xmodelvlm/utils.py

import os
import sys
import torch
import base64
import datetime
import logging
import requests
import logging.handlers
from PIL import Image
from io import BytesIO
from transformers import StoppingCriteria
from xmodelvlm.constants import LOGDIR, IMAGE_TOKEN_INDEX


server_error_msg = "**NETWORK ERROR DUE TO HIGH TRAFFIC. PLEASE REGENERATE OR REFRESH THIS PAGE.**"
moderation_msg = "YOUR INPUT VIOLATES OUR CONTENT MODERATION GUIDELINES. PLEASE TRY AGAIN."

handler = None


def load_image_from_base64(image):
    return Image.open(BytesIO(base64.b64decode(image)))


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


def process_images(images, image_processor, model_cfg):
    image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
    new_images = []
    if image_aspect_ratio == 'pad':
        for image in images:
            image = expand2square(image, tuple(int(x*255) for x in image_processor.image_mean))
            image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
            new_images.append(image)
    else:
        return image_processor(images, return_tensors='pt')['pixel_values']
    if all(x.shape == new_images[0].shape for x in new_images):
        new_images = torch.stack(new_images, dim=0)
    return new_images


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('<image>')]

    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 get_model_name_from_path(model_path):
    model_path = model_path.strip("/")
    model_paths = model_path.split("/")
    if model_paths[-1].startswith('checkpoint-'):
        return model_paths[-2] + "_" + model_paths[-1]
    else:
        return model_paths[-1]


class KeywordsStoppingCriteria(StoppingCriteria):
    def __init__(self, keywords, tokenizer, input_ids):
        self.keywords = keywords
        self.keyword_ids = []
        self.max_keyword_len = 0
        for keyword in keywords:
            cur_keyword_ids = tokenizer(keyword).input_ids
            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
                cur_keyword_ids = cur_keyword_ids[1:]
            if len(cur_keyword_ids) > self.max_keyword_len:
                self.max_keyword_len = len(cur_keyword_ids)
            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
        self.tokenizer = tokenizer
        self.start_len = input_ids.shape[1]

    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        assert output_ids.shape[0] == 1, "Only support batch size 1 (yet)"  # TODO
        offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
        for keyword_id in self.keyword_ids:
            if (output_ids[0, -keyword_id.shape[0]:] == keyword_id).all():
                return True
        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
        for keyword in self.keywords:
            if keyword in outputs:
                return True
        return False


def build_logger(logger_name, logger_filename):
    global handler

    formatter = logging.Formatter(
        fmt="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
        datefmt="%Y-%m-%d %H:%M:%S",
    )

    # Set the format of root handlers
    if not logging.getLogger().handlers:
        logging.basicConfig(level=logging.INFO)
    logging.getLogger().handlers[0].setFormatter(formatter)

    # Redirect stdout and stderr to loggers
    stdout_logger = logging.getLogger("stdout")
    stdout_logger.setLevel(logging.INFO)
    sl = StreamToLogger(stdout_logger, logging.INFO)
    sys.stdout = sl

    stderr_logger = logging.getLogger("stderr")
    stderr_logger.setLevel(logging.ERROR)
    sl = StreamToLogger(stderr_logger, logging.ERROR)
    sys.stderr = sl

    # Get logger
    logger = logging.getLogger(logger_name)
    logger.setLevel(logging.INFO)

    # Add a file handler for all loggers
    if handler is None:
        os.makedirs(LOGDIR, exist_ok=True)
        filename = os.path.join(LOGDIR, logger_filename)
        handler = logging.handlers.TimedRotatingFileHandler(
            filename, when='D', utc=True)
        handler.setFormatter(formatter)

        for name, item in logging.root.manager.loggerDict.items():
            if isinstance(item, logging.Logger):
                item.addHandler(handler)

    return logger


class StreamToLogger(object):
    """
    Fake file-like stream object that redirects writes to a logger instance.
    """
    def __init__(self, logger, log_level=logging.INFO):
        self.terminal = sys.stdout
        self.logger = logger
        self.log_level = log_level
        self.linebuf = ''

    def __getattr__(self, attr):
        return getattr(self.terminal, attr)

    def write(self, buf):
        temp_linebuf = self.linebuf + buf
        self.linebuf = ''
        for line in temp_linebuf.splitlines(True):
            # From the io.TextIOWrapper docs:
            #   On output, if newline is None, any '\n' characters written
            #   are translated to the system default line separator.
            # By default sys.stdout.write() expects '\n' newlines and then
            # translates them so this is still cross platform.
            if line[-1] == '\n':
                self.logger.log(self.log_level, line.rstrip())
            else:
                self.linebuf += line

    def flush(self):
        if self.linebuf != '':
            self.logger.log(self.log_level, self.linebuf.rstrip())
        self.linebuf = ''


def disable_torch_init():
    """
    Disable the redundant torch default initialization to accelerate model creation.
    """
    import torch
    setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
    setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)


def violates_moderation(text):
    """
    Check whether the text violates OpenAI moderation API.
    """
    url = "https://api.openai.com/v1/moderations"
    headers = {"Content-Type": "application/json",
               "Authorization": "Bearer " + os.environ["OPENAI_API_KEY"]}
    text = text.replace("\n", "")
    data = "{" + '"input": ' + f'"{text}"' + "}"
    data = data.encode("utf-8")
    try:
        ret = requests.post(url, headers=headers, data=data, timeout=5)
        flagged = ret.json()["results"][0]["flagged"]
    except requests.exceptions.RequestException as e:
        flagged = False
    except KeyError as e:
        flagged = False

    return flagged


def pretty_print_semaphore(semaphore):
    if semaphore is None:
        return "None"
    return f"Semaphore(value={semaphore._value}, locked={semaphore.locked()})"