.output = BlipTextSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ):
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert -> BlipText
+class BlipTextIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert -> BlipText
+class BlipTextOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class BlipTextLayer(nn.Module):
+    def __init__(self, config, layer_num):
+        super().__init__()
+        self.config = config
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = BlipTextAttention(config)
+        self.layer_num = layer_num
+        if self.config.is_decoder:
+            self.crossattention = BlipTextAttention(config, is_cross_attention=self.config.is_decoder)
+        self.intermediate = BlipTextIntermediate(config)
+        self.output = BlipTextOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        outputs = self_attention_outputs[1:-1]
+        present_key_value = self_attention_outputs[-1]
+
+        if encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+        layer_output = apply_chunking_to_forward(
+            self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
+        )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L386
+class BlipTextEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([BlipTextLayer(config, i) for i in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.is_decoder else None
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPooler with Bert->BlipText
+class BlipTextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertPredictionHeadTransform with Bert->BlipText
+class BlipTextPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertLMPredictionHead with Bert->BlipText
+class BlipTextLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = BlipTextPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOnlyMLMHead with Bert->BlipText
+class BlipTextOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = BlipTextLMPredictionHead(config)
+
+    def forward(self, sequence_output: torch.Tensor) -> torch.Tensor:
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L548
+class BlipTextPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = BlipTextConfig
+    base_model_prefix = "bert"
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        if isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/3a29b7410476bf5f2ba0955827390eb6ea1f4f9d/models/med.py#L571
+class BlipTextModel(BlipTextPreTrainedModel):
+    """
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in [Attention is
+    all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,
+    Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. argument and `is_decoder` set to `True`; an
+    `encoder_hidden_states` is then expected as an input to the forward pass.
+    """
+
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BlipTextEmbeddings(config)
+        self.encoder = BlipTextEncoder(config)
+        self.pooler = BlipTextPooler(config) if add_pooling_layer else None
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel._prune_heads
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self, attention_mask: Tensor, input_shape: Tuple[int], device: device, is_decoder: bool
+    ) -> Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (`Tuple[int]`):
+                The shape of the input to the model.
+            device: (`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - if the model is a decoder, apply a causal mask in addition to the padding mask
+            # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            if is_decoder:
+                batch_size, seq_length = input_shape
+
+                seq_ids = torch.arange(seq_length, device=device)
+                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                # in case past_key_values are used we need to add a prefix ones mask to the causal mask
+                # causal and attention masks must have same type with pytorch version < 1.3
+                causal_mask = causal_mask.to(attention_mask.dtype)
+
+                if causal_mask.shape[1] < attention_mask.shape[1]:
+                    prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]
+                    causal_mask = torch.cat(
+                        [
+                            torch.ones(
+                                (batch_size, seq_length, prefix_seq_len), device=device, dtype=causal_mask.dtype
+                            ),
+                            causal_mask,
+                        ],
+                        axis=-1,
+                    )
+
+                extended_attention_mask = causal_mask[:, None, :, :] * attention_mask[:, None, None, :]
+            else:
+                extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        is_decoder=False,
+    ):
+        r"""
+        encoder_hidden_states  (:
+            obj:*torch.FloatTensor* of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is
+            configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (:
+            obj:*tuple(tuple(torch.FloatTensor))* of length `config.n_layers` with each tuple having 4 tensors of shape
+            `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and value
+            hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the
+            user can optionally input only the last `decoder_input_ids` (those that don't have their past key value
+            states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape
+            `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
+            device = input_ids.device
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
+            device = inputs_embeds.device
+        elif encoder_embeds is not None:
+            input_shape = encoder_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
+            device = encoder_embeds.device
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)))
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+            attention_mask, input_shape, device, is_decoder
+        )
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if type(encoder_hidden_states) == list:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
+            else:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+
+            if type(encoder_attention_mask) == list:
+                encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+            else:
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        if encoder_embeds is None:
+            embedding_output = self.embeddings(
+                input_ids=input_ids,
+                position_ids=position_ids,
+                inputs_embeds=inputs_embeds,
+                past_key_values_length=past_key_values_length,
+            )
+        else:
+            embedding_output = encoder_embeds
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+# Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L811
+class BlipTextLMHeadModel(BlipTextPreTrainedModel):
+
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.bert = BlipTextModel(config, add_pooling_layer=False)
+        self.cls = BlipTextOnlyMLMHead(config)
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        labels=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        return_logits=False,
+        is_decoder=True,
+        reduction="mean",
+    ):
+        r"""
+        encoder_hidden_states  (:
+            obj:*torch.FloatTensor* of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is
+            configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
+            ignored (masked), the loss is only computed for the tokens with labels n `[0, ..., config.vocab_size]`
+        past_key_values (:
+            obj:*tuple(tuple(torch.FloatTensor))* of length `config.n_layers` with each tuple having 4 tensors of shape
+            `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and value
+            hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the
+            user can optionally input only the last `decoder_input_ids` (those that don't have their past key value
+            states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape
+            `(batch_size, sequence_length)`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        Returns:
+        Example:
+
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
+
+        outputs = self.bert(
+            input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            is_decoder=is_decoder,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        if return_logits:
+            return prediction_scores[:, :-1, :].contiguous()
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous().to(shifted_prediction_scores.device)
+            loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)
+            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            if reduction == "none":
+                lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_shape)
+
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "past_key_values": past,
+            "encoder_hidden_states": model_kwargs.get("encoder_hidden_states", None),
+            "encoder_attention_mask": model_kwargs.get("encoder_attention_mask", None),
+            "is_decoder": True,
+        }
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past

src/transformers/models/blip/processing_blip.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# 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.
+"""
+Processor class for Blip.
+"""
+
+from typing import List, Optional, Union
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding, PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from ...utils import TensorType
+
+
+class BlipProcessor(ProcessorMixin):
+    r"""
+    Constructs a BLIP processor which wraps a BERT tokenizer and BLIP image processor into a single processor.
+
+    [`BlipProcessor`] offers all the functionalities of [`BlipImageProcessor`] and [`BertTokenizerFast`]. See the
+    docstring of [`~BlipProcessor.__call__`] and [`~BlipProcessor.decode`] for more information.
+
+    Args:
+        image_processor (`BlipImageProcessor`):
+            An instance of [`BlipImageProcessor`]. The image processor is a required input.
+        tokenizer (`BertTokenizerFast`):
+            An instance of ['BertTokenizerFast`]. The tokenizer is a required input.
+    """
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "BlipImageProcessor"
+    tokenizer_class = ("BertTokenizer", "BertTokenizerFast")
+
+    def __init__(self, image_processor, tokenizer):
+        tokenizer.return_token_type_ids = False
+        super().__init__(image_processor, tokenizer)
+        self.current_processor = self.image_processor
+
+    def __call__(
+        self,
+        images=None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        add_special_tokens: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_token_type_ids: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs
+    ) -> BatchEncoding:
+        """
+        This method uses [`BlipImageProcessor.__call__`] method to prepare image(s) for the model, and
+        [`BertTokenizerFast.__call__`] to prepare text for the model.
+
+        Please refer to the docstring of the above two methods for more information.
+        """
+        if images is None and text is None:
+            raise ValueError("You have to specify either images or text.")
+
+        # Get only text
+        if images is None:
+
+            self.current_processor = self.tokenizer
+            text_encoding = self.tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+            return text_encoding
+
+        # add pixel_values
+        encoding_image_processor = self.image_processor(images, return_tensors=return_tensors)
+
+        if text is not None:
+            text_encoding = self.tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+        else:
+            text_encoding = None
+
+        if text_encoding is not None:
+            encoding_image_processor.update(text_encoding)
+
+        return encoding_image_processor
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

src/transformers/utils/dummy_pt_objects.py

@@ -1122,6 +1122,58 @@ class BlenderbotSmallPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+BLIP_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class BlipForConditionalGeneration(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BlipForImageTextRetrieval(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BlipForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BlipModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BlipPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BlipTextModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class BlipVisionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 BLOOM_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

src/transformers/utils/dummy_vision_objects.py

@@ -38,6 +38,13 @@ class BitImageProcessor(metaclass=DummyObject):
         requires_backends(self, ["vision"])
 
 
+class BlipImageProcessor(metaclass=DummyObject):
+    _backends = ["vision"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["vision"])
+
+
 class ChineseCLIPFeatureExtractor(metaclass=DummyObject):
     _backends = ["vision"]
 

tests/models/blip/test_image_processing_blip.py

+# coding=utf-8
+# Copyright 2022 HuggingFace Inc.
+#
+# 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.
+
+
+import unittest
+
+import numpy as np
+
+from transformers.testing_utils import require_torch, require_vision
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_feature_extraction_common import FeatureExtractionSavingTestMixin
+
+
+if is_torch_available():
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import BlipImageProcessor
+
+
+class BlipImageProcessingTester(unittest.TestCase):
+    def __init__(
+        self,
+        parent,
+        batch_size=7,
+        num_channels=3,
+        image_size=18,
+        min_resolution=30,
+        max_resolution=400,
+        do_resize=True,
+        size=None,
+        do_normalize=True,
+        do_pad=False,
+        image_mean=[0.48145466, 0.4578275, 0.40821073],
+        image_std=[0.26862954, 0.26130258, 0.27577711],
+        do_convert_rgb=True,
+    ):
+        size = size if size is not None else {"height": 20, "width": 20}
+        self.parent = parent
+        self.batch_size = batch_size
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.min_resolution = min_resolution
+        self.max_resolution = max_resolution
+        self.do_resize = do_resize
+        self.size = size
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_pad = do_pad
+        self.do_convert_rgb = do_convert_rgb
+
+    def prepare_feat_extract_dict(self):
+        return {
+            "do_resize": self.do_resize,
+            "size": self.size,
+            "do_normalize": self.do_normalize,
+            "image_mean": self.image_mean,
+            "image_std": self.image_std,
+            "do_convert_rgb": self.do_convert_rgb,
+            "do_pad": self.do_pad,
+        }
+
+    def prepare_inputs(self, equal_resolution=False, numpify=False, torchify=False):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        assert not (numpify and torchify), "You cannot specify both numpy and PyTorch tensors at the same time"
+
+        if equal_resolution:
+            image_inputs = []
+            for i in range(self.batch_size):
+                image_inputs.append(
+                    np.random.randint(
+                        255, size=(self.num_channels, self.max_resolution, self.max_resolution), dtype=np.uint8
+                    )
+                )
+        else:
+            image_inputs = []
+            for i in range(self.batch_size):
+                width, height = np.random.choice(np.arange(self.min_resolution, self.max_resolution), 2)
+                image_inputs.append(np.random.randint(255, size=(self.num_channels, width, height), dtype=np.uint8))
+
+        if not numpify and not torchify:
+            # PIL expects the channel dimension as last dimension
+            image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        if torchify:
+            image_inputs = [torch.from_numpy(x) for x in image_inputs]
+
+        return image_inputs
+
+
+@require_torch
+@require_vision
+class BlipImageProcessingTest(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = BlipImageProcessor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = BlipImageProcessingTester(self)
+
+    @property
+    def feat_extract_dict(self):
+        return self.feature_extract_tester.prepare_feat_extract_dict()
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+        self.assertTrue(hasattr(feature_extractor, "do_convert_rgb"))
+
+    def test_batch_feature(self):
+        pass
+
+    def test_call_pil(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PIL images
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+    def test_call_numpy(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random numpy tensors
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False, numpify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, np.ndarray)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+    def test_call_pytorch(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PyTorch tensors
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False, torchify=True)
+        for image in image_inputs:
+            self.assertIsInstance(image, torch.Tensor)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.feature_extract_tester.num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+
+@require_torch
+@require_vision
+class BlipImageProcessingTestFourChannels(FeatureExtractionSavingTestMixin, unittest.TestCase):
+
+    feature_extraction_class = BlipImageProcessor if is_vision_available() else None
+
+    def setUp(self):
+        self.feature_extract_tester = BlipImageProcessingTester(self, num_channels=4)
+        self.expected_encoded_image_num_channels = 3
+
+    @property
+    def feat_extract_dict(self):
+        return self.feature_extract_tester.prepare_feat_extract_dict()
+
+    def test_feat_extract_properties(self):
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        self.assertTrue(hasattr(feature_extractor, "do_resize"))
+        self.assertTrue(hasattr(feature_extractor, "size"))
+        self.assertTrue(hasattr(feature_extractor, "do_normalize"))
+        self.assertTrue(hasattr(feature_extractor, "image_mean"))
+        self.assertTrue(hasattr(feature_extractor, "image_std"))
+        self.assertTrue(hasattr(feature_extractor, "do_convert_rgb"))
+
+    def test_batch_feature(self):
+        pass
+
+    def test_call_pil_four_channels(self):
+        # Initialize feature_extractor
+        feature_extractor = self.feature_extraction_class(**self.feat_extract_dict)
+        # create random PIL images
+        image_inputs = self.feature_extract_tester.prepare_inputs(equal_resolution=False)
+        for image in image_inputs:
+            self.assertIsInstance(image, Image.Image)
+
+        # Test not batched input
+        encoded_images = feature_extractor(image_inputs[0], return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                1,
+                self.expected_encoded_image_num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )
+
+        # Test batched
+        encoded_images = feature_extractor(image_inputs, return_tensors="pt").pixel_values
+        self.assertEqual(
+            encoded_images.shape,
+            (
+                self.feature_extract_tester.batch_size,
+                self.expected_encoded_image_num_channels,
+                self.feature_extract_tester.size["height"],
+                self.feature_extract_tester.size["width"],
+            ),
+        )

tests/models/blip/test_modeling_blip.py

+# coding=utf-8
+# Copyright 2022 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.
+""" Testing suite for the PyTorch Blip model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+
+import numpy as np
+
+import requests
+from transformers import BlipConfig, BlipTextConfig, BlipVisionConfig
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import (
+        BlipForConditionalGeneration,
+        BlipForImageTextRetrieval,
+        BlipForQuestionAnswering,
+        BlipModel,
+        BlipTextModel,
+        BlipVisionModel,
+    )
+    from transformers.models.blip.modeling_blip import BLIP_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import BlipProcessor
+
+
+class BlipVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=1e-10,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+        # in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def get_config(self):
+        return BlipVisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = BlipVisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(pixel_values)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.image_size, self.image_size)
+        patch_size = (self.patch_size, self.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class BlipVisionModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as Blip does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (BlipVisionModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = BlipVisionModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=BlipVisionConfig, has_text_modality=False, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="Blip does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="BlipVisionModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="BlipVisionModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BlipVisionModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class BlipTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        bos_token_id=0,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+        self.bos_token_id = bos_token_id
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return BlipTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+            bos_token_id=self.bos_token_id,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = BlipTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class BlipTextModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (BlipTextModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = BlipTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=BlipTextConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="Blip does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BlipTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class BlipModelTester:
+    def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
+
+        if text_kwargs is None:
+            text_kwargs = {}
+        if vision_kwargs is None:
+            vision_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
+        self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return BlipConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = BlipModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+            "return_loss": True,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class BlipModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (BlipModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+
+    def setUp(self):
+        self.model_tester = BlipModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="BlipModel does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    # override as the `logit_scale` parameter initilization is different for Blip
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    # check if `logit_scale` is initilized as per the original implementation
+                    if name == "logit_scale":
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            np.log(1 / 0.07),
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                pixel_values = inputs_dict["pixel_values"]  # Blip needs pixel_values
+                traced_model = torch.jit.trace(model, (input_ids, pixel_values))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save BlipConfig and check if we can load BlipVisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save BlipConfig and check if we can load BlipTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BlipModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class BlipTextImageModelsModelTester:
+    def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
+
+        if text_kwargs is None:
+            text_kwargs = {}
+        if vision_kwargs is None:
+            vision_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
+        self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return BlipConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = BlipModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class BlipTextImageModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            BlipForConditionalGeneration,
+            BlipForQuestionAnswering,
+            BlipForImageTextRetrieval,
+        )
+        if is_torch_available()
+        else ()
+    )
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = BlipTextImageModelsModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="BlipModel does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            if model.config.is_encoder_decoder:
+                expected_arg_names = [
+                    "input_ids",
+                    "attention_mask",
+                    "decoder_input_ids",
+                    "decoder_attention_mask",
+                ]
+                expected_arg_names.extend(
+                    ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
+                    if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
+                    else ["encoder_outputs"]
+                )
+                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+            else:
+                expected_arg_names = ["input_ids"] if model_class != BlipForConditionalGeneration else ["pixel_values"]
+                self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_training(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes[:-1]:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            loss = model(**inputs).loss
+            loss.backward()
+
+    def test_training_gradient_checkpointing(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes[:-1]:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.use_cache = False
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.gradient_checkpointing_enable()
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+            loss = model(**inputs).loss
+            loss.backward()
+
+    # override as the `logit_scale` parameter initilization is different for Blip
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    # check if `logit_scale` is initilized as per the original implementation
+                    if name == "logit_scale":
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            np.log(1 / 0.07),
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                pixel_values = inputs_dict["pixel_values"]  # Blip needs pixel_values
+                traced_model = torch.jit.trace(model, (input_ids, pixel_values))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save BlipConfig and check if we can load BlipVisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save BlipConfig and check if we can load BlipTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BlipModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "https://storage.googleapis.com/sfr-vision-language-research/BLIP/demo.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@require_vision
+@require_torch
+@slow
+class BlipModelIntegrationTest(unittest.TestCase):
+    def test_inference_image_captioning(self):
+        model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(torch_device)
+        processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+        image = prepare_img()
+
+        # image only
+        inputs = processor(images=image, return_tensors="pt").to(torch_device)
+
+        predictions = model.generate(**inputs)
+
+        # Test output
+        self.assertEqual(predictions[0].tolist(), [30522, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102])
+
+        # image and context
+        context = ["a picture of"]
+        inputs = processor(images=image, text=context, return_tensors="pt").to(torch_device)
+
+        predictions = model.generate(**inputs)
+
+        # Test output
+        self.assertEqual(
+            predictions[0].tolist(),
+            [30522, 1037, 3861, 1997, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102],
+        )
+
+    def test_inference_image_captioning_fp16(self):
+        model = BlipForConditionalGeneration.from_pretrained(
+            "Salesforce/blip-image-captioning-base", torch_dtype=torch.float16
+        ).to(torch_device)
+        processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
+        image = prepare_img()
+
+        # image only
+        inputs = processor(images=image, return_tensors="pt").to(torch_device, torch.float16)
+
+        predictions = model.generate(**inputs)
+
+        # Test output
+        self.assertEqual(predictions[0].tolist(), [30522, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102])
+
+        # image and context
+        context = ["a picture of"]
+        inputs = processor(images=image, text=context, return_tensors="pt").to(torch_device, torch.float16)
+
+        predictions = model.generate(**inputs)
+
+        # Test output
+        self.assertEqual(
+            predictions[0].tolist(),
+            [30522, 1037, 3861, 1997, 1037, 2450, 3564, 2006, 1996, 3509, 2007, 2014, 3899, 102],
+        )
+
+    def test_inference_vqa(self):
+        model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base").to(torch_device)
+        processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
+
+        image = prepare_img()
+        text = "how many dogs are in the picture?"
+
+        inputs = processor(image, text=text, return_tensors="pt").to(torch_device)
+        out = model.generate(**inputs)
+
+        # Test output
+        self.assertEqual(out[0].tolist(), [30522, 1015, 102])
+
+    def test_inference_itm(self):
+        model = BlipForImageTextRetrieval.from_pretrained("Salesforce/blip-itm-base-coco").to(torch_device)
+        processor = BlipProcessor.from_pretrained("Salesforce/blip-itm-base-coco")
+
+        image = prepare_img()
+        text = "A woman and her dog sitting in a beach"
+
+        inputs = processor(image, text, return_tensors="pt").to(torch_device)
+
+        out_itm = model(**inputs)
+        out = model(**inputs, use_itm_head=False)
+
+        expected_scores = torch.Tensor([[0.9779, 0.0221]])
+
+        self.assertTrue(torch.allclose(torch.nn.Softmax()(out_itm[0].cpu()), expected_scores, atol=1e-3, rtol=1e-3))
+        self.assertTrue(torch.allclose(out[0].cpu(), torch.Tensor([[0.5053]]), atol=1e-3, rtol=1e-3))

tests/models/blip/test_modeling_blip_text.py

+# coding=utf-8
+# Copyright 2022 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.
+""" Testing suite for the PyTorch Blip model. """
+import unittest
+
+import numpy as np
+
+from transformers import BlipTextConfig
+from transformers.testing_utils import require_torch, slow, torch_device
+from transformers.utils import is_torch_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import BlipTextModel
+    from transformers.models.blip.modeling_blip import BLIP_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+class BlipTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        bos_token_id=0,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+        self.bos_token_id = bos_token_id
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return BlipTextConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+            bos_token_id=self.bos_token_id,
+        )
+
+    def create_and_check_model(self, config, input_ids, input_mask):
+        model = BlipTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(input_ids, attention_mask=input_mask)
+            result = model(input_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, input_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class BlipTextModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (BlipTextModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = BlipTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=BlipTextConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="Blip does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="BlipTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BlipTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)

tests/models/blip/test_processor_blip.py

+# Copyright 2022 The HuggingFace 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.
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers.testing_utils import require_vision
+from transformers.utils import is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoProcessor, BertTokenizer, BlipImageProcessor, BlipProcessor, PreTrainedTokenizerFast
+
+
+@require_vision
+class BlipProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = BlipImageProcessor()
+        tokenizer = BertTokenizer.from_pretrained("hf-internal-testing/tiny-random-BertModel")
+
+        processor = BlipProcessor(image_processor, tokenizer)
+
+        processor.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = BlipProcessor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
+
+        processor = BlipProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, PreTrainedTokenizerFast)
+
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.image_processor, BlipImageProcessor)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        image_input = self.prepare_image_inputs()
+
+        input_feat_extract = image_processor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str, return_token_type_ids=False)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["pixel_values", "input_ids", "attention_mask"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)
+
+    def test_model_input_names(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = BlipProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        # For now the processor supports only ['pixel_values', 'input_ids', 'attention_mask']
+        self.assertListEqual(list(inputs.keys()), ["pixel_values", "input_ids", "attention_mask"])

utils/check_repo.py

@@ -121,6 +121,7 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     "FlaxBertForCausalLM",  # Building part of bigger (tested) model. Tested implicitly through FlaxRobertaForCausalLM.
     "OPTDecoderWrapper",
     "TFSegformerDecodeHead",  # Not a regular model.
+    "BlipTextLMHeadModel",  # No need to test it as it is tested by BlipTextVision models
 ]
 
 # Update this list with test files that don't have a tester with a `all_model_classes` variable and which don't
@@ -147,6 +148,12 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     # models to ignore for model xxx mapping
+    "BlipForConditionalGeneration",
+    "BlipForImageTextRetrieval",
+    "BlipForQuestionAnswering",
+    "BlipVisionModel",
+    "BlipTextLMHeadModel",
+    "BlipTextModel",
     "Swin2SRForImageSuperResolution",
     "CLIPSegForImageSegmentation",
     "CLIPSegVisionModel",

utils/documentation_tests.txt

@@ -35,6 +35,7 @@ src/transformers/models/blenderbot/configuration_blenderbot.py
 src/transformers/models/blenderbot/modeling_blenderbot.py
 src/transformers/models/blenderbot_small/configuration_blenderbot_small.py
 src/transformers/models/blenderbot_small/modeling_blenderbot_small.py
+src/transformers/models/blip/modeling_blip.py
 src/transformers/models/bloom/configuration_bloom.py
 src/transformers/models/camembert/configuration_camembert.py
 src/transformers/models/canine/configuration_canine.py