processing_spec_vision.py

# coding=utf-8
# Copyright 2024 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.

"""
Processor class for Spec-Vision.
"""

import re
from typing import List, Optional, Union

import numpy as np
import torch
import torchvision
from PIL import Image
from transformers import AutoImageProcessor
from transformers.feature_extraction_utils import BatchFeature
from transformers.image_processing_utils import BaseImageProcessor
from transformers.image_transforms import convert_to_rgb
from transformers.image_utils import (OPENAI_CLIP_MEAN, OPENAI_CLIP_STD,
                                      ImageInput, make_list_of_images,
                                      valid_images)
from transformers.processing_utils import ProcessorMixin
from transformers.tokenization_utils_base import (PaddingStrategy, TextInput,
                                                  TruncationStrategy)
from transformers.utils import TensorType, is_vision_available, logging

logger = logging.get_logger(__name__)

def padding_336(image):
    """Apply padding to make height a multiple of 336 while preserving aspect ratio."""
    width, height = image.size
    target_height = int(np.ceil(height / 336) * 336)
    top_padding = int((target_height - height) / 2)
    bottom_padding = target_height - height - top_padding
    padded_image = torchvision.transforms.functional.pad(
        image, 
        [0, top_padding, 0, bottom_padding],
        fill=[255, 255, 255]
    )
    return padded_image

def calc_padded_size(width, height, padding_unit=336):
    """Calculate the padded dimensions for an image."""
    target_height = int(np.ceil(height / padding_unit) * padding_unit)
    padded_width = width
    padded_height = target_height
    return padded_width, padded_height

def hd_transform(img, hd_num=16):
    """Apply HD transformation with support for Spec-Vision's requirements."""
    width, height = img.size
    transposed = False
    
    # Handle portrait images by transposing
    if width < height:
        img = img.transpose(Image.TRANSPOSE)
        width, height = img.size
        transposed = True
        
    ratio = width / height
    scale = 1
    while scale * np.ceil(scale / ratio) <= hd_num:
        scale += 1
    scale -= 1
    
    new_width = int(scale * 336)
    new_height = int(new_width / ratio)
    
    # Resize and pad
    img = torchvision.transforms.functional.resize(img, [new_height, new_width])
    img = padding_336(img)
    
    # Restore original orientation if needed
    if transposed:
        img = img.transpose(Image.TRANSPOSE)
        
    return img

def pad_to_max_crops(images, max_crops=5):
    """Pad batch of images to have consistent number of crops."""
    B, _, H, W = images.shape
    if B < max_crops:
        padding = torch.zeros(max_crops - B, 3, H, W, dtype=images.dtype, device=images.device)
        images = torch.cat([images, padding], dim=0)
    return images

class SpecVisionImageProcessor(BaseImageProcessor):
    """
    Image processor for Spec-Vision model.
    
    This processor handles the preparation of images for the Spec-Vision model, including:
    - HD transformation for high-resolution image processing
    - Multi-crop processing with configurable number of crops
    - Normalization and padding
    """

    model_input_names = ["pixel_values"]

    def __init__(
        self,
        num_crops: int = 1,
        image_mean: Optional[Union[float, List[float]]] = None,
        image_std: Optional[Union[float, List[float]]] = None,
        do_convert_rgb: bool = True,
        hd_transform_order: str = "sub_glb",
        **kwargs,
    ) -> None:
        super().__init__(**kwargs)
        self.num_crops = num_crops
        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
        self.do_convert_rgb = do_convert_rgb
        self.hd_transform_order = hd_transform_order

    def calc_num_image_tokens(self, images: ImageInput) -> List[int]:
        """Calculate number of image tokens needed for each image."""
        images = make_list_of_images(images)
        if not valid_images(images):
            raise ValueError("Invalid image type provided")

        images = [image.convert('RGB') for image in images]
        transformed_images = [hd_transform(im, hd_num=self.num_crops) for im in images]
        shapes = [[im.size[1], im.size[0]] for im in transformed_images]
        
        # Calculate tokens based on Spec-Vision's architecture
        num_img_tokens = [
            int((h//336 * w//336 + 1) * 144 + 1 + (h//336 + 1) * 12)
            for h, w in shapes
        ]
        return num_img_tokens

    def preprocess(
        self,
        images: ImageInput,
        image_mean: Optional[Union[float, List[float]]] = None,
        image_std: Optional[Union[float, List[float]]] = None,
        do_convert_rgb: bool = None,
        return_tensors: Optional[Union[str, TensorType]] = None,
    ) -> BatchFeature:
        """
        Preprocess images for the Spec-Vision model.
        
        Handles HD transformation, normalization, and proper formatting of images
        according to Spec-Vision's requirements.
        """
        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

        # Validate and prepare images
        images = make_list_of_images(images)
        if not valid_images(images):
            raise ValueError("Invalid image type provided")

        if do_convert_rgb:
            images = [convert_to_rgb(image) for image in images]

        # Create image processor pipeline
        img_processor = torchvision.transforms.Compose([
            torchvision.transforms.ToTensor(),
            torchvision.transforms.Normalize(image_mean, image_std)
        ])

        # Process images according to Spec-Vision's HD transform requirements
        images = [image.convert('RGB') for image in images]
        transformed_images = [hd_transform(im, hd_num=self.num_crops) for im in images]
        
        # Convert to tensors and normalize
        hd_images = [img_processor(im) for im in transformed_images]
        
        # Create global views
        global_images = [
            torch.nn.functional.interpolate(
                im.unsqueeze(0).float(),
                size=(336, 336),
                mode='bicubic'
            ).to(im.dtype)
            for im in hd_images
        ]

        # Process shapes and calculate tokens
        shapes = [[im.size(1), im.size(2)] for im in hd_images]
        num_img_tokens = [
            int(((h//336) * (w//336) + 1) * 144 + 1 + (h//336 + 1) * 12)
            for h, w in shapes
        ]

        # Reshape images according to Spec-Vision's requirements
        hd_images_reshaped = [
            im.reshape(1, 3, h//336, 336, w//336, 336)
              .permute(0, 2, 4, 1, 3, 5)
              .reshape(-1, 3, 336, 336)
              .contiguous()
            for im, (h, w) in zip(hd_images, shapes)
        ]

        # Combine global and local views based on transform order
        if self.hd_transform_order == "sub_glb":
            processed_images = [
                torch.cat([_im, _global_image], dim=0)
                for _global_image, _im in zip(global_images, hd_images_reshaped)
            ]
        else:  # glb_sub
            processed_images = [
                torch.cat([_global_image, _im], dim=0)
                for _global_image, _im in zip(global_images, hd_images_reshaped)
            ]

        # Pad to consistent number of crops
        image_batch = [
            pad_to_max_crops(im, self.num_crops + 1)
            for im in processed_images
        ]
        image_batch = torch.stack(image_batch, dim=0)

        return BatchFeature(
            data={
                "pixel_values": image_batch,
                "image_sizes": shapes,
                "num_img_tokens": num_img_tokens
            },
            tensor_type=return_tensors
        )

class SpecVisionProcessor(ProcessorMixin):
    """
    Combined processor for Spec-Vision model, handling both image and text inputs.
    
    Combines SpecVisionImageProcessor for images and a tokenizer for text,
    coordinating their interaction for multi-modal inputs.
    """

    attributes = ["image_processor", "tokenizer"]
    image_processor_class = "SpecVisionImageProcessor"
    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
    special_image_token = "<|image|>"

    def __init__(self, image_processor, tokenizer):
        self.image_processor = image_processor
        self.tokenizer = tokenizer
        self.num_img_tokens = image_processor.num_crops
        self.img_tokens = [f"<|image_{i+1}|>" for i in range(1000000)]

    def __call__(
        self,
        text: Union[TextInput, List[TextInput]],
        images: ImageInput = None,
        padding: Union[bool, str, PaddingStrategy] = False,
        truncation: Union[bool, str, TruncationStrategy] = None,
        max_length=None,
        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
    ) -> BatchFeature:
        """Process both text and image inputs for the model."""
        if images is not None:
            image_features = self.image_processor(images, return_tensors=return_tensors)
        else:
            image_features = {}

        # Process combined inputs
        inputs = self._process_multimodal_inputs(
            image_features,
            text,
            padding=padding,
            truncation=truncation,
            max_length=max_length,
            return_tensors=return_tensors
        )
        
        return inputs

    def _process_multimodal_inputs(self, images, texts, **kwargs):
        """Process and combine image and text inputs."""
        if not images:
            return BatchFeature(data=self.tokenizer(
                texts,
                return_tensors=kwargs.get('return_tensors'),
                padding=kwargs.get('padding'),
                truncation=kwargs.get('truncation'),
                max_length=kwargs.get('max_length')
            ))

        # Process text chunks and image tags
        pattern = r"<\|image_\d+\|>"
        text_chunks = [
            self.tokenizer(chunk).input_ids
            for chunk in re.split(pattern, texts)
        ]

        # Handle image tokens
        num_img_tokens = (
            images['num_img_tokens']
            if 'num_img_tokens' in images
            else [self.num_img_tokens] * len(images['pixel_values'])
        )

        image_tags = re.findall(pattern, texts)
        image_ids = [int(tag.split("|")[1].split("_")[-1]) for tag in image_tags]
        
        # Validate image IDs
        unique_ids = sorted(set(image_ids))
        if unique_ids != list(range(1, len(unique_ids) + 1)):
            raise ValueError(
                f"Image IDs must be consecutive integers starting from 1, got {unique_ids}"
            )
        if len(unique_ids) != len(images['pixel_values']):
            raise ValueError(
                f"Number of image tags ({len(unique_ids)}) doesn't match "
                f"number of images ({len(images['pixel_values'])})"
            )

        # Create padded image IDs
        image_ids_padded = [
            [-iid] * num_img_tokens[iid-1]
            for iid in image_ids
        ]

        # Combine text and image tokens
        input_ids = []
        for x in self._interleave_sequences(text_chunks, image_ids_padded):
            input_ids.extend(x)

        input_ids = torch.tensor(input_ids, dtype=torch.long).unsqueeze(0)
        attention_mask = (input_ids > -1000000).to(torch.long)

        return BatchFeature(data={
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "pixel_values": images['pixel_values'],
            "image_sizes": images['image_sizes']
        })

    def _interleave_sequences(self, seq1, seq2):
        """Interleave two sequences, padding second sequence if needed."""
        if len(seq1) > len(seq2):
            seq2.append([])
        return [item for pair in zip(seq1, seq2) for item in pair]

    def batch_decode(self, *args, **kwargs):
        """Decode a batch of token IDs to text."""
        return self.tokenizer.batch_decode(*args, **kwargs)

    def decode(self, *args, **kwargs):
        """Decode token IDs to text."""
        return self.tokenizer.decode(*args, **kwargs)

    @property
    def model_input_names(self):
        """Get combined input names from both processors."""
        return list(dict.fromkeys(
            self.tokenizer.model_input_names +
            self.image_processor.model_input_names
        ))

# Register the processor with AutoImageProcessor
AutoImageProcessor.register("SpecVisionImageProcessor", SpecVisionImageProcessor)