llava/utils.py

import datetime
import logging
import logging.handlers
import os
import sys
import math
import random
import requests
import torch.distributed as dist

from llava.constants import LOGDIR

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 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, encoding='UTF-8')
        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()})"

def master_print(*args):
    import torch
    if torch.cuda.current_device() == 0:
        print(*args)

def is_dist_avail_and_initialized():
    if not dist.is_available():
        return False
    if not dist.is_initialized():
        return False
    return True

def get_world_size():
    if not is_dist_avail_and_initialized():
        return 1
    return dist.get_world_size()


def get_rank():
    if not is_dist_avail_and_initialized():
        return 0
    return dist.get_rank()

def is_main_process():
    return get_rank() == 0


class DatasetIter(object):
    def __init__(self, size, world_size, local_rank, num_workers=1):
        self.size = size
        self.world_size = world_size
        self.local_rank = local_rank
        # self.num_workers = 1 if num_workers == 0 else num_workers
        assert num_workers == 1, 'num workers must be 1'
        self.num_workers = num_workers
        self.per_worker = int(math.floor(self.size / float(self.world_size * self.num_workers)))
        self.worker_indexs = dict()

        for worker_id in range(self.num_workers):
            self.init_worker_index(worker_id)
    def init_worker_index(self, worker_id):

        start = self.per_worker * (self.local_rank * self.num_workers + worker_id)
        end = min(start + self.per_worker, self.size)
        rank_indexs = list(range(start, end))
        random.shuffle(rank_indexs)

        self.worker_indexs[worker_id] = rank_indexs

    def increment(self, worker_id):

        if len(self.worker_indexs[worker_id]) == 0:
            self.init_worker_index(worker_id)

        next_iter, self.worker_indexs[worker_id] = self.worker_indexs[worker_id][0], self.worker_indexs[worker_id][1:]
        return next_iter