util/misc.py

"""
Runtime helpers: distributed setup, metric logging, AMP loss scaling, and
checkpoint I/O.

Written for STP. The distributed-helper and metric-logger patterns are the
conventional ones for PyTorch DDP training scripts; the checkpoint format
matches what the released STP weights were saved with, so `--resume` works on
them directly.
"""
import datetime
import os
import time
from collections import defaultdict, deque
from pathlib import Path

import torch
import torch.distributed as dist


# ---------------------------------------------------------------------------
# metric logging
# ---------------------------------------------------------------------------
class SmoothedValue:
    """Running mean (and window mean) of a scalar."""

    def __init__(self, window_size=20, fmt=None):
        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):
        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:
    """Prints and averages a set of named meters, optionally across ranks."""

    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):
        parts = [f'{name}: {str(meter)}' for name, meter in self.meters.items()]
        return self.delimiter.join(parts)

    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(f'{header} Total time: {total_time_str} ({total_time / len(iterable):.4f} s / it)')


# ---------------------------------------------------------------------------
# distributed
# ---------------------------------------------------------------------------
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 init_distributed_mode(args):
    """Fill in args.distributed / rank / world_size / gpu from the launcher env.

    Works with `torchrun` (env:// rendezvous), SLURM, or single-process runs.
    """
    if '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.get('LOCAL_RANK', 0))
    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}', flush=True)
    torch.distributed.init_process_group(
        backend=args.dist_backend, init_method=args.dist_url,
        world_size=args.world_size, rank=args.rank)
    torch.distributed.barrier()
    setup_for_distributed(args.rank == 0)


def setup_for_distributed(is_master):
    """Silence non-master ranks' stdout, including inside tqdm."""
    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 all_reduce_mean(x):
    world_size = get_world_size()
    if world_size > 1:
        x_reduce = torch.tensor(x).cuda()
        dist.all_reduce(x_reduce)
        x_reduce /= world_size
        return x_reduce.item()
    return x


# ---------------------------------------------------------------------------
# mixed-precision loss scaling
# ---------------------------------------------------------------------------
class NativeScalerWithGradNormCount:
    """amp.GradScaler that also reports the (unscaled) gradient norm.

    Mirrors the interface of the MAE/DeiT training scripts so the training loop
    reads the same way, and so `scaler` round-trips through checkpoints.
    """

    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)
                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.)
    device = parameters[0].grad.device
    if norm_type == float('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)
    if torch.isnan(total_norm) or torch.isinf(total_norm):
        total_norm = torch.tensor(0., device=device)
    return total_norm


# ---------------------------------------------------------------------------
# checkpoints
# ---------------------------------------------------------------------------
def save_model(args, epoch, model, model_without_ddp, optimizer, loss_scaler):
    """Write `checkpoint-<epoch>.pth` holding {model, optimizer, epoch, scaler, args}.

    Same layout as the released `0.95_aug0.8.pth`, so those weights
    can be loaded with `load_model(--resume ...)` and training continued.
    """
    output_dir = Path(args.output_dir)
    checkpoint_path = output_dir / f'checkpoint-{epoch}.pth'
    to_save = {
        'model': model_without_ddp.state_dict(),
        'optimizer': optimizer.state_dict(),
        'epoch': epoch,
        'scaler': loss_scaler.state_dict() if loss_scaler is not None else None,
        'args': args,
    }
    save_on_master(to_save, checkpoint_path)


def load_model(args, model_without_ddp, optimizer, loss_scaler):
    if not args.resume:
        return
    if args.resume.startswith('https'):
        checkpoint = torch.hub.load_state_dict_from_url(args.resume, map_location='cpu', check_hash=True)
    else:
        from util.ckpt_io import load_checkpoint
        checkpoint = load_checkpoint(args.resume)

    model_without_ddp.load_state_dict(checkpoint['model'])
    print(f'Resume checkpoint {args.resume}')

    has_opt = 'optimizer' in checkpoint and 'epoch' in checkpoint
    if has_opt and not getattr(args, 'eval', False):
        optimizer.load_state_dict(checkpoint['optimizer'])
        args.start_epoch = checkpoint['epoch'] + 1
        if 'scaler' in checkpoint and checkpoint['scaler'] is not None and loss_scaler is not None:
            loss_scaler.load_state_dict(checkpoint['scaler'])
        print('With optim & sched!')