You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
pytorch-image-models/train.py

579 lines
25 KiB

import argparse
import time
import logging
import yaml
from datetime import datetime
try:
from apex import amp
from apex.parallel import DistributedDataParallel as DDP
from apex.parallel import convert_syncbn_model
has_apex = True
except ImportError:
from torch.nn.parallel import DistributedDataParallel as DDP
has_apex = False
from timm.data import Dataset, create_loader, resolve_data_config, FastCollateMixup, mixup_target
from timm.models import create_model, resume_checkpoint
from timm.utils import *
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from timm.optim import create_optimizer
from timm.scheduler import create_scheduler
import torch
import torch.nn as nn
import torchvision.utils
torch.backends.cudnn.benchmark = True
# 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='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='avg', type=str, metavar='POOL',
help='Type of global pool, "avg", "max", "avgmax", "avgmaxc" (default: "avg")')
parser.add_argument('--img-size', type=int, default=None, metavar='N',
help='Image patch size (default: None => model default)')
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('--drop', type=float, default=0.0, metavar='DROP',
help='Dropout rate (default: 0.)')
# Optimizer parameters
parser.add_argument('--opt', default='sgd', type=str, metavar='OPTIMIZER',
help='Optimizer (default: "sgd"')
parser.add_argument('--opt-eps', default=1e-8, type=float, metavar='EPSILON',
help='Optimizer Epsilon (default: 1e-8)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)')
parser.add_argument('--weight-decay', type=float, default=0.0001,
help='weight decay (default: 0.0001)')
# 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('--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=int, 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('--decay-rate', '--dr', type=float, default=0.1, metavar='RATE',
help='LR decay rate (default: 0.1)')
# Augmentation parameters
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('--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('--mixup', type=float, default=0.0,
help='mixup alpha, mixup enabled if > 0. (default: 0.)')
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)')
# 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)')
# 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 amp for mixed precision training')
parser.add_argument('--sync-bn', action='store_true',
help='enabling apex sync BN.')
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='prec1', type=str, metavar='EVAL_METRIC',
help='Best metric (default: "prec1"')
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)
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:
logging.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)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
args.world_size = torch.distributed.get_world_size()
args.rank = torch.distributed.get_rank()
assert args.rank >= 0
if args.distributed:
logging.info('Training in distributed mode with multiple processes, 1 GPU per process. Process %d, total %d.'
% (args.rank, args.world_size))
else:
logging.info('Training with a single process on %d GPUs.' % args.num_gpu)
torch.manual_seed(args.seed + args.rank)
model = create_model(
args.model,
pretrained=args.pretrained,
num_classes=args.num_classes,
drop_rate=args.drop,
global_pool=args.gp,
bn_tf=args.bn_tf,
bn_momentum=args.bn_momentum,
bn_eps=args.bn_eps,
checkpoint_path=args.initial_checkpoint)
if args.local_rank == 0:
logging.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)
if args.num_gpu > 1:
if args.amp:
logging.warning(
'AMP does not work well with nn.DataParallel, disabling. Use distributed mode for multi-GPU AMP.')
args.amp = False
model = nn.DataParallel(model, device_ids=list(range(args.num_gpu))).cuda()
else:
model.cuda()
optimizer = create_optimizer(args, model)
use_amp = False
if has_apex and args.amp:
model, optimizer = amp.initialize(model, optimizer, opt_level='O1')
use_amp = True
if args.local_rank == 0:
logging.info('NVIDIA APEX {}. AMP {}.'.format(
'installed' if has_apex else 'not installed', 'on' if use_amp else 'off'))
# optionally resume from a checkpoint
resume_state = {}
resume_epoch = None
if args.resume:
resume_state, resume_epoch = resume_checkpoint(model, args.resume)
if resume_state and not args.no_resume_opt:
if 'optimizer' in resume_state:
if args.local_rank == 0:
logging.info('Restoring Optimizer state from checkpoint')
optimizer.load_state_dict(resume_state['optimizer'])
if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
if args.local_rank == 0:
logging.info('Restoring NVIDIA AMP state from checkpoint')
amp.load_state_dict(resume_state['amp'])
resume_state = None
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:
try:
if has_apex:
model = convert_syncbn_model(model)
else:
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
if args.local_rank == 0:
logging.info('Converted model to use Synchronized BatchNorm.')
except Exception as e:
logging.error('Failed to enable Synchronized BatchNorm. Install Apex or Torch >= 1.1')
if has_apex:
model = DDP(model, delay_allreduce=True)
else:
if args.local_rank == 0:
logging.info("Using torch DistributedDataParallel. Install NVIDIA Apex for Apex DDP.")
model = DDP(model, device_ids=[args.local_rank]) # can use device str in Torch >= 1.1
# 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:
logging.info('Scheduled epochs: {}'.format(num_epochs))
train_dir = os.path.join(args.data, 'train')
if not os.path.exists(train_dir):
logging.error('Training folder does not exist at: {}'.format(train_dir))
exit(1)
dataset_train = Dataset(train_dir)
collate_fn = None
if args.prefetcher and args.mixup > 0:
collate_fn = FastCollateMixup(args.mixup, args.smoothing, args.num_classes)
loader_train = create_loader(
dataset_train,
input_size=data_config['input_size'],
batch_size=args.batch_size,
is_training=True,
use_prefetcher=args.prefetcher,
rand_erase_prob=args.reprob,
rand_erase_mode=args.remode,
rand_erase_count=args.recount,
color_jitter=args.color_jitter,
auto_augment=args.aa,
interpolation='random', # FIXME cleanly resolve this? data_config['interpolation'],
mean=data_config['mean'],
std=data_config['std'],
num_workers=args.workers,
distributed=args.distributed,
collate_fn=collate_fn,
)
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):
logging.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=4 * 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,
)
if args.mixup > 0.:
# smoothing is handled with mixup label transform
train_loss_fn = SoftTargetCrossEntropy().cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
elif args.smoothing:
train_loss_fn = LabelSmoothingCrossEntropy(smoothing=args.smoothing).cuda()
validate_loss_fn = nn.CrossEntropyLoss().cuda()
else:
train_loss_fn = nn.CrossEntropyLoss().cuda()
validate_loss_fn = train_loss_fn
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(checkpoint_dir=output_dir, decreasing=decreasing)
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,
use_amp=use_amp, model_ema=model_ema)
eval_metrics = validate(model, loader_eval, validate_loss_fn, args)
if model_ema is not None and not args.model_ema_force_cpu:
ema_eval_metrics = validate(
model_ema.ema, loader_eval, validate_loss_fn, args, 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(
6 years ago
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(
model, optimizer, args,
epoch=epoch, model_ema=model_ema, metric=save_metric, use_amp=use_amp)
except KeyboardInterrupt:
pass
if best_metric is not None:
logging.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='', use_amp=False, model_ema=None):
if args.prefetcher and args.mixup > 0 and loader.mixup_enabled:
if args.mixup_off_epoch and epoch >= args.mixup_off_epoch:
loader.mixup_enabled = False
batch_time_m = AverageMeter()
data_time_m = AverageMeter()
losses_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 args.mixup > 0.:
lam = 1.
if not args.mixup_off_epoch or epoch < args.mixup_off_epoch:
lam = np.random.beta(args.mixup, args.mixup)
input.mul_(lam).add_(1 - lam, input.flip(0))
target = mixup_target(target, args.num_classes, lam, args.smoothing)
output = model(input)
loss = loss_fn(output, target)
if not args.distributed:
losses_m.update(loss.item(), input.size(0))
optimizer.zero_grad()
if use_amp:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
torch.cuda.synchronize()
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:
logging.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))
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(
model, optimizer, args, epoch, model_ema=model_ema, use_amp=use_amp, 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, log_suffix=''):
batch_time_m = AverageMeter()
losses_m = AverageMeter()
prec1_m = AverageMeter()
prec5_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()
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)
prec1, prec5 = accuracy(output, target, topk=(1, 5))
if args.distributed:
reduced_loss = reduce_tensor(loss.data, args.world_size)
prec1 = reduce_tensor(prec1, args.world_size)
prec5 = reduce_tensor(prec5, args.world_size)
else:
reduced_loss = loss.data
torch.cuda.synchronize()
losses_m.update(reduced_loss.item(), input.size(0))
prec1_m.update(prec1.item(), output.size(0))
prec5_m.update(prec5.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
logging.info(
'{0}: [{1:>4d}/{2}] '
'Time: {batch_time.val:.3f} ({batch_time.avg:.3f}) '
'Loss: {loss.val:>7.4f} ({loss.avg:>6.4f}) '
'Prec@1: {top1.val:>7.4f} ({top1.avg:>7.4f}) '
'Prec@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=prec1_m, top5=prec5_m))
metrics = OrderedDict([('loss', losses_m.avg), ('prec1', prec1_m.avg), ('prec5', prec5_m.avg)])
return metrics
if __name__ == '__main__':
main()