# --------------------------------------------------------
# Image as a Foreign Language: BEiT Pretraining for Vision and Vision-Language Tasks (https://arxiv.org/abs/2208.10442)
# Github source: https://github.com/microsoft/unilm/tree/master/beit3
# Copyright (c) 2023 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# --------------------------------------------------------'

import argparse
import datetime
import io
import json
import math
import os
import time
from collections import defaultdict, deque
from pathlib import Path

import numpy as np
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
from timm.utils import get_state_dict
from torch import inf


def bool_flag(s):
    """
    Parse boolean arguments from the command line.
    """
    FALSY_STRINGS = {"off", "false", "0"}
    TRUTHY_STRINGS = {"on", "true", "1"}
    if s.lower() in FALSY_STRINGS:
        return False
    elif s.lower() in TRUTHY_STRINGS:
        return True
    else:
        raise argparse.ArgumentTypeError("invalid value for a boolean flag")


class SmoothedValue:
    """Track a series of values and provide access to smoothed values over a
    window or the global series average.
    """

    def __init__(self, window_size=20, fmt=None):
        if fmt is None:
            fmt = "{median:.4f} ({global_avg:.4f})"
        self.deque = deque(maxlen=window_size)
        self.total = 0.0
        self.count = 0
        self.fmt = fmt

    def update(self, value, n=1):
        self.deque.append(value)
        self.count += n
        self.total += value * n

    def synchronize_between_processes(self):
        """
        Warning: does not synchronize the deque!
        """
        if not is_dist_avail_and_initialized():
            return
        t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda')
        dist.barrier()
        dist.all_reduce(t)
        t = t.tolist()
        self.count = int(t[0])
        self.total = t[1]

    @property
    def median(self):
        d = torch.tensor(list(self.deque))
        return d.median().item()

    @property
    def avg(self):
        d = torch.tensor(list(self.deque), dtype=torch.float32)
        return d.mean().item()

    @property
    def global_avg(self):
        return self.total / self.count

    @property
    def max(self):
        return max(self.deque)

    @property
    def value(self):
        return self.deque[-1]

    def __str__(self):
        return self.fmt.format(
            median=self.median,
            avg=self.avg,
            global_avg=self.global_avg,
            max=self.max,
            value=self.value,
        )


class MetricLogger:
    def __init__(self, delimiter="\t"):
        self.meters = defaultdict(SmoothedValue)
        self.delimiter = delimiter

    def update(self, **kwargs):
        for k, v in kwargs.items():
            if v is None:
                continue
            if isinstance(v, torch.Tensor):
                v = v.item()
            assert isinstance(v, (float, int))
            self.meters[k].update(v)

    def __getattr__(self, attr):
        if attr in self.meters:
            return self.meters[attr]
        if attr in self.__dict__:
            return self.__dict__[attr]
        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{attr}'")

    def __str__(self):
        loss_str = []
        for name, meter in self.meters.items():
            loss_str.append(f"{name}: {str(meter)}")
        return self.delimiter.join(loss_str)

    def synchronize_between_processes(self):
        for meter in self.meters.values():
            meter.synchronize_between_processes()

    def add_meter(self, name, meter):
        self.meters[name] = meter

    def log_every(self, iterable, print_freq, header=None):
        i = 0
        if not header:
            header = ''
        start_time = time.time()
        end = time.time()
        iter_time = SmoothedValue(fmt='{avg:.4f}')
        data_time = SmoothedValue(fmt='{avg:.4f}')
        space_fmt = ':' + str(len(str(len(iterable)))) + 'd'
        log_msg = [
            header,
            '[{0' + space_fmt + '}/{1}]',
            'eta: {eta}',
            '{meters}',
            'time: {time}',
            'data: {data}',
        ]
        if torch.cuda.is_available():
            log_msg.append('max mem: {memory:.0f}')
        log_msg = self.delimiter.join(log_msg)
        MB = 1024.0 * 1024.0
        for obj in iterable:
            data_time.update(time.time() - end)
            yield obj
            iter_time.update(time.time() - end)
            if i % print_freq == 0 or i == len(iterable) - 1:
                eta_seconds = iter_time.global_avg * (len(iterable) - i)
                eta_string = str(datetime.timedelta(seconds=int(eta_seconds)))
                if torch.cuda.is_available():
                    print(
                        log_msg.format(
                            i,
                            len(iterable),
                            eta=eta_string,
                            meters=str(self),
                            time=str(iter_time),
                            data=str(data_time),
                            memory=torch.cuda.max_memory_allocated() / MB,
                        )
                    )
                else:
                    print(
                        log_msg.format(
                            i,
                            len(iterable),
                            eta=eta_string,
                            meters=str(self),
                            time=str(iter_time),
                            data=str(data_time),
                        )
                    )
            i += 1
            end = time.time()
        total_time = time.time() - start_time
        total_time_str = str(datetime.timedelta(seconds=int(total_time)))
        print(
            '{} Total time: {} ({:.4f} s / it)'.format(
                header, total_time_str, total_time / len(iterable)
            )
        )


def _load_checkpoint_for_ema(model_ema, checkpoint):
    """
    Workaround for ModelEma._load_checkpoint to accept an already-loaded object
    """
    mem_file = io.BytesIO()
    torch.save(checkpoint, mem_file)
    mem_file.seek(0)
    model_ema._load_checkpoint(mem_file)


def setup_for_distributed(is_master):
    """
    This function disables printing when not in master process
    """
    import builtins as __builtin__

    builtin_print = __builtin__.print

    def print(*args, **kwargs):
        force = kwargs.pop('force', False)
        if is_master or force:
            builtin_print(*args, **kwargs)

    __builtin__.print = print


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


def save_on_master(*args, **kwargs):
    if is_main_process():
        torch.save(*args, **kwargs)


def _get_rank_env():
    if "RANK" in os.environ:
        return int(os.environ["RANK"])
    else:
        return int(os.environ['OMPI_COMM_WORLD_RANK'])


def _get_local_rank_env():
    if "LOCAL_RANK" in os.environ:
        return int(os.environ["LOCAL_RANK"])
    else:
        return int(os.environ['OMPI_COMM_WORLD_LOCAL_RANK'])


def _get_world_size_env():
    if "WORLD_SIZE" in os.environ:
        return int(os.environ["WORLD_SIZE"])
    else:
        return int(os.environ['OMPI_COMM_WORLD_SIZE'])


# The implementation code is modified from DeiT (https://github.com/facebookresearch/deit.git)
def init_distributed_mode(args):
    if args.dist_on_itp:
        args.rank = _get_rank_env()
        args.world_size = _get_world_size_env()  # int(os.environ['OMPI_COMM_WORLD_SIZE'])
        args.gpu = _get_local_rank_env()
        args.dist_url = "tcp://{}:{}".format(os.environ['MASTER_ADDR'], os.environ['MASTER_PORT'])
        os.environ['LOCAL_RANK'] = str(args.gpu)
        os.environ['RANK'] = str(args.rank)
        os.environ['WORLD_SIZE'] = str(args.world_size)
        # ["RANK", "WORLD_SIZE", "MASTER_ADDR", "MASTER_PORT", "LOCAL_RANK"]
    elif 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:
        args.rank = int(os.environ["RANK"])
        args.world_size = int(os.environ['WORLD_SIZE'])
        args.gpu = int(os.environ['LOCAL_RANK'])
    elif 'SLURM_PROCID' in os.environ:
        args.rank = int(os.environ['SLURM_PROCID'])
        args.gpu = args.rank % torch.cuda.device_count()
    else:
        print('Not using distributed mode')
        args.distributed = False
        return

    args.distributed = True

    torch.cuda.set_device(args.gpu)
    args.dist_backend = 'nccl'
    print(
        f'| distributed init (rank {args.rank}): {args.dist_url}, gpu {args.gpu}',
        flush=True,
    )
    torch.distributed.init_process_group(
        backend=args.dist_backend,
        init_method=args.dist_url,
        world_size=args.world_size,
        rank=args.rank,
        timeout=datetime.timedelta(0, 7200),
    )
    torch.distributed.barrier()
    setup_for_distributed(args.rank == 0)


def load_state_dict(model, state_dict, prefix='', ignore_missing="relative_position_index"):
    missing_keys = []
    unexpected_keys = []
    error_msgs = []
    # copy state_dict so _load_from_state_dict can modify it
    metadata = getattr(state_dict, '_metadata', None)
    state_dict = state_dict.copy()
    if metadata is not None:
        state_dict._metadata = metadata

    def load(module, prefix=''):
        local_metadata = {} if metadata is None else metadata.get(prefix[:-1], {})
        module._load_from_state_dict(
            state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs
        )
        for name, child in module._modules.items():
            if child is not None:
                load(child, prefix + name + '.')

    load(model, prefix=prefix)

    warn_missing_keys = []
    ignore_missing_keys = []
    for key in missing_keys:
        keep_flag = True
        for ignore_key in ignore_missing.split('|'):
            if ignore_key in key:
                keep_flag = False
                break
        if keep_flag:
            warn_missing_keys.append(key)
        else:
            ignore_missing_keys.append(key)

    missing_keys = warn_missing_keys

    if len(missing_keys) > 0:
        print(
            "Weights of {} not initialized from pretrained model: {}".format(
                model.__class__.__name__, missing_keys
            )
        )
    if len(unexpected_keys) > 0:
        print(
            "Weights from pretrained model not used in {}: {}".format(
                model.__class__.__name__, unexpected_keys
            )
        )
    if len(ignore_missing_keys) > 0:
        print(
            "Ignored weights of {} not initialized from pretrained model: {}".format(
                model.__class__.__name__, ignore_missing_keys
            )
        )
    if len(error_msgs) > 0:
        print('\n'.join(error_msgs))


class NativeScalerWithGradNormCount:
    state_dict_key = "amp_scaler"

    def __init__(self):
        self._scaler = torch.cuda.amp.GradScaler()

    def __call__(
        self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True
    ):
        self._scaler.scale(loss).backward(create_graph=create_graph)
        if update_grad:
            if clip_grad is not None:
                assert parameters is not None
                self._scaler.unscale_(
                    optimizer
                )  # unscale the gradients of optimizer's assigned params in-place
                norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad)
            else:
                self._scaler.unscale_(optimizer)
                norm = get_grad_norm_(parameters)
            self._scaler.step(optimizer)
            self._scaler.update()
        else:
            norm = None
        return norm

    def state_dict(self):
        return self._scaler.state_dict()

    def load_state_dict(self, state_dict):
        self._scaler.load_state_dict(state_dict)


def get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor:
    if isinstance(parameters, torch.Tensor):
        parameters = [parameters]
    parameters = [p for p in parameters if p.grad is not None]
    norm_type = float(norm_type)
    if len(parameters) == 0:
        return torch.tensor(0.0)
    device = parameters[0].grad.device
    if norm_type == inf:
        total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters)
    else:
        total_norm = torch.norm(
            torch.stack([torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]),
            norm_type,
        )
    return total_norm


def cosine_scheduler(
    base_value,
    final_value,
    epochs,
    niter_per_ep,
    warmup_epochs=0,
    start_warmup_value=0,
    warmup_steps=-1,
    sched_type="cos",
):
    warmup_schedule = np.array([])
    warmup_iters = warmup_epochs * niter_per_ep
    if warmup_steps > 0:
        warmup_iters = warmup_steps
    print("Set warmup steps = %d" % warmup_iters)
    if warmup_epochs > 0:
        warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)

    if sched_type == "cos":
        iters = np.arange(epochs * niter_per_ep - warmup_iters)
        schedule = np.array(
            [
                final_value
                + 0.5 * (base_value - final_value) * (1 + math.cos(math.pi * i / (len(iters))))
                for i in iters
            ]
        )
    elif sched_type == "linear":
        schedule = np.linspace(base_value, final_value, epochs * niter_per_ep - warmup_iters)
    else:
        raise NotImplementedError()

    schedule = np.concatenate((warmup_schedule, schedule))

    assert len(schedule) == epochs * niter_per_ep
    return schedule


def save_model(args, epoch, model, model_without_ddp, optimizer, loss_scaler, model_ema=None):
    output_dir = Path(args.output_dir)
    if loss_scaler is not None:
        checkpoint_paths = [output_dir / ('checkpoint-%s.pth' % epoch)]
        for checkpoint_path in checkpoint_paths:
            to_save = {
                'model': model_without_ddp.state_dict(),
                'optimizer': optimizer.state_dict(),
                'epoch': epoch,
                'scaler': loss_scaler.state_dict(),
                'args': args,
            }

            if model_ema is not None:
                to_save['model_ema'] = get_state_dict(model_ema)

            save_on_master(to_save, checkpoint_path)
    else:
        client_state = {'epoch': epoch, "args": args}
        if model_ema is not None:
            client_state['model_ema'] = get_state_dict(model_ema)
        model.save_checkpoint(
            save_dir=args.output_dir, tag="checkpoint-%s" % epoch, client_state=client_state
        )


def auto_load_model(args, model, model_without_ddp, optimizer, loss_scaler, model_ema=None):
    output_dir = Path(args.output_dir)
    if loss_scaler is not None:
        # torch.amp
        if args.auto_resume and len(args.resume) == 0:
            import glob

            all_checkpoints = glob.glob(os.path.join(output_dir, 'checkpoint-*.pth'))
            latest_ckpt = -1
            for ckpt in all_checkpoints:
                t = ckpt.split('-')[-1].split('.')[0]
                if t.isdigit():
                    latest_ckpt = max(int(t), latest_ckpt)
            if latest_ckpt >= 0:
                args.resume = os.path.join(output_dir, 'checkpoint-%d.pth' % latest_ckpt)
            print("Auto resume checkpoint: %s" % args.resume)

        if args.resume:
            if args.resume.startswith('https'):
                checkpoint = torch.hub.load_state_dict_from_url(
                    args.resume, map_location='cpu', check_hash=True
                )
            else:
                checkpoint = torch.load(args.resume, map_location='cpu')
            model_without_ddp.load_state_dict(checkpoint['model'])
            print("Resume checkpoint %s" % args.resume)
            if 'optimizer' in checkpoint and 'epoch' in checkpoint:
                optimizer.load_state_dict(checkpoint['optimizer'])
                args.start_epoch = checkpoint['epoch'] + 1
                if hasattr(args, 'model_ema') and args.model_ema:
                    _load_checkpoint_for_ema(model_ema, checkpoint['model_ema'])
                if 'scaler' in checkpoint:
                    loss_scaler.load_state_dict(checkpoint['scaler'])
                print("With optim & sched!")
    else:
        # deepspeed, only support '--auto_resume'.
        if args.auto_resume:
            import glob

            all_checkpoints = glob.glob(os.path.join(output_dir, 'checkpoint-*'))
            latest_ckpt = -1
            for ckpt in all_checkpoints:
                t = ckpt.split('-')[-1].split('.')[0]
                if t.isdigit():
                    latest_ckpt = max(int(t), latest_ckpt)
            if latest_ckpt >= 0:
                args.resume = os.path.join(output_dir, 'checkpoint-%d' % latest_ckpt)
                print("Auto resume checkpoint: %d" % latest_ckpt)
                _, client_states = model.load_checkpoint(
                    args.output_dir, tag='checkpoint-%d' % latest_ckpt
                )
                args.start_epoch = client_states['epoch'] + 1
                if model_ema is not None:
                    if args.model_ema:
                        _load_checkpoint_for_ema(model_ema, client_states['model_ema'])


# The implementation code is modified from DeiT (https://github.com/facebookresearch/deit.git)
def load_model_and_may_interpolate(ckpt_path, model, model_key, model_prefix):
    if ckpt_path.startswith('https'):
        checkpoint = torch.hub.load_state_dict_from_url(
            ckpt_path, map_location='cpu', check_hash=True
        )
    else:
        checkpoint = torch.load(ckpt_path, map_location='cpu')

    print("Load ckpt from %s" % ckpt_path)
    checkpoint_model = None
    for model_key in model_key.split('|'):
        if model_key in checkpoint:
            checkpoint_model = checkpoint[model_key]
            print("Load state_dict by model_key = %s" % model_key)
            break

    if checkpoint_model is None:
        checkpoint_model = checkpoint

    state_dict = model.state_dict()
    for k in ['head.weight', 'head.bias']:
        if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape:
            print(f"Removing key {k} from pretrained checkpoint")
            del checkpoint_model[k]

    # interpolate position embedding
    for pos_embed_key in (
        "vision_pos_embed",
        "pos_embed",
        "beit3.encoder.embed_positions.A.weight",
    ):
        if pos_embed_key in checkpoint_model:
            pos_embed_checkpoint = checkpoint_model[pos_embed_key]
            embedding_size = pos_embed_checkpoint.shape[-1]
            if pos_embed_key == "beit3.encoder.embed_positions.A.weight":
                # being consistent with Fairseq, which starts from 2 for position embedding
                torchscale_model = True
                num_patches = model.beit3.vision_embed.num_patches
                num_extra_tokens = (
                    model.beit3.vision_embed.num_position_embeddings() + 2 - num_patches
                )
            else:
                torchscale_model = False
                num_patches = model.patch_embed.num_patches
                num_extra_tokens = getattr(model, pos_embed_key).shape[-2] - num_patches
            # height (== width) for the checkpoint position embedding
            orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
            # height (== width) for the new position embedding
            new_size = int(num_patches**0.5)
            # class_token and dist_token are kept unchanged
            if orig_size != new_size:
                print(
                    "Position interpolate from %dx%d to %dx%d"
                    % (orig_size, orig_size, new_size, new_size)
                )
                if torchscale_model:
                    extra_tokens = pos_embed_checkpoint[:num_extra_tokens].unsqueeze(0)
                    # only the position tokens are interpolated
                    pos_tokens = pos_embed_checkpoint[num_extra_tokens:]
                else:
                    extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
                    # only the position tokens are interpolated
                    pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
                pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(
                    0, 3, 1, 2
                )
                pos_tokens = torch.nn.functional.interpolate(
                    pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False
                )
                pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
                new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
                if torchscale_model:
                    new_pos_embed = new_pos_embed.squeeze(0)
                checkpoint_model[pos_embed_key] = new_pos_embed

    load_state_dict(model, checkpoint_model, prefix=model_prefix)


def create_ds_config(args):
    args.deepspeed_config = os.path.join(args.output_dir, "deepspeed_config.json")
    with open(args.deepspeed_config, mode="w") as writer:
        ds_config = {
            "train_batch_size": args.batch_size * args.update_freq * get_world_size(),
            "train_micro_batch_size_per_gpu": args.batch_size,
            "steps_per_print": 1000,
            "optimizer": {
                "type": "Adam",
                "adam_w_mode": True,
                "params": {
                    "lr": args.lr,
                    "weight_decay": args.weight_decay,
                    "bias_correction": True,
                    "betas": [args.opt_betas[0], args.opt_betas[1]],
                    "eps": args.opt_eps,
                },
            },
            "fp16": {
                "enabled": True,
                "loss_scale": 0,
                "initial_scale_power": getattr(args, "initial_scale_power", 12),
                "loss_scale_window": 1000,
                "hysteresis": 2,
                "min_loss_scale": 1,
            },
            "amp": {"enabled": False, "opt_level": "O2"},
        }

        if args.clip_grad is not None:
            ds_config.update({'gradient_clipping': args.clip_grad})

        if args.zero_stage == 1:
            ds_config.update(
                {"zero_optimization": {"stage": args.zero_stage, "reduce_bucket_size": 5e8}}
            )
        elif args.zero_stage > 1:
            raise NotImplementedError()

        writer.write(json.dumps(ds_config, indent=2))


def merge_batch_tensors_by_dict_key(batch):
    batch_tensors = {}
    for tensor_key in batch[0]:
        if isinstance(batch[0][tensor_key], torch.Tensor):
            batch_tensors[tensor_key] = torch.stack([d[tensor_key] for d in batch])
        else:
            batch_tensors[tensor_key] = torch.tensor(
                [d[tensor_key] for d in batch], dtype=torch.long
            )
    return batch_tensors


def get_loss_scale_for_deepspeed(model):
    optimizer = model.optimizer
    loss_scale = None
    if hasattr(optimizer, 'loss_scale'):
        loss_scale = optimizer.loss_scale
    elif hasattr(optimizer, 'cur_scale'):
        loss_scale = optimizer.cur_scale
    return loss_scale


class GatherLayer(torch.autograd.Function):
    """
    Gather tensors from all workers with support for backward propagation:
    This implementation does not cut the gradients as torch.distributed.all_gather does.
    """

    @staticmethod
    def forward(ctx, x):
        output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]
        dist.all_gather(output, x)
        return tuple(output)

    @staticmethod
    def backward(ctx, *grads):
        all_gradients = torch.stack(grads)
        dist.all_reduce(all_gradients)
        return all_gradients[dist.get_rank()]


def gather_features(
    image_features,
    text_features,
):
    gathered_image_features = GatherLayer.apply(image_features)
    gathered_text_features = GatherLayer.apply(text_features)
    all_image_features = torch.cat(gathered_image_features)
    all_text_features = torch.cat(gathered_text_features)

    return all_image_features, all_text_features


# The implementation code is modified from open_clip (https://github.com/mlfoundations/open_clip.git)
class ClipLoss(nn.Module):
    def __init__(
        self,
        cache_labels=False,
        rank=0,
        world_size=1,
    ):
        super().__init__()
        self.cache_labels = cache_labels
        self.rank = rank
        self.world_size = world_size

        # cache state
        self.prev_num_logits = 0
        self.labels = {}

    def forward(self, image_features, text_features, logit_scale):
        device = image_features.device
        if self.world_size > 1:
            all_image_features, all_text_features = gather_features(image_features, text_features)

            logits_per_image = logit_scale * image_features @ all_text_features.T
            logits_per_text = logit_scale * text_features @ all_image_features.T
        else:
            logits_per_image = logit_scale * image_features @ text_features.T
            logits_per_text = logit_scale * text_features @ image_features.T

        # calculated ground-truth and cache if enabled
        num_logits = logits_per_image.shape[0]
        if self.prev_num_logits != num_logits or device not in self.labels:
            labels = torch.arange(num_logits, device=device, dtype=torch.long)
            if self.world_size > 1:
                labels = labels + num_logits * self.rank
            if self.cache_labels:
                self.labels[device] = labels
                self.prev_num_logits = num_logits
        else:
            labels = self.labels[device]

        total_loss = (
            F.cross_entropy(logits_per_image, labels) + F.cross_entropy(logits_per_text, labels)
        ) / 2
        return total_loss, logits_per_image, logits_per_text


def write_result_to_jsonl(test_stats, result_file):
    with open(result_file, mode="w", encoding="utf-8") as writer:
        writer.write(json.dumps(test_stats, indent=None))


def read_result_from_jsonl(result_file):
    with open(result_file, encoding="utf-8") as reader:
        return json.load(reader)


class BertCaptioningLoss(nn.Module):
    def __init__(self, label_smoothing, drop_worst_ratio, drop_worst_after):
        super().__init__()
        self.label_smoothing = label_smoothing
        self.drop_worst_ratio = drop_worst_ratio
        self.drop_worst_after = drop_worst_after
        self.log_soft = nn.LogSoftmax(dim=1)
        self.kl = nn.KLDivLoss(reduction='none')
        self.iter = 0

    def forward(self, logits, target, iter):
        eps = self.label_smoothing
        n_class = logits.size(1)
        one_hot = torch.zeros_like(logits).scatter(1, target.view(-1, 1), 1)
        one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
        log_prb = self.log_soft(logits)
        loss = self.kl(log_prb, one_hot).sum(1)

        if self.drop_worst_ratio > 0 and iter > self.drop_worst_after:
            loss, _ = torch.topk(
                loss, k=int(loss.shape[0] * (1 - self.drop_worst_ratio)), largest=False
            )
        loss = loss.mean()

        return loss


class BeamHypotheses:
    def __init__(self, n_hyp, max_length, length_penalty, early_stopping):
        """
        Initialize n-best list of hypotheses.
        """
        self.max_length = max_length - 1  # ignoring bos_token
        self.length_penalty = length_penalty
        self.early_stopping = early_stopping
        self.n_hyp = n_hyp
        self.hyp = []
        self.worst_score = 1e9

    def __len__(self):
        """
        Number of hypotheses in the list.
        """
        return len(self.hyp)

    def add(self, hyp, sum_logprobs):
        """
        Add a new hypothesis to the list.
        """
        score = sum_logprobs / len(hyp) ** self.length_penalty
        if len(self) < self.n_hyp or score > self.worst_score:
            self.hyp.append((score, hyp))
            if len(self) > self.n_hyp:
                sorted_scores = sorted([(s, idx) for idx, (s, _) in enumerate(self.hyp)])
                del self.hyp[sorted_scores[0][1]]
                self.worst_score = sorted_scores[1][0]
            else:
                self.worst_score = min(score, self.worst_score)

    def is_done(self, best_sum_logprobs):
        """
        If there are enough hypotheses and that none of the hypotheses being generated
        can become better than the worst one in the heap, then we are done with this sentence.
        """
        if len(self) < self.n_hyp:
            return False
        elif self.early_stopping:
            return True
        else:
            return self.worst_score >= best_sum_logprobs / self.max_length**self.length_penalty


def dump_predictions(args, result, file_suffix):
    global_rank = get_rank()
    jsons = None
    if global_rank >= 0:
        output_file = os.path.join(args.task_cache_path, f"submit_{global_rank}_{file_suffix}.json")
        with open(output_file, "w") as fp:
            json.dump(result, fp, indent=2)
        torch.distributed.barrier()

        if global_rank == 0:
            world_size = get_world_size()
            jsons = []
            for i in range(world_size):
                each_file = os.path.join(args.task_cache_path, f"submit_{i}_{file_suffix}.json")
                with open(each_file) as fp:
                    jsons += json.load(fp)

            new_jsons = []
            res_dict = dict()
            if args.task in ["coco_captioning", "nocaps"]:
                qid_key = "image_id"
            else:
                # for VQAv2
                qid_key = "question_id"
            for item in jsons:
                if item[qid_key] in res_dict:
                    continue
                new_jsons.append(item)
                res_dict[item[qid_key]] = item
            jsons = new_jsons

        torch.distributed.barrier()
        os.remove(output_file)
    else:
        jsons = result

    result_file = os.path.join(args.output_dir, f"submit_{file_suffix}.json")
    if jsons is not None:
        with open(result_file, "w") as fp:
            json.dump(jsons, fp, indent=2)
        print("Infer %d examples into %s" % (len(jsons), result_file))
    return result_file