modeling.py

import torch
from torch import nn
import numpy as np
from typing import Optional, Tuple, List, Union
from transformers import Qwen2VLForConditionalGeneration
import logging
import warnings
from PIL import Image
from transformers.image_utils import load_image

logger = logging.getLogger(__name__)

LOGIT_BIAS = 2.65  # logit bias for sigmoid normalization

def load_images(images, lazy_load: bool = True):
    # Disable PIL DecompositionBomb threshold for reading large images.
    pil_max_px = Image.MAX_IMAGE_PIXELS
    Image.MAX_IMAGE_PIXELS = None

    images_batch = []
    for image in images:
        if isinstance(image, Image.Image):
            images_batch.append(image)
        else:
            pil_image = load_image(image)
            if lazy_load:
                images_batch.append(pil_image)
            else:
                # avoid Too many open files error
                images_batch.append(pil_image.copy())
                pil_image.close()
    Image.MAX_IMAGE_PIXELS = pil_max_px

    return images_batch


def formatting_prompts_func(
    query: str,
    doc: str,
    query_type: str = 'text',
    doc_type: str = 'text',
    prefix_str: str = '',
) -> str:
    """
    Format prompts for different combinations of query and content types.

    Args:
        query: Query text or image path
        doc: Content text or image path
        query_type: Whether query is an image
        doc_type: Whether content is an image
        prefix_str: Optional prefix string to add
    """
    # Format query part
    if query_type == 'image':
        query_part = "**Query**:\n<|vision_start|><|image_pad|><|vision_end|>"
    else:
        query_part = f"**Query**:\n{query}"

    # Format content part
    if doc_type == 'image':
        doc_part = "**Document**:\n<|vision_start|><|image_pad|><|vision_end|>"
    else:
        doc_part = f"**Document**:\n{doc}"

    # Combine parts
    prompt = doc_part + '\n' + query_part

    # Add prefix if provided
    if prefix_str:
        prompt = prefix_str + '\n' + prompt

    return prompt


class JinaVLForRanking(Qwen2VLForConditionalGeneration):
    def __init__(self, config):
        super().__init__(config)

        self.padding_side = "left"
        self.num_labels = 1  # config.num_labels

        # hack the lm_head to do nothing, since we only want the hidden states
        self.lm_head = nn.Identity()

        # copy the idea from `Qwen2ForRewardModel` to have a MLP layer to get the final score
        self.score = nn.Sequential(
            nn.Linear(config.hidden_size, config.hidden_size),
            nn.ReLU(),
            nn.Linear(config.hidden_size, self.num_labels),
        )

        # Initialize weights and apply final processing
        self.post_init()

        self.score_token_id = 100

    def forward(self, *args, **kwargs) -> torch.Tensor:
        # Delete output_hidden_states from kwargs
        kwargs.pop("output_hidden_states", None)
        kwargs.pop("use_cache", None)
        assert kwargs.pop("labels", None) is None, "labels should not be passed to forward()"

        outputs = super().forward(
            *args,
            use_cache=False,
            output_hidden_states=True,
            **kwargs,
        )

        # get the hidden states of the last layer
        hidden_states = outputs.hidden_states[-1]

        # IMPORTANT: the padding token must be on the left side
        # get the hidden states of the last token and apply the linear layer
        pooled_logits = self.score(hidden_states[:, -1])

        return pooled_logits.squeeze(-1)

    @torch.no_grad()
    def compute_score(
        self,
        pairs: Union[List[Tuple[str, str]], Tuple[str, str]],
        batch_size: int = 8,
        max_length: int = 10240,
        max_query_length: int = 512,
        max_doc_length: Optional[int] = None,
        query_type: str = 'text',
        doc_type: str = 'text',
        normalize_scores: bool = True,
        show_progress: bool = False,
    ) -> List[float]:

        if not hasattr(self, "_processor"):
            from transformers import AutoProcessor

            self._processor = AutoProcessor.from_pretrained(
                self.name_or_path, max_pixels=602112, min_pixels=3136, trust_remote_code=True
            )

        assert isinstance(pairs, list)

        if isinstance(pairs[0], str):
            pairs = [pairs]

        max_length = max_length or self.config.max_length

        if max_doc_length is None:
            max_doc_length = max(max_length - max_query_length, max_query_length)

        if max_doc_length < max_query_length:
            warnings.warn(
                f"max_doc_length={max_doc_length} should be greater than max_query_length={max_query_length}"
            )

        assert (
            max_doc_length + max_query_length <= max_length
        ), f"max_doc_length ({max_doc_length}) + max_query_length ({max_query_length}) should be less than max_length ({max_length})"

        max_length = max_length - 1

        all_scores = []

        device = next(self.parameters()).device

        batch_iter = range(0, len(pairs), batch_size)
        if show_progress:
            from tqdm import trange

            batch_iter = trange(0, len(pairs), batch_size, desc="Computing scores")

        for start_index in batch_iter:
            mini_batch = pairs[start_index : start_index + batch_size]

            batch_inputs = []
            for q, d in mini_batch:
                # TEMP FIX: Truncate long documents
                if doc_type == 'text':
                    tokens = self._processor.tokenizer(d, truncation=True, max_length=max_doc_length)
                    if len(tokens['input_ids']) >= max_doc_length:
                        d = self._processor.tokenizer.decode(tokens['input_ids'])

                batch_inputs.append(formatting_prompts_func(q, d, query_type=query_type, doc_type=doc_type))

            batch_images = None
            # if doc_type == 'image':
            #     batch_images = load_images([d for (q, d) in mini_batch])
            # elif query_type == 'image':
            #     batch_images = load_images([q for (q, d) in mini_batch])

            doc_images = []
            query_images = []
            if doc_type == 'image':
                doc_images = load_images([d for (q, d) in mini_batch])
            if query_type == 'image':
                query_images = load_images([q for (q, d) in mini_batch])

            if len(doc_images) == len(query_images) and len(doc_images) > 0:
                batch_images = [[d, q] for q, d in zip(query_images, doc_images)]
            elif len(doc_images) > 0:
                batch_images = doc_images
            elif len(query_images) > 0:
                batch_images = query_images

            batch = self._processor(
                text=batch_inputs,
                images=batch_images,
                return_tensors="pt",
                padding=True,
                truncation=True,
                max_length=max_length,
            )

            # append the reward token to the input_ids and attention_mask
            batch_size = batch["input_ids"].size(0)
            batch["input_ids"] = torch.cat(
                [
                    batch["input_ids"],
                    torch.full((batch_size, 1), self.score_token_id, device=batch["input_ids"].device),
                ],
                dim=1,
            )
            batch["attention_mask"] = torch.cat(
                [
                    batch["attention_mask"],
                    torch.ones((batch_size, 1), device=batch["attention_mask"].device),
                ],
                dim=1,
            )
            # move the batch to the correct device
            batch = {k: v.to(device) if isinstance(v, torch.Tensor) else v for k, v in batch.items()}

            scores = self.forward(**batch).view(-1).cpu().float().numpy()

            # normalize scores to [0, 1] with sigmoid with a scale
            scores = 1.0 / (1.0 + np.exp(-(scores - LOGIT_BIAS)))

            all_scores.extend(scores.tolist())

        if len(all_scores) == 1:
            return all_scores[0]
        return all_scores