fengshen/models/clip/modeling_taiyi_clip.py

import torch
from torch import nn
from transformers.models.clip.modeling_clip import (
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    CLIP_START_DOCSTRING,
    CLIP_TEXT_INPUTS_DOCSTRING,
    CLIP_VISION_INPUTS_DOCSTRING,
    CLIP_INPUTS_DOCSTRING,
    replace_return_docstrings,
    CLIPVisionConfig,
    CLIPPreTrainedModel,
    CLIPVisionTransformer,
    CLIPOutput,
    CLIPConfig,
    clip_loss,
)
from typing import Optional, Tuple, Union
# from transformers import MegatronBertConfig as BertConfig
# from transformers import MegatronBertModel as BertModel
from transformers.models.bert.modeling_bert import BertModel
from transformers.models.bert.configuration_bert import BertConfig
from .configuration_taiyi_clip import TaiyiCLIPConfig


@add_start_docstrings(CLIP_START_DOCSTRING)
class TaiyiCLIPModel(CLIPPreTrainedModel):
    config_class = TaiyiCLIPConfig

    def __init__(self, config: TaiyiCLIPConfig):
        super().__init__(config)

        if not isinstance(config.text_config, BertConfig):
            raise ValueError(
                "config.text_config is expected to be of type CLIPTextConfig but is of type"
                f" {type(config.text_config)}."
            )

        if not isinstance(config.vision_config, CLIPVisionConfig):
            raise ValueError(
                "config.vision_config is expected to be of type CLIPVisionConfig 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 = BertModel(text_config)
        self.vision_model = CLIPVisionTransformer(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(CLIP_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,
        token_type_ids: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = 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 [`CLIPTextModel`].

        Examples:

        ```python
        >>> from transformers import CLIPTokenizer, CLIPModel

        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        >>> tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")

        >>> inputs = tokenizer(["a photo of a cat", "a photo of a dog"], padding=True, return_tensors="pt")
        >>> text_features = model.get_text_features(**inputs)
        ```"""
        # Use CLIP 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

        text_outputs = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            token_type_ids=token_type_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        # pooled_output = text_outputs[1]
        pooled_output = text_outputs[0][:, 0, :]
        text_features = self.text_projection(pooled_output)

        return text_features

    @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING)
    def get_image_features(
        self,
        pixel_values: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = 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 [`CLIPVisionModel`].

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import CLIPProcessor, CLIPModel

        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        >>> processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")

        >>> 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)
        ```"""
        # Use CLIP 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,
        )

        pooled_output = vision_outputs[1]  # pooled_output
        image_features = self.visual_projection(pooled_output)

        return image_features

    @add_start_docstrings_to_model_forward(CLIP_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=CLIPOutput, config_class=CLIPConfig)
    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, CLIPOutput]:
        r"""
        Returns:

        Examples:

        ```python
        >>> from PIL import Image
        >>> import requests
        >>> from transformers import CLIPProcessor, CLIPModel

        >>> model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
        >>> processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")

        >>> 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 CLIP 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 = clip_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 CLIPOutput(
            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,
        )