internvl_chat.py

import torch
from transformers import AutoTokenizer, AutoModel, CLIPImageProcessor
import warnings
from PIL import Image
from .base import BaseModel
from ..smp import *
from ..dataset import DATASET_TYPE
import pandas as pd
import string
import torchvision.transforms as T
import transformers

from torchvision.transforms.functional import InterpolationMode
import random

IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)


def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=5, max_num=6, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images, target_aspect_ratio

def dynamic_preprocess2(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False, prior_aspect_ratio=None):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    new_target_ratios = []
    if prior_aspect_ratio is not None:
        for i in target_ratios:
            if prior_aspect_ratio[0]%i[0] !=0 or prior_aspect_ratio[1]%i[1] !=0:
                new_target_ratios.append(i)
            else:
                continue
    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, new_target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images

def load_image(image_file, input_size=448, min_num=1, max_num=6):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images, target_aspect_ratio = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, min_num=min_num, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values, target_aspect_ratio

def load_image2(image_file, input_size=448, target_aspect_ratio=(1,1), min_num=1, max_num=6):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess2(image, image_size=input_size, prior_aspect_ratio=target_aspect_ratio, use_thumbnail=True, min_num=min_num, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

class InternVLChat(BaseModel):

    INSTALL_REQ = False
    INTERLEAVE = False

    def __init__(self, model_path='OpenGVLab/InternVL-Chat-V1-5', load_in_8bit=False, **kwargs):
        assert model_path is not None
        assert version_cmp(transformers.__version__, '4.36.2', 'ge')
        self.model_path = model_path
        self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, use_fast=False)
        device = torch.cuda.current_device()
        self.device = device
        self.model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16,
                                               trust_remote_code=True,
                                               load_in_8bit=load_in_8bit).eval()
        if not load_in_8bit:
            self.model = self.model.to(device)
        self.image_size = self.model.config.vision_config.image_size

        if 'V1-1' in model_path:
            kwargs_default = dict(do_sample=False, max_new_tokens=1024, top_p=None, num_beams=5)
        else:
            kwargs_default = dict(do_sample=False, max_new_tokens=1024, top_p=None, num_beams=1)
        kwargs_default.update(kwargs)
        self.kwargs = kwargs_default
        warnings.warn(f'Following kwargs received: {self.kwargs}, will use as generation config. ')

    def use_custom_prompt(self, dataset):
        return True

    def build_multi_choice_prompt(self, line, dataset=None):
        question = line['question']
        hint = line['hint'] if ('hint' in line and not pd.isna(line['hint'])) else None
        if hint is not None:
            question = hint + '\n' + question

        options = {
            cand: line[cand]
            for cand in string.ascii_uppercase
            if cand in line and not pd.isna(line[cand])
        }
        for key, item in options.items():
            question += f'\n{key}. {item}'
        prompt = question

        if len(options):
            prompt += '\n请直接回答选项字母。' if cn_string(
                prompt) else "\nAnswer with the option's letter from the given choices directly."
        else:
            prompt += '\n请直接回答问题。' if cn_string(prompt) else '\nAnswer the question directly.'

        return prompt

    def build_prompt(self, line, dataset=None):
        assert self.use_custom_prompt(dataset)
        assert dataset is None or isinstance(dataset, str)
        tgt_path = self.dump_image(line, dataset)

        if 'V1-1' in self.model_path:
            kwargs_default = dict(do_sample=False, max_new_tokens=1024, top_p=None, num_beams=5)
        else:
            kwargs_default = dict(do_sample=False, max_new_tokens=1024, top_p=None, num_beams=1)
        self.kwargs = kwargs_default
        if dataset is not None and listinstr(['MME'], dataset):
            question = line['question']
            prompt = question + ' Answer the question using a single word or phrase.'
            if 'V1-2' not in self.model_path:
                self.kwargs = dict(do_sample=True, max_new_tokens=5, top_k=50, num_beams=5, top_p=0.9)
        elif dataset is not None and listinstr(['HallusionBench'], dataset):
            question = line['question']
            prompt = question + ' Please answer yes or no. Answer the question using a single word or phrase.'
        elif dataset is not None and DATASET_TYPE(dataset) == 'multi-choice':
            prompt = self.build_multi_choice_prompt(line, dataset)
        elif dataset is not None and DATASET_TYPE(dataset) == 'VQA':
            if 'MathVista' in dataset:
                prompt = line['question']
            elif listinstr(['LLaVABench'], dataset):
                question = line['question']
                prompt = question + '\nAnswer this question in detail.'
            elif listinstr(['MMVet'], dataset):
                prompt = line['question']
            else:
                question = line['question']
                prompt = question + '\nAnswer the question using a single word or phrase.'
        else:
            prompt = line['question']

        message = [dict(type='text', value=prompt)]
        message.extend([dict(type='image', value=s) for s in tgt_path])

        return message

    def generate(self, message, dataset=None):
        prompt, image_path = self.message_to_promptimg(message)
        if dataset is not None and listinstr(['ChartQA_TEST'], dataset):
            self.max_num = 12
            self.max_num2 = 3
        elif dataset is not None and listinstr(['DocVQA_VAL', 'DocVQA_TEST', 'TextVQA_VAL'], dataset):
            self.max_num = 23
            self.max_num2 = 15
            self.min_num = 14
            self.min_num2 = 5
        elif dataset is not None and listinstr(['InfoVQA_VAL', 'InfoVQA_TEST'], dataset):
            self.max_num = 23
            self.max_num2 = 5
            self.min_num = 15
            self.min_num2 = 3
        elif dataset is not None and listinstr(['OCRBench'], dataset):
            self.max_num = 24
            self.max_num2 = 8
            self.min_num = 9
            self.min_num2 = 5
        else:
            self.max_num = 8
            self.max_num2 = 4
            self.min_num = 3
            self.min_num2 = 1
        pixel_values, target_aspect_ratio = load_image(image_path, min_num=self.min_num, max_num=self.max_num)
        pixel_values = pixel_values.cuda().to(torch.bfloat16)
        pixel_values2 = load_image2(image_path, target_aspect_ratio=target_aspect_ratio, min_num=self.min_num2, max_num=self.max_num2)
        pixel_values2 = pixel_values2.cuda().to(torch.bfloat16)
        pixel_values = torch.cat((pixel_values[:-1],  pixel_values2[:-1], pixel_values[-1:]), 0)

        with torch.no_grad():
            response = self.model.chat(self.tokenizer, pixel_values=pixel_values, target_aspect_ratio=target_aspect_ratio,
                                       question=prompt, generation_config=self.kwargs)
        response = response.split('[UNUSED_TOKEN_145]')[0]

        return response

    def generate_inner(self, message, dataset=None):
        return self.generate(message, dataset)