记录一下，非常不错的train模板。通过这个模板可以学到很多。
链接：https://github.com/huawei-noah/Efficient-AI-Backbones/blob/master/augvit_pytorch/train.py
# 2022.02.14-Changed for main script for wavemlp model
# Huawei Technologies Co., Ltd. <foss@huawei.com>
#!/usr/bin/env python
""" ImageNet Training Script
This is intended to be a lean and easily modifiable ImageNet training script that reproduces ImageNet
training results with some of the latest networks and training techniques. It favours canonical PyTorch
and standard Python style over trying to be able to 'do it all.' That said, it offers quite a few speed
and training result improvements over the usual PyTorch example scripts. Repurpose as you see fit.
This script was started from an early version of the PyTorch ImageNet example
(https://github.com/pytorch/examples/tree/master/imagenet)
NVIDIA CUDA specific speedups adopted from NVIDIA Apex examples
(https://github.com/NVIDIA/apex/tree/master/examples/imagenet)
Hacked together by / Copyright 2020 Ross Wightman (https://github.com/rwightman)
"""
import warnings
warnings.filterwarnings('ignore')
import argparse
import time
import yaml
import os
import logging
from collections import OrderedDict
from contextlib import suppress
from datetime import datetime
import math
import torch
import torch.nn as nn
import torchvision.utils
from torch.nn.parallel import DistributedDataParallel as NativeDDP
from timm.data import Dataset, resolve_data_config, Mixup, FastCollateMixup, AugMixDataset
from timm.models import create_model, resume_checkpoint, convert_splitbn_model
from timm.utils import *
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy, JsdCrossEntropy
from timm.optim import create_optimizer
from timm.scheduler import create_scheduler
from timm.utils import ApexScaler, NativeScaler
from data.myloader import create_loader
import models.wavemlp
try:
from apex import amp
from apex.parallel import DistributedDataParallel as ApexDDP
from apex.parallel import convert_syncbn_model
has_apex = True
except ImportError:
has_apex = False
has_native_amp = False
try:
if getattr(torch.cuda.amp, 'autocast') is not None:
has_native_amp = True
except AttributeError:
pass
torch.backends.cudnn.benchmark = True
_logger = logging.getLogger('train')
# The first arg parser parses out only the --config argument, this argument is used to
# load a yaml file containing key-values that override the defaults for the main parser below
config_parser = parser = argparse.ArgumentParser(description='Training Config', add_help=False)
parser.add_argument('-c', '--config', default='', type=str, metavar='FILE',
help='YAML config file specifying default arguments')
parser = argparse.ArgumentParser(description='PyTorch ImageNet Training')
# Dataset / Model parameters
parser.add_argument('data', metavar='DIR',
help='path to dataset')
parser.add_argument('--model', default='resnet101', type=str, metavar='MODEL',
help='Name of model to train (default: "countception"')
parser.add_argument('--pretrained', action='store_true', default=False,
help='Start with pretrained version of specified network (if avail)')
parser.add_argument('--initial-checkpoint', default='', type=str, metavar='PATH',
help='Initialize model from this checkpoint (default: none)')
parser.add_argument('--resume', default='', type=str, metavar='PATH',
help='Resume full model and optimizer state from checkpoint (default: none)')
parser.add_argument('--no-resume-opt', action='store_true', default=False,
help='prevent resume of optimizer state when resuming model')
parser.add_argument('--num-classes', type=int, default=1000, metavar='N',
help='number of label classes (default: 1000)')
parser.add_argument('--gp', default=None, type=str, metavar='POOL',
help='Global pool type, one of (fast, avg, max, avgmax, avgmaxc). Model default if None.')
parser.add_argument('--img-size', type=int, default=None, metavar='N',
help='Image patch size (default: None => model default)')
parser.add_argument('--crop-pct', default=None, type=float,
metavar='N', help='Input image center crop percent (for validation only)')
parser.add_argument('--mean', type=float, nargs='+', default=None, metavar='MEAN',
help='Override mean pixel value of dataset')
parser.add_argument('--std', type=float, nargs='+', default=None, metavar='STD',
help='Override std deviation of of dataset')
parser.add_argument('--interpolation', default='', type=str, metavar='NAME',
help='Image resize interpolation type (overrides model)')
parser.add_argument('-b', '--batch-size', type=int, default=32, metavar='N',
help='input batch size for training (default: 32)')
parser.add_argument('-vb', '--validation-batch-size-multiplier', type=int, default=1, metavar='N',
help='ratio of validation batch size to training batch size (default: 1)')
# Optimizer parameters
parser.add_argument('--opt', default='sgd', type=str, metavar='OPTIMIZER',
help='Optimizer (default: "sgd"')
parser.add_argument('--opt-eps', default=None, type=float, metavar='EPSILON',
help='Optimizer Epsilon (default: None, use opt default)')
parser.add_argument('--opt-betas', default=None, type=float, nargs='+', metavar='BETA',
help='Optimizer Betas (default: None, use opt default)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='Optimizer momentum (default: 0.9)')
parser.add_argument('--weight-decay', type=float, default=0.0001,
help='weight decay (default: 0.0001)')
parser.add_argument('--clip-grad', type=float, default=None, metavar='NORM',
help='Clip gradient norm (default: None, no clipping)')
# Learning rate schedule parameters
parser.add_argument('--sched', default='step', type=str, metavar='SCHEDULER',
help='LR scheduler (default: "step"')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--lr-noise', type=float, nargs='+', default=None, metavar='pct, pct',
help='learning rate noise on/off epoch percentages')
parser.add_argument('--lr-noise-pct', type=float, default=0.67, metavar='PERCENT',
help='learning rate noise limit percent (default: 0.67)')
parser.add_argument('--lr-noise-std', type=float, default=1.0, metavar='STDDEV',
help='learning rate noise std-dev (default: 1.0)')
parser.add_argument('--lr-cycle-mul', type=float, default=1.0, metavar='MULT',
help='learning rate cycle len multiplier (default: 1.0)')
parser.add_argument('--lr-cycle-limit', type=int, default=1, metavar='N',
help='learning rate cycle limit')
parser.add_argument('--warmup-lr', type=float, default=0.0001, metavar='LR',
help='warmup learning rate (default: 0.0001)')
parser.add_argument('--min-lr', type=float, default=1e-5, metavar='LR',
help='lower lr bound for cyclic schedulers that hit 0 (1e-5)')
parser.add_argument('--epochs', type=int, default=200, metavar='N',
help='number of epochs to train (default: 2)')
parser.add_argument('--start-epoch', default=None, type=int, metavar='N',
help='manual epoch number (useful on restarts)')
parser.add_argument('--decay-epochs', type=float, default=30, metavar='N',
help='epoch interval to decay LR')
parser.add_argument('--warmup-epochs', type=int, default=3, metavar='N',
help='epochs to warmup LR, if scheduler supports')
parser.add_argument('--cooldown-epochs', type=int, default=10, metavar='N',
help='epochs to cooldown LR at min_lr, after cyclic schedule ends')
parser.add_argument('--patience-epochs', type=int, default=10, metavar='N',
help='patience epochs for Plateau LR scheduler (default: 10')
parser.add_argument('--decay-rate', '--dr', type=float, default=0.1, metavar='RATE',
help='LR decay rate (default: 0.1)')
# Augmentation & regularization parameters
parser.add_argument('--no-aug', action='store_true', default=False,
help='Disable all training augmentation, override other train aug args')
parser.add_argument('--repeated-aug', action='store_true')
parser.add_argument('--scale', type=float, nargs='+', default=[0.08, 1.0], metavar='PCT',
help='Random resize scale (default: 0.08 1.0)')
parser.add_argument('--ratio', type=float, nargs='+', default=[3./4., 4./3.], metavar='RATIO',
help='Random resize aspect ratio (default: 0.75 1.33)')
parser.add_argument('--hflip', type=float, default=0.5,
help='Horizontal flip training aug probability')
parser.add_argument('--vflip', type=float, default=0.,
help='Vertical flip training aug probability')
parser.add_argument('--color-jitter', type=float, default=0.4, metavar='PCT',
help='Color jitter factor (default: 0.4)')
parser.add_argument('--aa', type=str, default=None, metavar='NAME',
help='Use AutoAugment policy. "v0" or "original". (default: None)'),
parser.add_argument('--aug-splits', type=int, default=0,
help='Number of augmentation splits (default: 0, valid: 0 or >=2)')
parser.add_argument('--jsd', action='store_true', default=False,
help='Enable Jensen-Shannon Divergence + CE loss. Use with `--aug-splits`.')
parser.add_argument('--reprob', type=float, default=0., metavar='PCT',
help='Random erase prob (default: 0.)')
parser.add_argument('--remode', type=str, default='const',
help='Random erase mode (default: "const")')
parser.add_argument('--recount', type=int, default=1,
help='Random erase count (default: 1)')
parser.add_argument('--resplit', action='store_true', default=False,
help='Do not random erase first (clean) augmentation split')
parser.add_argument('--mixup', type=float, default=0.0,
help='mixup alpha, mixup enabled if > 0. (default: 0.)')
parser.add_argument('--cutmix', type=float, default=0.0,
help='cutmix alpha, cutmix enabled if > 0. (default: 0.)')
parser.add_argument('--cutmix-minmax', type=float, nargs='+', default=None,
help='cutmix min/max ratio, overrides alpha and enables cutmix if set (default: None)')
parser.add_argument('--mixup-prob', type=float, default=1.0,
help='Probability of performing mixup or cutmix when either/both is enabled')
parser.add_argument('--mixup-switch-prob', type=float, default=0.5,
help='Probability of switching to cutmix when both mixup and cutmix enabled')
parser.add_argument('--mixup-mode', type=str, default='batch',
help='How to apply mixup/cutmix params. Per "batch", "pair", or "elem"')
parser.add_argument('--mixup-off-epoch', default=0, type=int, metavar='N',
help='Turn off mixup after this epoch, disabled if 0 (default: 0)')
parser.add_argument('--smoothing', type=float, default=0.1,
help='Label smoothing (default: 0.1)')
parser.add_argument('--train-interpolation', type=str, default='random',
help='Training interpolation (random, bilinear, bicubic default: "random")')
parser.add_argument('--drop', type=float, default=0.0, metavar='PCT',
help='Dropout rate (default: 0.)')
parser.add_argument('--drop-connect', type=float, default=None, metavar='PCT',
help='Drop connect rate, DEPRECATED, use drop-path (default: None)')
parser.add_argument('--drop-path', type=float, default=None, metavar='PCT',
help='Drop path rate (default: None)')
parser.add_argument('--drop-block', type=float, default=None, metavar='PCT',
help='Drop block rate (default: None)')
# Batch norm parameters (only works with gen_efficientnet based models currently)
parser.add_argument('--bn-tf', action='store_true', default=False,
help='Use Tensorflow BatchNorm defaults for models that support it (default: False)')
parser.add_argument('--bn-momentum', type=float, default=None,
help='BatchNorm momentum override (if not None)')
parser.add_argument('--bn-eps', type=float, default=None,
help='BatchNorm epsilon override (if not None)')
parser.add_argument('--sync-bn', action='store_true',
help='Enable NVIDIA Apex or Torch synchronized BatchNorm.')
parser.add_argument('--dist-bn', type=str, default='',
help='Distribute BatchNorm stats between nodes after each epoch ("broadcast", "reduce", or "")')
parser.add_argument('--split-bn', action='store_true',
help='Enable separate BN layers per augmentation split.')
# Model Exponential Moving Average
parser.add_argument('--model-ema', action='store_true', default=False,
help='Enable tracking moving average of model weights')
parser.add_argument('--model-ema-force-cpu', action='store_true', default=False,
help='Force ema to be tracked on CPU, rank=0 node only. Disables EMA validation.')
parser.add_argument('--model-ema-decay', type=float, default=0.9998,
help='decay factor for model weights moving average (default: 0.9998)')
# Misc
parser.add_argument('--seed', type=int, default=42, metavar='S',
help='random seed (default: 42)')
parser.add_argument('--log-interval', type=int, default=50, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--recovery-interval', type=int, default=0, metavar='N',
help='how many batches to wait before writing recovery checkpoint')
parser.add_argument('-j', '--workers', type=int, default=4, metavar='N',
help='how many training processes to use (default: 1)')
parser.add_argument('--num-gpu', type=int, default=1,
help='Number of GPUS to use')
parser.add_argument('--save-images', action='store_true', default=False,
help='save images of input bathes every log interval for debugging')
parser.add_argument('--amp', action='store_true', default=False,
help='use NVIDIA Apex AMP or Native AMP for mixed precision training')
parser.add_argument('--apex-amp', action='store_true', default=False,
help='Use NVIDIA Apex AMP mixed precision')
parser.add_argument('--native-amp', action='store_true', default=False,
help='Use Native Torch AMP mixed precision')
parser.add_argument('--channels-last', action='store_true', default=False,
help='Use channels_last memory layout')
parser.add_argument('--pin-mem', action='store_true', default=False,
help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
parser.add_argument('--no-prefetcher', action='store_true', default=False,
help='disable fast prefetcher')
parser.add_argument('--output', default='', type=str, metavar='PATH',
help='path to output folder (default: none, current dir)')
parser.add_argument('--eval-metric', default='top1', type=str, metavar='EVAL_METRIC',
help='Best metric (default: "top1"')
parser.add_argument('--tta', type=int, default=0, metavar='N',
help='Test/inference time augmentation (oversampling) factor. 0=None (default: 0)')
parser.add_argument("--local_rank", default=0, type=int)
parser.add_argument('--use-multi-epochs-loader', action='store_true', default=False,
help='use the multi-epochs-loader to save time at the beginning of every epoch')
parser.add_argument("--init_method", default='env://', type=str)
def _parse_args():
# Do we have a config file to parse?
args_config, remaining = config_parser.parse_known_args()
if args_config.config:
with open(args_config.config, 'r') as f:
cfg = yaml.safe_load(f)
parser.set_defaults(**cfg)
# The main arg parser parses the rest of the args, the usual
# defaults will have been overridden if config file specified.
args = parser.parse_args(remaining)
# Cache the args as a text string to save them in the output dir later
args_text = yaml.safe_dump(args.__dict__, default_flow_style=False)
return args, args_text
def main():
setup_default_logging()
args, args_text = _parse_args()
args.prefetcher = not args.no_prefetcher
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
if args.distributed and args.num_gpu > 1:
_logger.warning(
'Using more than one GPU per process in distributed mode is not allowed.Setting num_gpu to 1.')
args.num_gpu = 1
args.device = 'cuda:0'
args.world_size = 1
args.rank = 0 # global rank
if args.distributed:
args.num_gpu = 1
args.device = 'cuda:%d' % args.local_rank
torch.cuda.set_device(args.local_rank)
args.world_size = int(os.environ['WORLD_SIZE'])
args.rank = int(os.environ['RANK'])
torch.distributed.init_process_group(backend='nccl', init_method=args.init_method, rank=args.rank, world_size=args.world_size)
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
_logger.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
_logger.info('Training with a single process on %d GPUs.' % args.num_gpu)
torch.manual_seed(args.seed + args.rank)
extra_dict = {}
extra_dict['args'] = args
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
drop_connect_rate=args.drop_connect,
drop_path_rate=args.drop_path,
drop_block_rate=args.drop_block,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint,
**extra_dict)
################### flops #################
if args.local_rank == 0:
print(model)
from thop import profile
from thop.vision.basic_hooks import zero_ops
if hasattr(model, 'default_cfg'):
default_cfg = model.default_cfg
input_size = [1] + list(default_cfg['input_size'])
else:
input_size = [1, 3, 224, 224]
input = torch.randn(input_size)#.cuda()
_, params = profile(model, inputs=(input, ), custom_ops={nn.BatchNorm2d: zero_ops})
print('params:',params)
from torchprofile import profile_macs
macs = profile_macs(model, input)
print('model flops:', macs, 'input_size:', input_size)
if args.local_rank == 0:
_logger.info('Model %s created, param count: %d' %
(args.model, sum([m.numel() for m in model.parameters()])))
data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0)
num_aug_splits = 0
if args.aug_splits > 0:
assert args.aug_splits > 1, 'A split of 1 makes no sense'
num_aug_splits = args.aug_splits
if args.split_bn:
assert num_aug_splits > 1 or args.resplit
model = convert_splitbn_model(model, max(num_aug_splits, 2))
use_amp = None
if args.amp:
# for backwards compat, `--amp` arg tries apex before native amp
if has_apex:
args.apex_amp = True
elif has_native_amp:
args.native_amp = True
if args.apex_amp and has_apex:
use_amp = 'apex'
elif args.native_amp and has_native_amp:
use_amp = 'native'
elif args.apex_amp or args.native_amp:
_logger.warning("Neither APEX or native Torch AMP is available, using float32. "
"Install NVIDA apex or upgrade to PyTorch 1.6")
if args.num_gpu > 1:
if use_amp == 'apex':
_logger.warning(
'Apex AMP does not work well with nn.DataParallel, disabling. Use DDP or Torch AMP.')
use_amp = None
model = nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
assert not args.channels_last, "Channels last not supported with DP, use DDP."
else:
model.cuda()
if args.channels_last:
model = model.to(memory_format=torch.channels_last)
optimizer = create_optimizer(args, model)
amp_autocast = suppress # do nothing
loss_scaler = None
if use_amp == 'apex':
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
loss_scaler = ApexScaler()
if args.local_rank == 0:
_logger.info('Using NVIDIA APEX AMP. Training in mixed precision.')
elif use_amp == 'native':
amp_autocast = torch.cuda.amp.autocast
loss_scaler = NativeScaler()
if args.local_rank == 0:
_logger.info('Using native Torch AMP. Training in mixed precision.')
else:
if args.local_rank == 0:
_logger.info('AMP not enabled. Training in float32.')
# optionally resume from a checkpoint
resume_epoch = None
if args.resume:
resume_epoch = resume_checkpoint(
model, args.resume,
optimizer=None if args.no_resume_opt else optimizer,
loss_scaler=None if args.no_resume_opt else loss_scaler,
log_info=args.local_rank == 0)
model_ema = None
if args.model_ema:
# Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
model_ema = ModelEma(
model,
decay=args.model_ema_decay,
device='cpu' if args.model_ema_force_cpu else '',
resume=args.resume)
if args.distributed:
if args.sync_bn:
assert not args.split_bn
try:
if has_apex and use_amp != 'native':
# Apex SyncBN preferred unless native amp is activated
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
_logger.info(
'Converted model to use Synchronized BatchNorm. WARNING: You may have issues if using '
'zero initialized BN layers (enabled by default for ResNets) while sync-bn enabled.')
except Exception as e:
_logger.error('Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1')
if has_apex and use_amp != 'native':
# Apex DDP preferred unless native amp is activated
if args.local_rank == 0:
_logger.info("Using NVIDIA APEX DistributedDataParallel.")
model = ApexDDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
_logger.info("Using native Torch DistributedDataParallel.")
#model = NativeDDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
model = NativeDDP(model, device_ids=[args.local_rank],find_unused_parameters=True)
# NOTE: EMA model does not need to be wrapped by DDP
lr_scheduler, num_epochs = create_scheduler(args, optimizer)
start_epoch = 0
if args.start_epoch is not None:
# a specified start_epoch will always override the resume epoch
start_epoch = args.start_epoch
elif resume_epoch is not None:
start_epoch = resume_epoch
if lr_scheduler is not None and start_epoch > 0:
lr_scheduler.step(start_epoch)
if args.local_rank == 0:
_logger.info('Scheduled epochs: {}'.format(num_epochs))
if 'cifar100' in args.data:
dataset_train = torchvision.datasets.CIFAR100(root=args.data, train=True, download=False, transform=None)
elif 'cifar10' in args.data:
dataset_train = torchvision.datasets.CIFAR10(root=args.data, train=True, download=False, transform=None)
else:
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
_logger.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
mixup_fn = None
mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
if mixup_active:
mixup_args = dict(
mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax,
prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode,
label_smoothing=args.smoothing, num_classes=args.num_classes)
if args.prefetcher:
assert not num_aug_splits # collate conflict (need to support deinterleaving in collate mixup)
collate_fn = FastCollateMixup(**mixup_args)
else:
mixup_fn = Mixup(**mixup_args)
if num_aug_splits > 1:
dataset_train = AugMixDataset(dataset_train, num_splits=num_aug_splits)
train_interpolation = args.train_interpolation
if args.no_aug or not train_interpolation:
train_interpolation = data_config['interpolation']
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
no_aug=args.no_aug,
re_prob=args.reprob,
re_mode=args.remode,
re_count=args.recount,
re_split=args.resplit,
scale=args.scale,
ratio=args.ratio,
hflip=args.hflip,
vflip=args.vflip,
color_jitter=args.color_jitter,
auto_augment=args.aa,
num_aug_splits=num_aug_splits,
interpolation=train_interpolation,
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
pin_memory=args.pin_mem,
use_multi_epochs_loader=args.use_multi_epochs_loader,
repeated_aug=args.repeated_aug
)
if 'cifar100' in args.data:
dataset_eval = torchvision.datasets.CIFAR100(root=args.data, train=False, download=False, transform=None)
elif 'cifar10' in args.data:
dataset_eval = torchvision.datasets.CIFAR10(root=args.data, train=False, download=False, transform=None)
else:
eval_dir = os.path.join(args.data, 'val')
if not os.path.isdir(eval_dir):
eval_dir = os.path.join(args.data, 'validation')
if not os.path.isdir(eval_dir):
_logger.error('Validation folder does not exist at: {}'.format(eval_dir))
exit(1)
dataset_eval = Dataset(eval_dir)
loader_eval = create_loader(
dataset_eval,
input_size=data_config['input_size'],
batch_size=args.validation_batch_size_multiplier * args.batch_size,
is_training=False,
use_prefetcher=args.prefetcher,
interpolation=data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
crop_pct=data_config['crop_pct'],
pin_memory=args.pin_mem,
)
if args.jsd:
assert num_aug_splits > 1 # JSD only valid with aug splits set
train_loss_fn = JsdCrossEntropy(num_splits=num_aug_splits, smoothing=args.smoothing).cuda()
elif mixup_active:
# smoothing is handled with mixup target transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
eval_metric = args.eval_metric
best_metric = None
best_epoch = None
saver = None
output_dir = ''
if args.local_rank == 0:
output_base = args.output if args.output else './output'
exp_name = '-'.join([
datetime.now().strftime("%Y%m%d-%H%M%S"),
args.model,
str(data_config['input_size'][-1])
])
output_dir = get_outdir(output_base, 'train', exp_name)
decreasing = True if eval_metric == 'loss' else False
saver = CheckpointSaver(
model=model, optimizer=optimizer, args=args, model_ema=model_ema, amp_scaler=loss_scaler,
checkpoint_dir=output_dir, recovery_dir=output_dir, decreasing=decreasing,max_history=args.max_history)
###max_history 最多保存的checkpoint数
with open(os.path.join(output_dir, 'args.yaml'), 'w') as f:
f.write(args_text)
try:
for epoch in range(start_epoch, num_epochs):
if args.distributed:
loader_train.sampler.set_epoch(epoch)
train_metrics = train_epoch(
epoch, model, loader_train, optimizer, train_loss_fn, args,
lr_scheduler=lr_scheduler, saver=saver, output_dir=output_dir,
amp_autocast=amp_autocast, loss_scaler=loss_scaler, model_ema=model_ema, mixup_fn=mixup_fn)
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
if args.local_rank == 0:
_logger.info("Distributing BatchNorm running means and vars")
distribute_bn(model, args.world_size, args.dist_bn == 'reduce')
eval_metrics = validate(model, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast)
if model_ema is not None and not args.model_ema_force_cpu:
if args.distributed and args.dist_bn in ('broadcast', 'reduce'):
distribute_bn(model_ema, args.world_size, args.dist_bn == 'reduce')
ema_eval_metrics = validate(
model_ema.ema, loader_eval, validate_loss_fn, args, amp_autocast=amp_autocast, log_suffix=' (EMA)')
eval_metrics = ema_eval_metrics
if lr_scheduler is not None:
# step LR for next epoch
lr_scheduler.step(epoch + 1, eval_metrics[eval_metric])
update_summary(
epoch, train_metrics, eval_metrics, os.path.join(output_dir, 'summary.csv'),
write_header=best_metric is None)
if saver is not None:
# save proper checkpoint with eval metric
save_metric = eval_metrics[eval_metric]
best_metric, best_epoch = saver.save_checkpoint(epoch, metric=save_metric)
except KeyboardInterrupt:
pass
if best_metric is not None:
_logger.info('*** Best metric: {0} (epoch {1})'.format(best_metric, best_epoch))
def train_epoch(
epoch, model, loader, optimizer, loss_fn, args,
lr_scheduler=None, saver=None, output_dir='', amp_autocast=suppress,
loss_scaler=None, model_ema=None, mixup_fn=None):
if args.mixup_off_epoch and epoch >= args.mixup_off_epoch:
if args.prefetcher and loader.mixup_enabled:
loader.mixup_enabled = False
elif mixup_fn is not None:
mixup_fn.mixup_enabled = False
second_order = hasattr(optimizer, 'is_second_order') and optimizer.is_second_order
batch_time_m = AverageMeter()
data_time_m = AverageMeter()
losses_m = AverageMeter()
top1_m = AverageMeter()
top5_m = AverageMeter()
model.train()
end = time.time()
last_idx = len(loader) - 1
num_updates = epoch * len(loader)
for batch_idx, (input, target) in enumerate(loader):
last_batch = batch_idx == last_idx
data_time_m.update(time.time() - end)
if not args.prefetcher:
input, target = input.cuda(), target.cuda()
if mixup_fn is not None:
input, target = mixup_fn(input, target)
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
with amp_autocast():
output = model(input)
loss = loss_fn(output, target)
acc1, acc5 = accuracy(output, target.argmax(dim=-1), topk=(1, 5))
if args.distributed:
acc1 = reduce_tensor(acc1, args.world_size)
acc5 = reduce_tensor(acc5, args.world_size)
if not args.distributed:
losses_m.update(loss.item(), input.size(0))
optimizer.zero_grad()
if loss_scaler is not None:
loss_scaler(
loss, optimizer, clip_grad=args.clip_grad, parameters=model.parameters(), create_graph=second_order)
else:
loss.backward(create_graph=second_order)
if args.clip_grad is not None:
torch.nn.utils.clip_grad_norm_(model.parameters(), args.clip_grad)
optimizer.step()
torch.cuda.synchronize()
top1_m.update(acc1.item(), output.size(0))
top5_m.update(acc5.item(), output.size(0))
if model_ema is not None:
model_ema.update(model)
num_updates += 1
batch_time_m.update(time.time() - end)
if last_batch or batch_idx % args.log_interval == 0:
lrl = [param_group['lr'] for param_group in optimizer.param_groups]
lr = sum(lrl) / len(lrl)
if args.distributed:
reduced_loss = reduce_tensor(loss.data, args.world_size)
losses_m.update(reduced_loss.item(), input.size(0))
if args.local_rank == 0:
_logger.info(
'Train: {} [{:>4d}/{} ({:>3.0f}%)] '
'Loss: {loss.val:>9.6f} ({loss.avg:>6.4f}) '
'Time: {batch_time.val:.3f}s, {rate:>7.2f}/s '
'({batch_time.avg:.3f}s, {rate_avg:>7.2f}/s) '
'LR: {lr:.3e} '
'Data: {data_time.val:.3f} ({data_time.avg:.3f})'.format(
epoch,
batch_idx, len(loader),
100. * batch_idx / last_idx,
loss=losses_m,
batch_time=batch_time_m,
rate=input.size(0) * args.world_size / batch_time_m.val,
rate_avg=input.size(0) * args.world_size / batch_time_m.avg,
lr=lr,
data_time=data_time_m))
_logger.info(
'Acc@1: {top1.val:>7.4f} ({top1.avg:>7.4f}) '
'Acc@5: {top5.val:>7.4f} ({top5.avg:>7.4f})'.format(top1=top1_m, top5=top5_m))
if args.save_images and output_dir:
torchvision.utils.save_image(
input,
os.path.join(output_dir, 'train-batch-%d.jpg' % batch_idx),
padding=0,
normalize=True)
if saver is not None and args.recovery_interval and (
last_batch or (batch_idx + 1) % args.recovery_interval == 0):
saver.save_recovery(epoch, batch_idx=batch_idx)
if lr_scheduler is not None:
lr_scheduler.step_update(num_updates=num_updates, metric=losses_m.avg)
end = time.time()
# end for
if hasattr(optimizer, 'sync_lookahead'):
optimizer.sync_lookahead()
return OrderedDict([('loss', losses_m.avg)])
def validate(model, loader, loss_fn, args, amp_autocast=suppress, log_suffix=''):
batch_time_m = AverageMeter()
losses_m = AverageMeter()
top1_m = AverageMeter()
top5_m = AverageMeter()
model.eval()
end = time.time()
last_idx = len(loader) - 1
with torch.no_grad():
for batch_idx, (input, target) in enumerate(loader):
last_batch = batch_idx == last_idx
if not args.prefetcher:
input = input.cuda()
target = target.cuda()
if args.channels_last:
input = input.contiguous(memory_format=torch.channels_last)
with amp_autocast():
output = model(input)
if isinstance(output, (tuple, list)):
output = output[0]
# augmentation reduction
reduce_factor = args.tta
if reduce_factor > 1:
output = output.unfold(0, reduce_factor, reduce_factor).mean(dim=2)
target = target[0:target.size(0):reduce_factor]
loss = loss_fn(output, target)
acc1, acc5 = accuracy(output, target, topk=(1, 5))
if args.distributed:
reduced_loss = reduce_tensor(loss.data, args.world_size)
acc1 = reduce_tensor(acc1, args.world_size)
acc5 = reduce_tensor(acc5, args.world_size)
else:
reduced_loss = loss.data
torch.cuda.synchronize()
losses_m.update(reduced_loss.item(), input.size(0))
top1_m.update(acc1.item(), output.size(0))
top5_m.update(acc5.item(), output.size(0))
batch_time_m.update(time.time() - end)
end = time.time()
if args.local_rank == 0 and (last_batch or batch_idx % args.log_interval == 0):
log_name = 'Test' + log_suffix
_logger.info(
'{0}: [{1:>4d}/{2}] '
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Acc@1: {top1.val:>7.4f} ({top1.avg:>7.4f}) '
'Acc@5: {top5.val:>7.4f} ({top5.avg:>7.4f})'.format(
log_name, batch_idx, last_idx, batch_time=batch_time_m,
loss=losses_m, top1=top1_m, top5=top5_m))
metrics = OrderedDict([('loss', losses_m.avg), ('top1', top1_m.avg), ('top5', top5_m.avg)])
return metrics
if __name__ == '__main__':
main()
文章目录
- 摘要
- YoloV8改进策略：基于分层注意力的FasterViT，让YoloV8实现性能的飞跃
- YoloV8改进策略：InceptionNext主干替换YoloV8和YoloV5的主干
- YoloV8改进策略：轻量级的CloFormer助力Yolov8在速度和精度上实现双双提升
- YoloV8改进策略：InceptionNeXt和YoloV8完美结合，让YoloV8大放异彩
- YoloV8改进策略：新出炉的EMA注意力机制助力YoloV8更加强大
- YoloV8改进策略：VanillaNet极简主义网络，大大降低YoloV8的参数
- YoloV8改进策略：RFAConv模块即插即用，实现YoloV8丝滑上分
- YoloV8改进策略：让SeaFormer走进Yolov8的视野，轻量高效的注意力模块展现出无与伦比的魅力
- YoloV8改进策略：将DCN v1与v2运用到YoloV8中，化身成上分小黑子
- YoloV8改进策略：基于双层路由注意力的视觉Transformer提升YoloV8的检测能力
- YoloV8改进策略：来自谷歌最新的优化器——Lion，在速度和精度上双双提升。Adam表示年轻人不讲武德
- YoloV8改进策略：Conv2Former与YoloV8深度融合，极简网络，极高性能
- YoloV8改进策略：将ConvNextV2与YoloV8激情碰撞，能迸发出什么样的火花呢？
- YoloV8改进策略：将CIoU替换成Wise-IoU，幸福涨点，值得拥有，还支持EIoU、GIoU、DIoU、SIoU无缝替换。
- YoloV8改进策略：增加分支，减少漏检
- YoloV8改进策略：将FasterNet与YoloV8深度融合，打造更快更强的检测网络
- Yolov8网络详解与实战（附数据集）
摘要
本专栏是讲解如何改进Yolov8的专栏。改进方法采用了最新的论文提到的方法。改进的方法包括：增加注意力机制、更换卷积、更换block、更换backbone、更换head、更换优化器等；每篇文章提供了一种到N种改进方法。
评测用的数据集是我自己标注的数据集，里面包含32种飞机。每种改进方法我都做了测评，并与官方的模型做对比。
代码和PDF版本的文章，我在验证无误后会上传到百度网盘中，方便大家下载使用。
这个专栏，求质不求量，争取尽心尽力打造精品专栏！！！
谢谢大家支持！！！
YoloV8改进策略：基于分层注意力的FasterViT，让YoloV8实现性能的飞跃
YoloV8改进策略：基于分层注意力的FasterViT，让YoloV8实现性能的飞跃
这篇文章向大家展示如何使用FasterViT改进YoloV8，我尝试了几种方法，选出了三种效果比较好的方法推荐给大家。
FasterViT结合了cnn的快速局部表示学习和ViT的全局建模特性的优点。新提出的分层注意力(HAT)方法将具有二次复杂度的全局自注意力分解为具有减少计算成本的多级注意力。我们受益于基于窗口的高效自我关注。每个窗口都可以访问参与局部和全局表示学习的专用载体Token。在高层次上，全局的自我关注使高效的跨窗口通信能够以较低的成本实现。FasterViT在精度与图像吞吐量方面达到了SOTA Pareto-front。
YoloV8改进策略：InceptionNext主干替换YoloV8和YoloV5的主干
YoloV8改进策略：InceptionNext主干替换YoloV8和YoloV5的主干
这篇文章主要讲解如何使用InceptionNext主干网络替换YoloV8和YoloV5的主干。更改了InceptionNext网络结构，和Yolov5、YoloV8的架构。
YoloV8改进策略：轻量级的CloFormer助力Yolov8在速度和精度上实现双双提升
YoloV8改进策略：轻量级的CloFormer助力Yolov8在速度和精度上实现双双提升
CloFormer是清华大学在今年发表的轻量级主干网络，引入了AttnConv，一种attention风格的卷积算子。所提出的AttnConv使用共享权重来聚合局部信息，并配置精心设计的上下文感知权重来增强局部特征。AttnConv和普通attention的结合使用池化来减少CloFormer中的FLOPs，使模型能够感知高频和低频信息。
YoloV8改进策略：InceptionNeXt和YoloV8完美结合，让YoloV8大放异彩
YoloV8改进策略：InceptionNeXt和YoloV8完美结合，让YoloV8大放异彩
InceptionNeXt是今年颜水成团队发布的一篇论文，将ConvNext和Inception的思想融合，即IncepitonNeXt。InceptionNeXt-T实现了比convnext - t高1.6倍的训练吞吐量，并在ImageNet- 1K上实现了0.2%的top-1精度提高。
YoloV8改进策略：新出炉的EMA注意力机制助力YoloV8更加强大
YoloV8改进策略：新出炉的EMA注意力机制助力YoloV8更加强大
EMA注意力机制是今年新的高效的多尺度注意力模块。以保留每个通道上的信息和降低计算开销为目标，将部分通道重塑为批量维度，并将通道维度分组为多个子特征，使空间语义特征在每个特征组中均匀分布。具体来说，除了对全局信息进行编码以重新校准每个并行分支中的通道权重外，还通过跨维度交互进一步聚合两个并行分支的输出特征，以捕获像素级成对关系。
YoloV8改进策略：VanillaNet极简主义网络，大大降低YoloV8的参数
YoloV8改进策略：VanillaNet极简主义网络，大大降低YoloV8的参数
VanillaNet，一个包含优雅设计的神经网络架构。通过避免高深度，shotcut和复杂的操作，如自主意力，VanillaNet令人耳目一新的简洁，但非常强大。每一层都被精心制作得紧凑而直接，非线性激活函数在训练后被修剪以恢复原始结构。VanillaNet克服了固有复杂性的挑战，使其成为资源受限环境的理想选择。其易于理解和高度简化的架构为高效部署提供了新的可能性。大量的实验表明，VanillaNet提供的性能与著名的深度神经网络和视觉转换器相当，展示了极简主义在深度学习中的力量。VanillaNet的这一富有远见的旅程具有重新定义景观和挑战基础模型现状的巨大潜力，为优雅有效的模型设计开辟了一条新的道路。
YoloV8改进策略：RFAConv模块即插即用，实现YoloV8丝滑上分
YoloV8改进策略：RFAConv模块即插即用，实现YoloV8丝滑上分
RFAConv是一种新的注意力机制，称为感受野注意力（RFA）。卷积块注意力模块（CBAM）和协调注意力模块（CA）只关注空间特征，不能完全解决卷积核参数共享的问题，但在RFA中，感受野空间特征不仅集中，而且为大尺寸卷积核提供了良好的注意力权重。RFA设计的感受野注意力卷积运算（RFAConv）可以被认为是取代标准卷积的一种新方法，它带来的计算成本和许多参数几乎可以忽略不计。由于作者没有开源我自己复现了一版，并尝试将其加入到YoloV8网络中。
YoloV8改进策略：让SeaFormer走进Yolov8的视野，轻量高效的注意力模块展现出无与伦比的魅力
YoloV8改进策略：让SeaFormer走进Yolov8的视野，轻量高效的注意力模块展现出无与伦比的魅力
SeaFormer使用压缩轴和细节增强的方法设计了一个通用的注意力块。它可以进一步用于创建一系列具有卓越成本效益的骨干体系结构。再加上一个轻分割头，我们在基于arm的移动设备上在ADE20K和cityscape数据集上实现了分割精度和延迟之间的最佳权衡。关键的是，我们以更好的性能和更低的延迟击败了适合移动设备的竞争对手和基于transformer的对手，而且没有花哨的东西。
YoloV8改进策略：将DCN v1与v2运用到YoloV8中，化身成上分小黑子
YoloV8改进策略：将DCN v1与v2运用到YoloV8中，化身成上分小黑子
尝试用DCNv1与DCNv2代替普通的卷积！
YoloV8改进策略：基于双层路由注意力的视觉Transformer提升YoloV8的检测能力
YoloV8改进策略：基于双层路由注意力的视觉Transformer提升YoloV8的检测能力
双层路由注意力实现具有内容感知的更灵活的计算分配。利用稀疏性来节省计算和内存，同时只涉及适用于GPU的密集矩阵乘法。用所提出的双层路由注意力建立了一个新的通用视觉transformer，称为BiFormer。
YoloV8改进策略：来自谷歌最新的优化器——Lion，在速度和精度上双双提升。Adam表示年轻人不讲武德
YoloV8改进策略：来自谷歌最新的优化器——Lion，在速度和精度上双双提升。Adam表示年轻人不讲武德
Lion将ViT在ImageNet上的准确率提高了2%，并在JFT上节省了高达5倍的预训练计算。在视觉-语言对比学习方面，在ImageNet上实现了88.3%的零样本和91.1%的微调精度，分别超过了之前的最佳结果2%和0.1%。在扩散模型上，Lion通过获得更好的FID分数并将训练计算量减少了2.3倍，超越了Adam。在自回归、掩码语言建模和微调方面，Lion表现出与Adam类似或更好的性能。对Lion的分析表明，其性能增益随着训练批大小的增加而增长。由于符号函数产生的更新范数更大，它还需要比Adam更小的学习率。
YoloV8改进策略：Conv2Former与YoloV8深度融合，极简网络，极高性能
YoloV8改进策略：Conv2Former与YoloV8深度融合，极简网络，极高性能
Conv2Former是在ConvNeXt基础上，做了进一步的优化，性能得到了提升。
YoloV8改进策略：将ConvNextV2与YoloV8激情碰撞，能迸发出什么样的火花呢？
YoloV8改进策略：将ConvNextV2与YoloV8激情碰撞，能迸发出什么样的火花呢
ConvNextV2将一个全卷积掩码自编码器框架和一个新的全局响应归一化(GRN)层，可以添加到ConvNeXt架构中，以增强通道间的特征竞争，它显著提高了纯ConvNets在各种识别基准上的性能，包括ImageNet分类、COCO检测和ADE20K分割。
YoloV8改进策略：将CIoU替换成Wise-IoU，幸福涨点，值得拥有，还支持EIoU、GIoU、DIoU、SIoU无缝替换。
YoloV8改进策略：将CIoU替换成Wise-IoU，幸福涨点，值得拥有，还支持EIoU、GIoU、DIoU、SIoU无缝替换。
这篇文章讲述如何在yolov8中，使用Wise-IoU涨点。首先，翻译了论文，让大家了解什么是Wise IoU，以及Wise IoU的三个版本。接下来讲解如何在yolov8中添加Wise IoU。
YoloV8改进策略：增加分支，减少漏检
通过增加一个分支，来提高小目标的检测
YoloV8改进策略：将FasterNet与YoloV8深度融合，打造更快更强的检测网络
YoloV8改进策略：将FasterNet与YoloV8深度融合，打造更快更强的检测网络
fastternet，这是一种新的神经网络家族，它在各种设备上获得了比其他网络更高的运行速度，而不影响各种视觉任务的准确性。
734
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
