modeling_mplug_owl2_huggingface.py

#    Copyright 2023 Haotian Liu & Qinghao Ye (Modified from LLaVA)
#
#    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 os
import sys
from abc import ABC, abstractmethod
from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn import CrossEntropyLoss
from PIL import Image

dir_path = os.path.dirname(os.path.realpath(__file__))
sys.path.insert(0, dir_path)

from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    LlamaForCausalLM,
    LlamaModel,
)
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.models.clip.image_processing_clip import CLIPImageProcessor

from .configuration_mplug_owl2 import (
    MPLUGOwl2Config,
    MplugOwlVisionConfig,
    MplugOwlVisualAbstractorConfig
)
from .modeling_attn_mask_utils import _prepare_4d_causal_attention_mask  # Force copy
from .modeling_llama2 import replace_llama_modality_adaptive
from .visual_encoder import MplugOwlVisionModel, MplugOwlVisualAbstractorModel

IGNORE_INDEX = -100
IMAGE_TOKEN_INDEX = -200
DEFAULT_IMAGE_TOKEN = "<|image|>"


def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split(DEFAULT_IMAGE_TOKEN)]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids


def expand2square(pil_img, background_color):
    width, height = pil_img.size
    if width == height:
        return pil_img
    elif width > height:
        result = Image.new(pil_img.mode, (width, width), background_color)
        result.paste(pil_img, (0, (width - height) // 2))
        return result
    else:
        result = Image.new(pil_img.mode, (height, height), background_color)
        result.paste(pil_img, ((height - width) // 2, 0))
        return result


class MPLUGOwl2MetaModel:
    def __init__(self, config):
        super(MPLUGOwl2MetaModel, self).__init__(config)
        self.vision_model = MplugOwlVisionModel(
            MplugOwlVisionConfig(**config.visual_config["visual_model"])
        )
        self.visual_abstractor = MplugOwlVisualAbstractorModel(
            MplugOwlVisualAbstractorConfig(**config.visual_config["visual_abstractor"]),
            config.hidden_size,
        )

    def get_vision_tower(self):
        vision_model = getattr(self, "vision_model", None)
        if type(vision_model) is list:
            vision_model = vision_model[0]
        return vision_model

    def get_visual_abstractor(self):
        visual_abstractor = getattr(self, "visual_abstractor", None)
        if type(visual_abstractor) is list:
            visual_abstractor = visual_abstractor[0]
        return visual_abstractor


class MPLUGOwl2MetaForCausalLM(ABC):
    @abstractmethod
    def get_model(self):
        pass

    def encode_images(self, images):
        image_features = self.get_model().vision_model(images).last_hidden_state
        image_features = (
            self.get_model()
            .visual_abstractor(encoder_hidden_states=image_features)
            .last_hidden_state
        )
        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, attention_mask, past_key_values, labels, images
    ):
        if images is None or input_ids.shape[1] == 1:
            if (
                past_key_values is not None
                and images is not None
                and input_ids.shape[1] == 1
            ):
                attention_mask = torch.ones(
                    (attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1),
                    dtype=attention_mask.dtype,
                    device=attention_mask.device,
                )
            multiway_indices = torch.zeros_like(input_ids).long().to(self.device)
            return (
                input_ids,
                multiway_indices,
                attention_mask,
                past_key_values,
                None,
                labels,
            )

        if type(images) is list or images.ndim == 5:
            concat_images = torch.cat([image for image in images], dim=0)
            image_features = self.encode_images(concat_images)
            split_sizes = [image.shape[0] for image in images]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1) for x in image_features]
        else:
            image_features = self.encode_images(images)

        new_input_embeds = []
        new_modality_indicators = []
        new_labels = [] if labels is not None else None
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
                half_len = cur_input_ids.shape[0] // 2
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(
                    cur_input_ids[:half_len]
                )
                cur_input_embeds_2 = self.get_model().embed_tokens(
                    cur_input_ids[half_len:]
                )
                cur_input_embeds = torch.cat(
                    [cur_input_embeds_1, cur_image_features[0:0], cur_input_embeds_2],
                    dim=0,
                )
                new_input_embeds.append(cur_input_embeds)

                cur_modality_indicators = (
                    torch.zeros(len(cur_input_embeds)).long().to(self.device)
                )
                new_modality_indicators.append(cur_modality_indicators)
                if labels is not None:
                    new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue
            image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            cur_new_input_embeds = []
            cur_modality_indicators = []
            if labels is not None:
                cur_labels = labels[batch_idx]
                cur_new_labels = []
                assert cur_labels.shape == cur_input_ids.shape
            while image_token_indices.numel() > 0:
                cur_image_features = image_features[cur_image_idx]
                image_token_start = image_token_indices[0]
                cur_new_input_embeds.append(
                    self.get_model().embed_tokens(cur_input_ids[:image_token_start])
                )
                cur_new_input_embeds.append(cur_image_features)

                cur_modality_indicators.append(
                    torch.zeros(len(cur_input_ids[:image_token_start])).long()
                )
                cur_modality_indicators.append(
                    torch.ones(len(cur_image_features)).long()
                )

                if labels is not None:
                    cur_new_labels.append(cur_labels[:image_token_start])
                    cur_new_labels.append(
                        torch.full(
                            (cur_image_features.shape[0],),
                            IGNORE_INDEX,
                            device=labels.device,
                            dtype=labels.dtype,
                        )
                    )
                    cur_labels = cur_labels[image_token_start + 1 :]
                cur_image_idx += 1
                cur_input_ids = cur_input_ids[image_token_start + 1 :]
                image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            if cur_input_ids.numel() > 0:
                cur_new_input_embeds.append(
                    self.get_model().embed_tokens(cur_input_ids)
                )
                cur_modality_indicators.append(torch.zeros(len(cur_input_ids)).long())
                if labels is not None:
                    cur_new_labels.append(cur_labels)
            cur_new_input_embeds = [
                x.to(device=self.device) for x in cur_new_input_embeds
            ]
            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
            new_input_embeds.append(cur_new_input_embeds)

            cur_modality_indicators = [
                x.to(device=self.device) for x in cur_modality_indicators
            ]
            cur_modality_indicators = torch.cat(cur_modality_indicators, dim=0)
            new_modality_indicators.append(cur_modality_indicators)

            if labels is not None:
                cur_new_labels = torch.cat(cur_new_labels, dim=0)
                new_labels.append(cur_new_labels)

        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
            max_len = max(x.shape[0] for x in new_input_embeds)

            new_input_embeds_align = []
            for cur_new_embed in new_input_embeds:
                cur_new_embed = torch.cat(
                    (
                        cur_new_embed,
                        torch.zeros(
                            (max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]),
                            dtype=cur_new_embed.dtype,
                            device=cur_new_embed.device,
                        ),
                    ),
                    dim=0,
                )
                new_input_embeds_align.append(cur_new_embed)
            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)

            new_modality_indicators_align = []
            for cur_modality_indicator in new_modality_indicators:
                cur_new_embed = torch.cat(
                    (
                        cur_modality_indicator,
                        torch.zeros(
                            max_len - cur_modality_indicator.shape[0],
                            dtype=cur_modality_indicator.dtype,
                            device=cur_modality_indicator.device,
                        ),
                    ),
                    dim=0,
                )
                new_modality_indicators_align.append(cur_new_embed)
            new_modality_indicators = torch.stack(new_modality_indicators_align, dim=0)

            if labels is not None:
                new_labels_align = []
                _new_labels = new_labels
                for cur_new_label in new_labels:
                    cur_new_label = torch.cat(
                        (
                            cur_new_label,
                            torch.full(
                                (max_len - cur_new_label.shape[0],),
                                IGNORE_INDEX,
                                dtype=cur_new_label.dtype,
                                device=cur_new_label.device,
                            ),
                        ),
                        dim=0,
                    )
                    new_labels_align.append(cur_new_label)
                new_labels = torch.stack(new_labels_align, dim=0)

            if attention_mask is not None:
                new_attention_mask = []
                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(
                    attention_mask, _new_labels, new_labels
                ):
                    new_attn_mask_pad_left = torch.full(
                        (cur_new_labels.shape[0] - labels.shape[1],),
                        True,
                        dtype=attention_mask.dtype,
                        device=attention_mask.device,
                    )
                    new_attn_mask_pad_right = torch.full(
                        (cur_new_labels_align.shape[0] - cur_new_labels.shape[0],),
                        False,
                        dtype=attention_mask.dtype,
                        device=attention_mask.device,
                    )
                    cur_new_attention_mask = torch.cat(
                        (
                            new_attn_mask_pad_left,
                            cur_attention_mask,
                            new_attn_mask_pad_right,
                        ),
                        dim=0,
                    )
                    new_attention_mask.append(cur_new_attention_mask)
                attention_mask = torch.stack(new_attention_mask, dim=0)
                assert attention_mask.shape == new_labels.shape
        else:
            new_input_embeds = torch.stack(new_input_embeds, dim=0)
            new_modality_indicators = torch.stack(new_modality_indicators, dim=0)
            if labels is not None:
                new_labels = torch.stack(new_labels, dim=0)

            if attention_mask is not None:
                new_attn_mask_pad_left = torch.full(
                    (
                        attention_mask.shape[0],
                        new_input_embeds.shape[1] - input_ids.shape[1],
                    ),
                    True,
                    dtype=attention_mask.dtype,
                    device=attention_mask.device,
                )
                attention_mask = torch.cat(
                    (new_attn_mask_pad_left, attention_mask), dim=1
                )
                assert attention_mask.shape == new_input_embeds.shape[:2]
        return (
            None,
            new_modality_indicators,
            attention_mask,
            past_key_values,
            new_input_embeds,
            new_labels,
        )


class MPLUGOwl2LlamaModel(MPLUGOwl2MetaModel, LlamaModel):
    config_class = MPLUGOwl2Config

    def __init__(self, config: MPLUGOwl2Config):
        super(MPLUGOwl2LlamaModel, self).__init__(config)


class MPLUGOwl2LlamaForCausalLM(LlamaForCausalLM, MPLUGOwl2MetaForCausalLM):
    config_class = MPLUGOwl2Config

    def __init__(self, config):
        super(LlamaForCausalLM, self).__init__(config)
        self.model = MPLUGOwl2LlamaModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

        # Tokenizer and image processor will be initialized lazily in score()
        self._tokenizer = None
        self._image_processor = None
        self._preferential_ids = None

        self.post_init()

    def _init_processors(self):
        """Lazily initialize tokenizer and image processor from the model's directory."""
        if self._tokenizer is None:
            # Use the model's name_or_path from config, fallback to HF repo name
            model_path = getattr(self.config, '_name_or_path', None)
            if model_path is None or model_path == './' or not model_path.startswith(('/', 'http', 'mapo80')):
                model_path = "mapo80/DeQA-Doc-Color"
            self._tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
            self._image_processor = CLIPImageProcessor.from_pretrained(model_path)
            self._preferential_ids = [id_[1] for id_ in self._tokenizer(
                ["excellent", "good", "fair", "poor", "bad"]
            )["input_ids"]]

    @property
    def tokenizer(self):
        self._init_processors()
        return self._tokenizer

    @property
    def image_processor(self):
        self._init_processors()
        return self._image_processor

    @property
    def preferential_ids_(self):
        self._init_processors()
        return self._preferential_ids

    def get_model(self):
        return self.model

    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        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
        )
        (
            input_ids,
            modality_indicators,
            attention_mask,
            past_key_values,
            inputs_embeds,
            labels,
        ) = self.prepare_inputs_labels_for_multimodal(
            input_ids, attention_mask, past_key_values, labels, images
        )

        outputs = self.model(
            input_ids=input_ids,
            modality_indicators=modality_indicators,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)

        loss = None
        if labels is not None:
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def score(
        self,
        images: List[Image.Image],
        task_: str = "quality",
        input_: str = "image",
    ) -> torch.Tensor:
        """
        Score images based on quality assessment.

        Args:
            images: List of PIL Images to score
            task_: Type of assessment (default: "quality")
            input_: Input type - "image" or "video" (default: "image")

        Returns:
            torch.Tensor: Quality scores (1-5 scale)
        """
        if not hasattr(self, "weight_tensor"):
            self.weight_tensor = torch.Tensor([5., 4., 3., 2., 1.]).half().to(self.device)

        prompt = "USER: How would you rate the {} of this {}?\n<|image|>\nASSISTANT: The {} of the {} is".format(
            task_, input_, task_, input_
        )

        if input_ == "image":
            # Process single images
            images = [expand2square(img, tuple(int(x*255) for x in self.image_processor.image_mean)) for img in images]
            input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)

            with torch.inference_mode():
                image_tensor = self.image_processor.preprocess(images, return_tensors="pt")["pixel_values"].half().to(self.device)
                output_logits = self(
                    input_ids=input_ids.repeat(image_tensor.shape[0], 1),
                    images=image_tensor
                )["logits"][:, -1, self.preferential_ids_]

                return torch.softmax(output_logits, -1) @ self.weight_tensor
        else:
            # Process videos (list of frame sequences)
            video = [[expand2square(frame, tuple(int(x*255) for x in self.image_processor.image_mean)) for frame in vid] for vid in images]
            input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)

            with torch.inference_mode():
                video_tensors = [
                    self.image_processor.preprocess(vid, return_tensors="pt")["pixel_values"].half().to(self.device)
                    for vid in video
                ]
                output_logits = self(
                    input_ids=input_ids.repeat(len(video_tensors), 1),
                    images=video_tensors
                )["logits"][:, -1, self.preferential_ids_]

                return torch.softmax(output_logits, -1) @ self.weight_tensor

    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        inputs_embeds=None,
        images=None,
        **kwargs,
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "images": images,
            }
        )
        return model_inputs


AutoConfig.register("mplug_owl2", MPLUGOwl2Config)
AutoModelForCausalLM.register(MPLUGOwl2Config, MPLUGOwl2LlamaForCausalLM)

replace_llama_modality_adaptive()