vlmeval/api/taichu.py

from vlmeval.smp import *
from vlmeval.api.base import BaseAPI
import os
import re
import json

from PIL import Image
import base64
from io import BytesIO
import copy


class ChatResponse(dict):
    def __getattr__(self, name):
        value = self.get(name)
        if isinstance(value, dict):
            return ChatResponse(value)  # 如果值是字典，递归包装成 DotDict
        elif isinstance(value, list):
            return [ChatResponse(v) if isinstance(v, dict) else v for v in value]  # 如果值是列表，处理其中的字典
        return value

    def __setattr__(self, name, value):
        self[name] = value

    def __delattr__(self, name):
        del self[name]


from ..dataset import DATASET_TYPE


class TaichuVLWrapper(BaseAPI):
    is_api: bool = True

    def __init__(self,
                 model: str = 'Taichu-VL-2B',
                 retry: int = 5,
                 wait: int = 5,
                 verbose: bool = True,
                 temperature: float = 0.0,
                 system_prompt: str = None,
                 max_tokens: int = 4096,
                 key: str = None,
                 url: str = None,
                 **kwargs):

        self.model = model
        self.kwargs = kwargs
        self.max_tokens = max_tokens

        self.system_prompt = '[sys]You are a helpful assistant.[/sys]'
        self.hint_prompt = '|<Hint>|'
        self.mcq_prompt = '|<MCQ>|'

        self.datasets_use_system = ['MMVet']
        self.datasets_use_multichoice = [
            'MathVista', 'MathVision']

        openai_key = os.environ.get('OPENAI_API_KEY', None)
        use_openai = os.environ.get('USE_OPENAI_EVAL', True)
        self.use_openai_evaluate = (isinstance(openai_key, str) and openai_key.startswith('sk-') and use_openai)

        self.api_key = os.environ.get('TAICHU_API_KEY', key)
        self.api_url = url

        assert self.api_key is not None, 'Please set the API Key'

        super().__init__(wait=wait, retry=retry, system_prompt=self.system_prompt, verbose=verbose, **kwargs)

    def set_dump_image(self, dump_image_func):
        self.dump_image_func = dump_image_func

    def dump_image(self, line, dataset):
        return self.dump_image_func(line)

    def use_custom_prompt(self, dataset):
        if listinstr(['MCQ', 'VQA'], DATASET_TYPE(dataset)):
            return True
        elif dataset is not None and listinstr(['HallusionBench'], dataset):
            return True
        return False

    def clear_prompt(self, prompt):
        prompt = re.sub(r"Hint:.*?Question:", "", prompt, flags=re.S).strip()
        prompt = re.sub(r"\nChoices:\n.*", "", prompt, flags=re.S).strip()
        return prompt

    def encode_image(self, pil_image):
        buffer = BytesIO()
        pil_image.save(buffer, format='PNG')
        base64_str = base64.b64encode(buffer.getvalue()).decode("utf-8")
        return base64_str

    def build_prompt(self, line, dataset=None):
        if isinstance(line, int):
            line = self.data.iloc[line]

        tgt_path = self.dump_image(line, dataset)
        question = line['question']
        hint = None
        if listinstr(self.datasets_use_system, dataset):
            system_prompt = self.system_prompt
        else:
            system_prompt = ''
        mcq = False
        if DATASET_TYPE(dataset) == 'MCQ' or listinstr(self.datasets_use_multichoice, dataset):
            options = {
                cand: line[cand]
                for cand in string.ascii_uppercase
                if cand in line and not pd.isna(line[cand])
            }
            if listinstr(self.datasets_use_multichoice, dataset):
                options = {}
                if not pd.isna(line['choices']):
                    for i, c in enumerate(eval(line['choices'])):
                        options[string.ascii_uppercase[i]] = c
                question = self.clear_prompt(question)

            # support chinese
            if listinstr(['_CN', '_cn'], dataset):
                options_prompt = '\n选项：\n'
            else:
                options_prompt = '\nOPTIONS:\n'
            options_prompt += '\n'.join(f"{key}:{value}" for key, value in options.items())
            hint = line['hint'] if ('hint' in line and not pd.isna(line['hint'])) else None
            mcq = True if len(options) else False
            if len(options):
                prompt = question + options_prompt
            else:
                prompt = question
        else:
            prompt = question

        msgs = []
        if system_prompt:
            msgs.append(dict(type='text', value=system_prompt))

        if isinstance(tgt_path, list):
            msgs.extend([dict(type='image', value=p) for p in tgt_path])
        else:
            msgs.append(dict(type='image', value=tgt_path))

        if hint:
            prompt = 'Hint: ' + hint + '\n' + prompt
        msgs.append(dict(type='text', value=prompt))

        if mcq:
            msgs.append(dict(type='text', value=self.mcq_prompt))
        return msgs

    def prompt_to_request_messages(self, inputs):

        messages = [
            {'role': 'user', 'content': []}
        ]
        is_mcq = False
        for x in inputs:
            if x['type'] == 'text':
                if x['value'] == self.system_prompt:
                    messages = [{'role': 'system', 'content': [{"type": "text", "text": x['value']}]}] + messages
                elif self.mcq_prompt == x['value']:
                    is_mcq = True
                else:
                    messages[-1]['content'].append(
                        {"type": "text", "text": x['value']},
                    )
            if x['type'] == 'image':
                _url = self.encode_image(Image.open(x['value']))
                messages[-1]['content'].append(
                    {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{_url}"}},
                )
            else:
                continue

        return messages, is_mcq

    def generate_inner(self, inputs, **kwargs) -> str:
        messages, is_mcq = self.prompt_to_request_messages(inputs)

        data = {
            "model": self.model,
            "messages": messages,
            "max_tokens": self.max_tokens,
            "temperature": 0,
            "top_p": 0.8,
            "stream": False,
            "extra_body": {
                "repetition_penalty": 1
            }
        }

        headers = {
            'Authorization': self.api_key,
            'Content-Type': 'application/json'
        }

        try:
            chat_response = requests.post(self.api_url, json=data, headers=headers)
            response = ChatResponse(json.loads(chat_response.content))
            result = response.choices[0].message.content
            # Extract index to exact matching when ChatGPT is unavailable.
            if self.use_openai_evaluate is False and is_mcq is True:
                try:
                    result = result[0]
                except:
                    result = 'A'
            return 0, result, 'Succeeded! '
        except Exception as err:
            if self.verbose:
                self.logger.error(f'{type(err)}: {err}')
                self.logger.error(f'The input messages are {inputs}.')
            return -1, '', ''


class TaichuVLAPI(TaichuVLWrapper):

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


class TaichuVLRWrapper(BaseAPI):
    is_api: bool = True

    def __init__(self,
                 model: str = 'taichu_vlr_3b',
                 retry: int = 5,
                 wait: int = 5,
                 verbose: bool = True,
                 temperature: float = 0.0,
                 system_prompt: str = None,
                 max_tokens: int = 4096,
                 use_reasoning_prompt: bool = True,
                 post_process: bool = True,
                 key: str = None,
                 url: str = None,
                 **kwargs):

        self.model = model
        self.kwargs = kwargs
        self.max_tokens = max_tokens
        self.system_prompt = system_prompt
        self.use_reasoning_prompt = use_reasoning_prompt
        self.post_process = post_process
        self.verbose = verbose

        openai_key = os.environ.get('OPENAI_API_KEY', None)
        use_openai = os.environ.get('USE_OPENAI_EVAL', True)
        self.use_openai_evaluate = (isinstance(openai_key, str) and openai_key.startswith('sk-') and use_openai)

        self.api_key = os.environ.get('TAICHU_API_KEY', key)
        self.api_url = url

        assert self.api_key is not None, 'Please set the API Key'

        super().__init__(wait=wait, retry=retry, system_prompt=self.system_prompt, verbose=verbose, **kwargs)

    def use_custom_prompt(self, dataset):
        return False

    def encode_image(self, pil_image):
        buffer = BytesIO()
        pil_image.save(buffer, format='PNG')
        base64_str = base64.b64encode(buffer.getvalue()).decode("utf-8")
        return base64_str

    def post_process_func(self, response):
        resp = response.split('\\boxed{')[-1]
        lt = len(resp)
        counter, end = 1, None
        for i in range(lt):
            if resp[i] == '{':
                counter += 1
            elif resp[i] == '}':
                counter -= 1
            if counter == 0:
                end = i
                break
            elif i == lt - 1:
                end = lt
                break
        if end is not None:
            response = resp[:end]
        return response

    def prompt_to_request_messages(self, inputs):

        messages = [
            {'role': 'user', 'content': []}
        ]
        for x in inputs:
            if x['type'] == 'text':
                messages[-1]['content'].append(
                    {"type": "text", "text": x['value']},
                )
            if x['type'] == 'image':
                _url = self.encode_image(Image.open(x['value']))
                messages[-1]['content'].append(
                    {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{_url}"}},
                )
            else:
                continue
        
        PROMPT = (
            "First thinks about the reasoning process in the mind and then provides the user with the answer. "
            "Put your final answer within \\boxed{}. "
            "The response of reasoning and answer are formatted in <think> reasoning </think><answer> \\boxed{answer here} </answer>.\n"
        )

        if self.use_reasoning_prompt:
            for content in messages[0]['content']:
                if content['type'] == 'text':
                    content['text'] = PROMPT + content['text']
                    break

        return messages

    def generate_inner(self, inputs, **kwargs) -> str:
        messages = self.prompt_to_request_messages(inputs)
        if self.verbose:
            verbose_messages = copy.deepcopy(messages)
            for mess in verbose_messages:
                if mess['role'] == 'user':
                    for content in mess['content']:
                        if content['type'] == 'image_url':
                            content['image_url']['url'] = ''
            print(f'\033[31m{verbose_messages}\033[0m')

        data = {
            "model": self.model,
            "messages": messages,
            "max_tokens": self.max_tokens,
            "temperature": 0,
            "top_p": 0.8,
            "stream": False,
            "repetition_penalty": 1.0
        }

        headers = {
            'Authorization': f"Bearer {self.api_key}",
            'Content-Type': 'application/json'
        }

        try:
            chat_response = requests.post(self.api_url, json=data, headers=headers)
            response = ChatResponse(json.loads(chat_response.content))
            result = response.choices[0].message.content
            if self.post_process:
                result = self.post_process_func(result)
            if self.verbose:
                print(f'\033[32m{result}\033[0m')
            
            return 0, result, 'Succeeded! '
        except Exception as err:
            if self.verbose:
                self.logger.error(f'{type(err)}: {err}')
                self.logger.error(f'The input messages are {inputs}.')
            return -1, '', ''


class TaichuVLRAPI(TaichuVLRWrapper):

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