YOLOv8-16bit/ultralytics/yolo/utils/torch_utils.py

# Ultralytics YOLO 🚀, GPL-3.0 license

import math
import os
import platform
import random
import time
from contextlib import contextmanager
from copy import deepcopy
from pathlib import Path

import numpy as np
import thop
import torch
import torch.distributed as dist
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.parallel import DistributedDataParallel as DDP

from ultralytics.yolo.utils import DEFAULT_CFG_DICT, DEFAULT_CFG_KEYS, LOGGER, __version__
from ultralytics.yolo.utils.checks import check_version

LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1))  # https://pytorch.org/docs/stable/elastic/run.html
RANK = int(os.getenv('RANK', -1))
WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))

TORCH_1_9 = check_version(torch.__version__, '1.9.0')
TORCH_1_11 = check_version(torch.__version__, '1.11.0')
TORCH_1_12 = check_version(torch.__version__, '1.12.0')


@contextmanager
def torch_distributed_zero_first(local_rank: int):
    # Decorator to make all processes in distributed training wait for each local_master to do something
    initialized = torch.distributed.is_available() and torch.distributed.is_initialized()
    if initialized and local_rank not in {-1, 0}:
        dist.barrier(device_ids=[local_rank])
    yield
    if initialized and local_rank == 0:
        dist.barrier(device_ids=[0])


def smart_inference_mode():
    # Applies torch.inference_mode() decorator if torch>=1.9.0 else torch.no_grad() decorator
    def decorate(fn):
        return (torch.inference_mode if TORCH_1_9 else torch.no_grad)()(fn)

    return decorate


def DDP_model(model):
    # Model DDP creation with checks
    assert not check_version(torch.__version__, '1.12.0', pinned=True), \
        'torch==1.12.0 torchvision==0.13.0 DDP training is not supported due to a known issue. ' \
        'Please upgrade or downgrade torch to use DDP. See https://github.com/ultralytics/yolov5/issues/8395'
    if TORCH_1_11:
        return DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK, static_graph=True)
    else:
        return DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK)


def select_device(device='', batch=0, newline=False, verbose=True):
    # device = None or 'cpu' or 0 or '0' or '0,1,2,3'
    s = f'Ultralytics YOLOv{__version__} 🚀 Python-{platform.python_version()} torch-{torch.__version__} '
    device = str(device).lower()
    for remove in 'cuda:', 'none', '(', ')', '[', ']', "'", ' ':
        device = device.replace(remove, '')  # to string, 'cuda:0' -> '0' and '(0, 1)' -> '0,1'
    cpu = device == 'cpu'
    mps = device == 'mps'  # Apple Metal Performance Shaders (MPS)
    if cpu or mps:
        os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # force torch.cuda.is_available() = False
    elif device:  # non-cpu device requested
        visible = os.environ.get('CUDA_VISIBLE_DEVICES', None)
        os.environ['CUDA_VISIBLE_DEVICES'] = device  # set environment variable - must be before assert is_available()
        if not (torch.cuda.is_available() and torch.cuda.device_count() >= len(device.replace(',', ''))):
            LOGGER.info(s)
            install = 'See https://pytorch.org/get-started/locally/ for up-to-date torch install instructions if no ' \
                      'CUDA devices are seen by torch.\n' if torch.cuda.device_count() == 0 else ''
            raise ValueError(f"Invalid CUDA 'device={device}' requested."
                             f" Use 'device=cpu' or pass valid CUDA device(s) if available,"
                             f" i.e. 'device=0' or 'device=0,1,2,3' for Multi-GPU.\n"
                             f'\ntorch.cuda.is_available(): {torch.cuda.is_available()}'
                             f'\ntorch.cuda.device_count(): {torch.cuda.device_count()}'
                             f"\nos.environ['CUDA_VISIBLE_DEVICES']: {visible}\n"
                             f'{install}')

    if not cpu and not mps and torch.cuda.is_available():  # prefer GPU if available
        devices = device.split(',') if device else '0'  # range(torch.cuda.device_count())  # i.e. 0,1,6,7
        n = len(devices)  # device count
        if n > 1 and batch > 0 and batch % n != 0:  # check batch_size is divisible by device_count
            raise ValueError(f"'batch={batch}' must be a multiple of GPU count {n}. Try 'batch={batch // n * n}' or "
                             f"'batch={batch // n * n + n}', the nearest batch sizes evenly divisible by {n}.")
        space = ' ' * (len(s) + 1)
        for i, d in enumerate(devices):
            p = torch.cuda.get_device_properties(i)
            s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / (1 << 20):.0f}MiB)\n"  # bytes to MB
        arg = 'cuda:0'
    elif mps and getattr(torch, 'has_mps', False) and torch.backends.mps.is_available():  # prefer MPS if available
        s += 'MPS\n'
        arg = 'mps'
    else:  # revert to CPU
        s += 'CPU\n'
        arg = 'cpu'

    if verbose and RANK == -1:
        LOGGER.info(s if newline else s.rstrip())
    return torch.device(arg)


def time_sync():
    # PyTorch-accurate time
    if torch.cuda.is_available():
        torch.cuda.synchronize()
    return time.time()


def fuse_conv_and_bn(conv, bn):
    # Fuse Conv2d() and BatchNorm2d() layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
    fusedconv = nn.Conv2d(conv.in_channels,
                          conv.out_channels,
                          kernel_size=conv.kernel_size,
                          stride=conv.stride,
                          padding=conv.padding,
                          dilation=conv.dilation,
                          groups=conv.groups,
                          bias=True).requires_grad_(False).to(conv.weight.device)

    # Prepare filters
    w_conv = conv.weight.clone().view(conv.out_channels, -1)
    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))

    # Prepare spatial bias
    b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)

    return fusedconv


def fuse_deconv_and_bn(deconv, bn):
    fuseddconv = nn.ConvTranspose2d(deconv.in_channels,
                                    deconv.out_channels,
                                    kernel_size=deconv.kernel_size,
                                    stride=deconv.stride,
                                    padding=deconv.padding,
                                    output_padding=deconv.output_padding,
                                    dilation=deconv.dilation,
                                    groups=deconv.groups,
                                    bias=True).requires_grad_(False).to(deconv.weight.device)

    # prepare filters
    w_deconv = deconv.weight.clone().view(deconv.out_channels, -1)
    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
    fuseddconv.weight.copy_(torch.mm(w_bn, w_deconv).view(fuseddconv.weight.shape))

    # Prepare spatial bias
    b_conv = torch.zeros(deconv.weight.size(1), device=deconv.weight.device) if deconv.bias is None else deconv.bias
    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
    fuseddconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)

    return fuseddconv


def model_info(model, detailed=False, verbose=True, imgsz=640):
    # Model information. imgsz may be int or list, i.e. imgsz=640 or imgsz=[640, 320]
    if not verbose:
        return
    n_p = get_num_params(model)
    n_g = get_num_gradients(model)  # number gradients
    if detailed:
        LOGGER.info(
            f"{'layer':>5} {'name':>40} {'gradient':>9} {'parameters':>12} {'shape':>20} {'mu':>10} {'sigma':>10}")
        for i, (name, p) in enumerate(model.named_parameters()):
            name = name.replace('module_list.', '')
            LOGGER.info('%5g %40s %9s %12g %20s %10.3g %10.3g' %
                        (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))

    flops = get_flops(model, imgsz)
    fused = ' (fused)' if model.is_fused() else ''
    fs = f', {flops:.1f} GFLOPs' if flops else ''
    m = Path(getattr(model, 'yaml_file', '') or model.yaml.get('yaml_file', '')).stem.replace('yolo', 'YOLO') or 'Model'
    LOGGER.info(f'{m} summary{fused}: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}')


def get_num_params(model):
    return sum(x.numel() for x in model.parameters())


def get_num_gradients(model):
    return sum(x.numel() for x in model.parameters() if x.requires_grad)


def get_flops(model, imgsz=640):
    try:
        model = de_parallel(model)
        p = next(model.parameters())
        stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32  # max stride
        im = torch.empty((1, p.shape[1], stride, stride), device=p.device)  # input image in BCHW format
        flops = thop.profile(deepcopy(model), inputs=[im], verbose=False)[0] / 1E9 * 2  # stride GFLOPs
        imgsz = imgsz if isinstance(imgsz, list) else [imgsz, imgsz]  # expand if int/float
        flops = flops * imgsz[0] / stride * imgsz[1] / stride  # 640x640 GFLOPs
        return flops
    except Exception:
        return 0


def initialize_weights(model):
    for m in model.modules():
        t = type(m)
        if t is nn.Conv2d:
            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
        elif t is nn.BatchNorm2d:
            m.eps = 1e-3
            m.momentum = 0.03
        elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
            m.inplace = True


def scale_img(img, ratio=1.0, same_shape=False, gs=32):  # img(16,3,256,416)
    # Scales img(bs,3,y,x) by ratio constrained to gs-multiple
    if ratio == 1.0:
        return img
    h, w = img.shape[2:]
    s = (int(h * ratio), int(w * ratio))  # new size
    img = F.interpolate(img, size=s, mode='bilinear', align_corners=False)  # resize
    if not same_shape:  # pad/crop img
        h, w = (math.ceil(x * ratio / gs) * gs for x in (h, w))
    return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean


def make_divisible(x, divisor):
    # Returns nearest x divisible by divisor
    if isinstance(divisor, torch.Tensor):
        divisor = int(divisor.max())  # to int
    return math.ceil(x / divisor) * divisor


def copy_attr(a, b, include=(), exclude=()):
    # Copy attributes from b to a, options to only include [...] and to exclude [...]
    for k, v in b.__dict__.items():
        if (len(include) and k not in include) or k.startswith('_') or k in exclude:
            continue
        else:
            setattr(a, k, v)


def get_latest_opset():
    # Return max supported ONNX opset by this version of torch
    return max(int(k[14:]) for k in vars(torch.onnx) if 'symbolic_opset' in k)  # opset


def intersect_dicts(da, db, exclude=()):
    # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values
    return {k: v for k, v in da.items() if k in db and all(x not in k for x in exclude) and v.shape == db[k].shape}


def is_parallel(model):
    # Returns True if model is of type DP or DDP
    return isinstance(model, (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel))


def de_parallel(model):
    # De-parallelize a model: returns single-GPU model if model is of type DP or DDP
    return model.module if is_parallel(model) else model


def one_cycle(y1=0.0, y2=1.0, steps=100):
    # lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf
    return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1


def init_seeds(seed=0, deterministic=False):
    # Initialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)  # for Multi-GPU, exception safe
    # torch.backends.cudnn.benchmark = True  # AutoBatch problem https://github.com/ultralytics/yolov5/issues/9287
    if deterministic and TORCH_1_12:  # https://github.com/ultralytics/yolov5/pull/8213
        torch.use_deterministic_algorithms(True)
        torch.backends.cudnn.deterministic = True
        os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8'
        os.environ['PYTHONHASHSEED'] = str(seed)


class ModelEMA:
    """ Updated Exponential Moving Average (EMA) from https://github.com/rwightman/pytorch-image-models
    Keeps a moving average of everything in the model state_dict (parameters and buffers)
    For EMA details see https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage
    To disable EMA set the `enabled` attribute to `False`.
    """

    def __init__(self, model, decay=0.9999, tau=2000, updates=0):
        # Create EMA
        self.ema = deepcopy(de_parallel(model)).eval()  # FP32 EMA
        self.updates = updates  # number of EMA updates
        self.decay = lambda x: decay * (1 - math.exp(-x / tau))  # decay exponential ramp (to help early epochs)
        for p in self.ema.parameters():
            p.requires_grad_(False)
        self.enabled = True

    def update(self, model):
        # Update EMA parameters
        if self.enabled:
            self.updates += 1
            d = self.decay(self.updates)

            msd = de_parallel(model).state_dict()  # model state_dict
            for k, v in self.ema.state_dict().items():
                if v.dtype.is_floating_point:  # true for FP16 and FP32
                    v *= d
                    v += (1 - d) * msd[k].detach()
                    # assert v.dtype == msd[k].dtype == torch.float32, f'{k}: EMA {v.dtype},  model {msd[k].dtype}'

    def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):
        # Update EMA attributes
        if self.enabled:
            copy_attr(self.ema, model, include, exclude)


def strip_optimizer(f='best.pt', s=''):
    """
    Strip optimizer from 'f' to finalize training, optionally save as 's'.

    Usage:
        from ultralytics.yolo.utils.torch_utils import strip_optimizer
        from pathlib import Path
        for f in Path('/Users/glennjocher/Downloads/weights').glob('*.pt'):
            strip_optimizer(f)

    Args:
        f (str): file path to model to strip the optimizer from. Default is 'best.pt'.
        s (str): file path to save the model with stripped optimizer to. If not provided, 'f' will be overwritten.

    Returns:
        None
    """
    x = torch.load(f, map_location=torch.device('cpu'))
    args = {**DEFAULT_CFG_DICT, **x['train_args']}  # combine model args with default args, preferring model args
    if x.get('ema'):
        x['model'] = x['ema']  # replace model with ema
    for k in 'optimizer', 'best_fitness', 'ema', 'updates':  # keys
        x[k] = None
    x['epoch'] = -1
    x['model'].half()  # to FP16
    for p in x['model'].parameters():
        p.requires_grad = False
    x['train_args'] = {k: v for k, v in args.items() if k in DEFAULT_CFG_KEYS}  # strip non-default keys
    # x['model'].args = x['train_args']
    torch.save(x, s or f)
    mb = os.path.getsize(s or f) / 1E6  # filesize
    LOGGER.info(f"Optimizer stripped from {f},{f' saved as {s},' if s else ''} {mb:.1f}MB")


def profile(input, ops, n=10, device=None):
    """ YOLOv8 speed/memory/FLOPs profiler
    Usage:
        input = torch.randn(16, 3, 640, 640)
        m1 = lambda x: x * torch.sigmoid(x)
        m2 = nn.SiLU()
        profile(input, [m1, m2], n=100)  # profile over 100 iterations
    """
    results = []
    if not isinstance(device, torch.device):
        device = select_device(device)
    LOGGER.info(f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}"
                f"{'input':>24s}{'output':>24s}")

    for x in input if isinstance(input, list) else [input]:
        x = x.to(device)
        x.requires_grad = True
        for m in ops if isinstance(ops, list) else [ops]:
            m = m.to(device) if hasattr(m, 'to') else m  # device
            m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m
            tf, tb, t = 0, 0, [0, 0, 0]  # dt forward, backward
            try:
                flops = thop.profile(m, inputs=[x], verbose=False)[0] / 1E9 * 2  # GFLOPs
            except Exception:
                flops = 0

            try:
                for _ in range(n):
                    t[0] = time_sync()
                    y = m(x)
                    t[1] = time_sync()
                    try:
                        _ = (sum(yi.sum() for yi in y) if isinstance(y, list) else y).sum().backward()
                        t[2] = time_sync()
                    except Exception:  # no backward method
                        # print(e)  # for debug
                        t[2] = float('nan')
                    tf += (t[1] - t[0]) * 1000 / n  # ms per op forward
                    tb += (t[2] - t[1]) * 1000 / n  # ms per op backward
                mem = torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0  # (GB)
                s_in, s_out = (tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' for x in (x, y))  # shapes
                p = sum(x.numel() for x in m.parameters()) if isinstance(m, nn.Module) else 0  # parameters
                LOGGER.info(f'{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}')
                results.append([p, flops, mem, tf, tb, s_in, s_out])
            except Exception as e:
                LOGGER.info(e)
                results.append(None)
            torch.cuda.empty_cache()
    return results


class EarlyStopping:
    """
    Early stopping class that stops training when a specified number of epochs have passed without improvement.
    """

    def __init__(self, patience=50):
        """
        Initialize early stopping object

        Args:
            patience (int, optional): Number of epochs to wait after fitness stops improving before stopping.
        """
        self.best_fitness = 0.0  # i.e. mAP
        self.best_epoch = 0
        self.patience = patience or float('inf')  # epochs to wait after fitness stops improving to stop
        self.possible_stop = False  # possible stop may occur next epoch

    def __call__(self, epoch, fitness):
        """
        Check whether to stop training

        Args:
            epoch (int): Current epoch of training
            fitness (float): Fitness value of current epoch

        Returns:
            bool: True if training should stop, False otherwise
        """
        if fitness is None:  # check if fitness=None (happens when val=False)
            return False

        if fitness >= self.best_fitness:  # >= 0 to allow for early zero-fitness stage of training
            self.best_epoch = epoch
            self.best_fitness = fitness
        delta = epoch - self.best_epoch  # epochs without improvement
        self.possible_stop = delta >= (self.patience - 1)  # possible stop may occur next epoch
        stop = delta >= self.patience  # stop training if patience exceeded
        if stop:
            LOGGER.info(f'Stopping training early as no improvement observed in last {self.patience} epochs. '
                        f'Best results observed at epoch {self.best_epoch}, best model saved as best.pt.\n'
                        f'To update EarlyStopping(patience={self.patience}) pass a new patience value, '
                        f'i.e. `patience=300` or use `patience=0` to disable EarlyStopping.')
        return stop