# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
"""
PyTorch utils
"""
import math
import os
import platform # 提供获取操作系统相关信息的模块
import subprocess # 子进程定义及操作的模块
import time # 时间模块 更底层
import warnings
from contextlib import contextmanager
from copy import deepcopy # 实现深度复制的模块
from pathlib import Path # Path将str转换为Path对象 使字符串路径易于操作的模块
# 以下是一些基本的torch相关的类
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 utils.general import LOGGER, check_version, colorstr, file_date, git_describe
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))
try:
import thop # for FLOPs computation # 用于Pytorch模型的FLOPS计算工具模块
except ImportError:
thop = None
# Suppress PyTorch warnings
warnings.filterwarnings('ignore', message='User provided device_type of \'cuda\', but CUDA is not available. Disabling')
def smart_DDP(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 check_version(torch.__version__, '1.11.0'):
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)
@contextmanager
def torch_distributed_zero_first(local_rank: int):
"""train.py
用于处理模型进行分布式训练时同步问题
基于torch.distributed.barrier()函数的上下文管理器，为了完成数据的正常同步操作（yolov5中拥有大量的多线程并行操作）
Decorator to make all processes in distributed training wait for each local_master to do something.
:params local_rank: 代表当前进程号 0代表主进程 1、2、3代表子进程
"""
# Decorator to make all processes in distributed training wait for each local_master to do something
if local_rank not in [-1, 0]:
# 如果执行create_dataloader()函数的进程不是主进程，即rank不等于0或者-1，
# 上下文管理器会执行相应的torch.distributed.barrier()，设置一个阻塞栅栏，
# 让此进程处于等待状态，等待所有进程到达栅栏处（包括主进程数据处理完毕）；
dist.barrier(device_ids=[local_rank])
yield
if local_rank == 0:
# 如果执行create_dataloader()函数的进程是主进程，其会直接去读取数据并处理，
# 然后其处理结束之后会接着遇到torch.distributed.barrier()，
# 此时，所有进程都到达了当前的栅栏处，这样所有进程就达到了同步，并同时得到释放。
dist.barrier(device_ids=[0])
def device_count():
# Returns number of CUDA devices available. Safe version of torch.cuda.device_count(). Supports Linux and Windows
assert platform.system() in ('Linux', 'Windows'), 'device_count() only supported on Linux or Windows'
try:
cmd = 'nvidia-smi -L | wc -l' if platform.system() == 'Linux' else 'nvidia-smi -L | find /c /v ""' # Windows
return int(subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1])
except Exception:
return 0
# 完成自动选择系统设备的操作，在select_device函数中会调用git_describe函数和date_modified函数。
def select_device(device='', batch_size=0, newline=True):
# device = None or 'cpu' or 0 or '0' or '0,1,2,3'
s = f'YOLOv5 🚀 {git_describe() or file_date()} Python-{platform.python_version()} torch-{torch.__version__} '
device = str(device).strip().lower().replace('cuda:', '').replace('none', '') # to string, 'cuda:0' to '0'
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
os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable - must be before assert is_available()
assert torch.cuda.is_available() and torch.cuda.device_count() >= len(device.replace(',', '')), \
f"Invalid CUDA '--device {device}' requested, use '--device cpu' or pass valid CUDA device(s)"
if not (cpu or 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_size > 0: # check batch_size is divisible by device_count
assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {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 not newline:
s = s.rstrip()
LOGGER.info(s)
return torch.device(arg)
# 精确的获取当前时间
def time_sync():
# PyTorch-accurate time
if torch.cuda.is_available():
torch.cuda.synchronize()
return time.time()
# 主要用于输出模型的一些信息，如所有层数量, 模型总参数量等。
def profile(input, ops, n=10, device=None):
# YOLOv5 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)
print(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
print(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:
print(e)
results.append(None)
torch.cuda.empty_cache()
return results
# 用于判断模型是否支持并行 Returns True if model is of type DP or DDP
def is_parallel(model):
# Returns True if model is of type DP or DDP
return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel)
# 判断单卡还是多卡(能否并行) 多卡返回model.module 单卡返回model
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
# 函数是用来初始化模型权重的，会在yolo.py的Model类中的init函数被调用，如下：
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 find_modules(model, mclass=nn.Conv2d):
# Finds layer indices matching module class 'mclass'
return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)]
# 矩阵的行压缩存储
def sparsity(model):
# Return global model sparsity
a, b = 0, 0
for p in model.parameters():
a += p.numel()
b += (p == 0).sum()
return b / a
def prune(model, amount=0.3):
# Prune model to requested global sparsity
import torch.nn.utils.prune as prune
print('Pruning model... ', end='')
for name, m in model.named_modules():
if isinstance(m, nn.Conv2d):
prune.l1_unstructured(m, name='weight', amount=amount) # prune
prune.remove(m, 'weight') # make permanent
print(' %.3g global sparsity' % sparsity(model))
# 函数增强
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,
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
# 模型的信息
def model_info(model, verbose=False, img_size=640):
"""用于yolo.py文件的Model类的info函数
输出模型的所有信息 包括: 所有层数量, 模型总参数量, 需要求梯度的总参数量, img_size大小的model的浮点计算量GFLOPs
:params model: 模型
:params verbose: 是否输出每一层的参数parameters的相关信息
:params img_size: int or list i.e. img_size=640 or img_size=[640, 320]
"""
# Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320]
n_p = sum(x.numel() for x in model.parameters()) # number parameters
n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients
if verbose:
print(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.', '')
print('%5g %40s %9s %12g %20s %10.3g %10.3g' %
(i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std()))
try: # FLOPs
from thop import profile
stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32
img = torch.zeros((1, model.yaml.get('ch', 3), stride, stride), device=next(model.parameters()).device) # input
flops = profile(deepcopy(model), inputs=(img,), verbose=False)[0] / 1E9 * 2 # stride GFLOPs
img_size = img_size if isinstance(img_size, list) else [img_size, img_size] # expand if int/float
fs = ', %.1f GFLOPs' % (flops * img_size[0] / stride * img_size[1] / stride) # 640x640 GFLOPs
except Exception:
fs = ''
name = Path(model.yaml_file).stem.replace('yolov5', 'YOLOv5') if hasattr(model, 'yaml_file') else 'Model'
LOGGER.info(f"{name} summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}")
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
# 用于模型复制
def copy_attr(a, b, include=(), exclude=()):
"""在ModelEMA函数和yolo.py中Model类的autoshape函数中调用
复制b的属性(这个属性必须在include中而不在exclude中)给a
:params a: 对象a(待赋值)
:params b: 对象b(赋值)
:params include: 可以赋值的属性
:params exclude: 不能赋值的属性
"""
# 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)
# 选择优化器
def smart_optimizer(model, name='Adam', lr=0.001, momentum=0.9, decay=1e-5):
# YOLOv5 3-param group optimizer: 0) weights with decay, 1) weights no decay, 2) biases no decay
g = [], [], [] # optimizer parameter groups
bn = tuple(v for k, v in nn.__dict__.items() if 'Norm' in k) # normalization layers, i.e. BatchNorm2d()
for v in model.modules():
if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias (no decay)
g[2].append(v.bias)
if isinstance(v, bn): # weight (no decay)
g[1].append(v.weight)
elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): # weight (with decay)
g[0].append(v.weight)
if name == 'Adam':
optimizer = torch.optim.Adam(g[2], lr=lr, betas=(momentum, 0.999)) # adjust beta1 to momentum
elif name == 'AdamW':
optimizer = torch.optim.AdamW(g[2], lr=lr, betas=(momentum, 0.999), weight_decay=0.0)
elif name == 'RMSProp':
optimizer = torch.optim.RMSprop(g[2], lr=lr, momentum=momentum)
elif name == 'SGD':
optimizer = torch.optim.SGD(g[2], lr=lr, momentum=momentum, nesterov=True)
else:
raise NotImplementedError(f'Optimizer {name} not implemented.')
optimizer.add_param_group({'params': g[0], 'weight_decay': decay}) # add g0 with weight_decay
optimizer.add_param_group({'params': g[1], 'weight_decay': 0.0}) # add g1 (BatchNorm2d weights)
LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__}(lr={lr}) with parameter groups "
f"{len(g[1])} weight(decay=0.0), {len(g[0])} weight(decay={decay}), {len(g[2])} bias")
return optimizer
# 断点恢复
def smart_resume(ckpt, optimizer, ema=None, weights='yolov5s.pt', epochs=300, resume=True):
# Resume training from a partially trained checkpoint
best_fitness = 0.0
start_epoch = ckpt['epoch'] + 1
if ckpt['optimizer'] is not None:
optimizer.load_state_dict(ckpt['optimizer']) # optimizer
best_fitness = ckpt['best_fitness']
if ema and ckpt.get('ema'):
ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) # EMA
ema.updates = ckpt['updates']
if resume:
assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.\n' \
f"Start a new training without --resume, i.e. 'python train.py --weights {weights}'"
LOGGER.info(f'Resuming training from {weights} from epoch {start_epoch} to {epochs} total epochs')
if epochs < start_epoch:
LOGGER.info(f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs.")
epochs += ckpt['epoch'] # finetune additional epochs
return best_fitness, start_epoch, epochs
class EarlyStopping:
# YOLOv5 simple early stopper
def __init__(self, patience=30):
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):
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. `python train.py --patience 300` or use `--patience 0` to disable EarlyStopping.')
return stop
# 模型ema
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
"""
def __init__(self, model, decay=0.9999, tau=2000, updates=0):
# Create EMA 创建ema模型
self.ema = deepcopy(de_parallel(model)).eval() # FP32 EMA
# if next(model.parameters()).device.type != 'cpu':
# self.ema.half() # FP16 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)
# 所有参数取消设置梯度(测试 model.val)
for p in self.ema.parameters():
p.requires_grad_(False)
def update(self, model):
# Update EMA parameters
with torch.no_grad():
self.updates += 1 # ema更新次数 + 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:
v *= d
v += (1 - d) * msd[k].detach()
def update_attr(self, model, include=(), exclude=('process_group', 'reducer')):
# Update EMA attributes
copy_attr(self.ema, model, include, exclude)
