def main():
parser = argparse.ArgumentParser(description='Benchmark dataloading')
parser.add_argument('config', help='train config file path')
args = parser.parse_args()
cfg = Config.fromfile(args.config)
# init logger before other steps
logger = get_root_logger()
logger.info(f'MMPose Version: {__version__}')
logger.info(f'Config: {cfg.text}')
dataset = build_dataset(cfg.data.train)
data_loader = build_dataloader(
dataset,
samples_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=False,
shuffle=False)
# Start progress bar after first 5 batches
prog_bar = mmcv.ProgressBar(
len(dataset) - 5 * cfg.data.samples_per_gpu, start=False)
for i, data in enumerate(data_loader):
if i == 5:
prog_bar.start()
for _ in data['img']:
if i < 5:
continue
prog_bar.update()
def main():
args = parse_args()
cfg = mmcv.Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
args.work_dir = osp.join('./work_dirs',
osp.splitext(osp.basename(args.config))[0])
mmcv.mkdir_or_exist(osp.abspath(args.work_dir))
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
init_dist(args.launcher, **cfg.dist_params)
# build the dataloader
# TODO: support multiple images per gpu (only minor changes are needed)
dataset = build_dataset(cfg.data.test, dict(test_mode=True))
data_loader = build_dataloader(dataset,
samples_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
# build the model and load checkpoint
model = build_posenet(cfg.model)
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
wrap_fp16_model(model)
_ = load_checkpoint(model, args.checkpoint, map_location='cpu')
# for backward compatibility
if not distributed:
model = MMDataParallel(model, device_ids=[0])
outputs = single_gpu_test(model, data_loader)
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
outputs = multi_gpu_test(model, data_loader, args.tmpdir,
args.gpu_collect)
rank, _ = get_dist_info()
eval_config = cfg.get('eval_config', {})
eval_config = merge_configs(eval_config, dict(metrics=args.eval))
if rank == 0:
if args.out:
print(f'\nwriting results to {args.out}')
mmcv.dump(outputs, args.out)
dataset.evaluate(outputs, args.work_dir, **eval_config)
def main():
args = parse_args()
cfg = Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
# build the dataloader
dataset = build_dataset(cfg.data.val)
data_loader = build_dataloader(
dataset,
samples_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=False,
shuffle=False)
# build the model and load checkpoint
model = build_posenet(cfg.model)
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
wrap_fp16_model(model)
if args.fuse_conv_bn:
model = fuse_conv_bn(model)
model = MMDataParallel(model, device_ids=[0])
# the first several iterations may be very slow so skip them
num_warmup = 5
pure_inf_time = 0
# benchmark with total batch and take the average
for i, data in enumerate(data_loader):
torch.cuda.synchronize()
start_time = time.perf_counter()
with torch.no_grad():
model(return_loss=False, **data)
torch.cuda.synchronize()
elapsed = time.perf_counter() - start_time
if i >= num_warmup:
pure_inf_time += elapsed
if (i + 1) % args.log_interval == 0:
its = (i + 1 - num_warmup) / pure_inf_time
print(f'Done item [{i + 1:<3}], {its:.2f} items / s')
print(f'Overall average: {its:.2f} items / s')
print(f'Total time: {pure_inf_time:.2f} s')
def main():
args = parse_args()
cfg = Config.fromfile(args.config)
if args.options is not None:
cfg.merge_from_dict(args.options)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
# work_dir is determined in this priority: CLI > segment in file > filename
if args.work_dir is not None:
# update configs according to CLI args if args.work_dir is not None
cfg.work_dir = args.work_dir
elif cfg.get('work_dir', None) is None:
# use config filename as default work_dir if cfg.work_dir is None
cfg.work_dir = osp.join('./work_dirs',
osp.splitext(osp.basename(args.config))[0])
if args.resume_from is not None:
cfg.resume_from = args.resume_from
if args.gpu_ids is not None:
cfg.gpu_ids = args.gpu_ids
else:
cfg.gpu_ids = range(1) if args.gpus is None else range(args.gpus)
if args.autoscale_lr:
# apply the linear scaling rule (https://arxiv.org/abs/1706.02677)
cfg.optimizer['lr'] = cfg.optimizer['lr'] * len(cfg.gpu_ids) / 8
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
init_dist(args.launcher, **cfg.dist_params)
# create work_dir
mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
# init the logger before other steps
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, f'{timestamp}.log')
logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
# init the meta dict to record some important information such as
# environment info and seed, which will be logged
meta = dict()
# log env info
env_info_dict = collect_env()
env_info = '\n'.join([(f'{k}: {v}') for k, v in env_info_dict.items()])
dash_line = '-' * 60 + '\n'
logger.info('Environment info:\n' + dash_line + env_info + '\n' +
dash_line)
meta['env_info'] = env_info
# log some basic info
logger.info(f'Distributed training: {distributed}')
logger.info(f'Config:\n{cfg.pretty_text}')
# set random seeds
if args.seed is not None:
logger.info(f'Set random seed to {args.seed}, '
f'deterministic: {args.deterministic}')
set_random_seed(args.seed, deterministic=args.deterministic)
cfg.seed = args.seed
meta['seed'] = args.seed
model = build_posenet(cfg.model)
datasets = [build_dataset(cfg.data.train)]
if len(cfg.workflow) == 2:
val_dataset = copy.deepcopy(cfg.data.val)
val_dataset.pipeline = cfg.data.train.pipeline
datasets.append(build_dataset(val_dataset))
if cfg.checkpoint_config is not None:
# save mmpose version, config file content
# checkpoints as meta data
cfg.checkpoint_config.meta = dict(
mmpose_version=__version__ + get_git_hash(digits=7),
config=cfg.pretty_text,
)
train_model(
model,
datasets,
cfg,
distributed=distributed,
validate=(not args.no_validate),
timestamp=timestamp,
meta=meta)
def train_model(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
"""Train model entry function.
Args:
model (nn.Module): The model to be trained.
dataset (Dataset): Train dataset.
cfg (dict): The config dict for training.
distributed (bool): Whether to use distributed training.
Default: False.
validate (bool): Whether to do evaluation. Default: False.
timestamp (str | None): Local time for runner. Default: None.
meta (dict | None): Meta dict to record some important information.
Default: None
"""
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
# step 1: give default values and override (if exist) from cfg.data
loader_cfg = {
**dict(
seed=cfg.get('seed'),
drop_last=False,
dist=distributed,
num_gpus=len(cfg.gpu_ids)),
**({} if torch.__version__ != 'parrots' else dict(
prefetch_num=2,
pin_memory=False,
)),
**dict((k, cfg.data[k]) for k in [
'samples_per_gpu',
'workers_per_gpu',
'shuffle',
'seed',
'drop_last',
'prefetch_num',
'pin_memory',
'persistent_workers',
] if k in cfg.data)
}
# step 2: cfg.data.train_dataloader has highest priority
train_loader_cfg = dict(loader_cfg, **cfg.data.get('train_dataloader', {}))
data_loaders = [build_dataloader(ds, **train_loader_cfg) for ds in dataset]
# determine whether use adversarial training precess or not
use_adverserial_train = cfg.get('use_adversarial_train', False)
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', True)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
if use_adverserial_train:
# Use DistributedDataParallelWrapper for adversarial training
model = DistributedDataParallelWrapper(
model,
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(
model.cuda(cfg.gpu_ids[0]), device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizers(model, cfg.optimizer)
runner = EpochBasedRunner(
model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
if use_adverserial_train:
# The optimizer step process is included in the train_step function
# of the model, so the runner should NOT include optimizer hook.
optimizer_config = None
else:
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(
**cfg.optimizer_config, **fp16_cfg, distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
eval_cfg = cfg.get('evaluation', {})
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader_setting = dict(
samples_per_gpu=1,
workers_per_gpu=cfg.data.get('workers_per_gpu', 1),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
drop_last=False,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('val_dataloader', {}))
val_dataloader = build_dataloader(val_dataset, **dataloader_setting)
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
img = model.show_result(img,
result,
skeleton,
radius=radius,
pose_kpt_color=pose_kpt_color,
pose_limb_color=pose_limb_color,
kpt_score_thr=kpt_score_thr,
show=show,
out_file=out_file)
return img
cfg = Config.fromfile('configs/top_down/hrnet/cowacar/cowacar.py')
dataset = build_dataset(cfg.data.test, dict(test_mode=True))
data_loader = build_dataloader(dataset,
samples_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=False,
shuffle=False)
# for d in datasets:
# img = (255 * d['img'].permute(1, 2, 0).numpy())[:, :, ::-1]
# hm = d['target'].max(axis=0, keepdims=True).transpose(1, 2, 0)
# hm = cv2.resize(hm, None, fx=4, fy=4, interpolation=cv2.INTER_LINEAR)
# img *= 0.8
# img[..., -1] += 255 * 0.2 * hm
# cv2.imshow('0', img.astype(np.uint8))
# cv2.waitKey(0)
def main():
args = parse_args()
cfg = Config.fromfile(args.config)
if args.cfg_options is not None:
cfg.merge_from_dict(args.cfg_options)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
args.work_dir = osp.join('./work_dirs',
osp.splitext(osp.basename(args.config))[0])
mmcv.mkdir_or_exist(osp.abspath(args.work_dir))
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
init_dist(args.launcher, **cfg.dist_params)
# build the dataloader
dataset = build_dataset(cfg.data.test, dict(test_mode=True))
dataloader_setting = dict(samples_per_gpu=1,
workers_per_gpu=cfg.data.get(
'workers_per_gpu', 1),
dist=distributed,
shuffle=False,
drop_last=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('test_dataloader', {}))
data_loader = build_dataloader(dataset, **dataloader_setting)
# print(data_loader)
# build the model and load checkpoint
model = build_posenet(cfg.model)
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
wrap_fp16_model(model)
load_checkpoint(model, args.checkpoint, map_location='cpu')
if args.fuse_conv_bn:
model = fuse_conv_bn(model)
if not distributed:
model = MMDataParallel(model, device_ids=[0])
outputs = single_gpu_test(model, data_loader)
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
outputs = multi_gpu_test(model, data_loader, args.tmpdir,
args.gpu_collect)
rank, _ = get_dist_info()
eval_config = cfg.get('evaluation', {})
eval_config = merge_configs(eval_config, dict(metric=args.eval))
if rank == 0:
if args.out:
print(f'\nwriting results to {args.out}')
mmcv.dump(outputs, args.out)
results = dataset.evaluate(outputs, args.work_dir, **eval_config)
for k, v in sorted(results.items()):
print(f'{k}: {v}')
def main():
args = parse_args()
cfg = Config.fromfile(args.config)
if args.cfg_options is not None:
cfg.merge_from_dict(args.cfg_options)
# set multi-process settings
setup_multi_processes(cfg)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
# work_dir is determined in this priority: CLI > segment in file > filename
if args.work_dir is not None:
# update configs according to CLI args if args.work_dir is not None
cfg.work_dir = args.work_dir
elif cfg.get('work_dir', None) is None:
# use config filename as default work_dir if cfg.work_dir is None
cfg.work_dir = osp.join('./work_dirs',
osp.splitext(osp.basename(args.config))[0])
mmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
init_dist(args.launcher, **cfg.dist_params)
# build the dataloader
dataset = build_dataset(cfg.data.test, dict(test_mode=True))
# step 1: give default values and override (if exist) from cfg.data
loader_cfg = {
**dict(seed=cfg.get('seed'), drop_last=False, dist=distributed),
**({} if torch.__version__ != 'parrots' else dict(
prefetch_num=2,
pin_memory=False,
)),
**dict((k, cfg.data[k]) for k in [
'seed',
'prefetch_num',
'pin_memory',
'persistent_workers',
] if k in cfg.data)
}
# step2: cfg.data.test_dataloader has higher priority
test_loader_cfg = {
**loader_cfg,
**dict(shuffle=False, drop_last=False),
**dict(workers_per_gpu=cfg.data.get('workers_per_gpu', 1)),
**dict(samples_per_gpu=cfg.data.get('samples_per_gpu', 1)),
**cfg.data.get('test_dataloader', {})
}
data_loader = build_dataloader(dataset, **test_loader_cfg)
# build the model and load checkpoint
model = build_posenet(cfg.model)
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
wrap_fp16_model(model)
load_checkpoint(model, args.checkpoint, map_location='cpu')
if args.fuse_conv_bn:
model = fuse_conv_bn(model)
if not distributed:
model = MMDataParallel(model, device_ids=[args.gpu_id])
outputs = single_gpu_test(model, data_loader)
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
outputs = multi_gpu_test(model, data_loader, args.tmpdir,
args.gpu_collect)
rank, _ = get_dist_info()
eval_config = cfg.get('evaluation', {})
eval_config = merge_configs(eval_config, dict(metric=args.eval))
if rank == 0:
if args.out:
print(f'\nwriting results to {args.out}')
mmcv.dump(outputs, args.out)
results = dataset.evaluate(outputs, cfg.work_dir, **eval_config)
for k, v in sorted(results.items()):
print(f'{k}: {v}')
def main():
args = parse_args()
if 'cuda' in args.device.lower():
if torch.cuda.is_available():
with_cuda = True
else:
raise RuntimeError('No CUDA device found, please check it again.')
else:
with_cuda = False
if args.root_work_dir is None:
# get the current time stamp
now = datetime.now()
ts = now.strftime('%Y_%m_%d_%H_%M')
args.root_work_dir = f'work_dirs/inference_speed_test_{ts}'
mmcv.mkdir_or_exist(osp.abspath(args.root_work_dir))
cfg = mmcv.load(args.config)
dummy_datasets = mmcv.load(args.dummy_dataset_config)['dummy_datasets']
results = []
for i in range(args.priority + 1):
models = cfg['model_list'][f'P{i}']
for cur_model in models:
cfg_file = cur_model['config']
model_cfg = Config.fromfile(cfg_file)
test_dataset = model_cfg['data']['test']
dummy_dataset = dummy_datasets[test_dataset['type']]
test_dataset.update(dummy_dataset)
dataset = build_dataset(test_dataset)
data_loader = build_dataloader(
dataset,
samples_per_gpu=args.batch_size,
workers_per_gpu=model_cfg.data.workers_per_gpu,
dist=False,
shuffle=False)
data_loader = IterLoader(data_loader)
if 'pretrained' in model_cfg.model.keys():
del model_cfg.model['pretrained']
model = init_pose_model(model_cfg, device=args.device.lower())
fp16_cfg = model_cfg.get('fp16', None)
if fp16_cfg is not None:
wrap_fp16_model(model)
if args.fuse_conv_bn:
model = fuse_conv_bn(model)
# benchmark with several iterations and take the average
pure_inf_time = 0
speed = []
for iteration in range(args.num_iters + args.num_warmup):
data = next(data_loader)
data['img'] = data['img'].to(args.device.lower())
data['img_metas'] = data['img_metas'].data[0]
if with_cuda:
torch.cuda.synchronize()
start_time = time.perf_counter()
with torch.no_grad():
model(return_loss=False, **data)
if with_cuda:
torch.cuda.synchronize()
elapsed = time.perf_counter() - start_time
if iteration >= args.num_warmup:
pure_inf_time += elapsed
speed.append(1 / elapsed)
speed_mean = np.mean(speed)
speed_std = np.std(speed)
split_line = '=' * 30
result = f'{split_line}\nModel config:{cfg_file}\n' \
f'Device: {args.device}\n' \
f'Batch size: {args.batch_size}\n' \
f'Overall average speed: {speed_mean:.2f} \u00B1 ' \
f'{speed_std:.2f} items / s\n' \
f'Total iters: {args.num_iters}\n'\
f'Total time: {pure_inf_time:.2f} s \n{split_line}\n'\
print(result)
results.append(result)
print('!!!Please be cautious if you use the results in papers. '
'You may need to check if all ops are included and verify that the '
'speed computation is correct.')
with open(osp.join(args.root_work_dir, 'inference_speed.txt'), 'w') as f:
for res in results:
f.write(res)
def train_model(model,
dataset,
cfg,
distributed=False,
validate=False,
timestamp=None,
meta=None):
"""Train model entry function.
Args:
model (nn.Module): The model to be trained.
dataset (Dataset): Train dataset.
cfg (dict): The config dict for training.
distributed (bool): Whether to use distributed training.
Default: False.
validate (bool): Whether to do evaluation. Default: False.
timestamp (str | None): Local time for runner. Default: None.
meta (dict | None): Meta dict to record some important information.
Default: None
"""
logger = get_root_logger(cfg.log_level)
# prepare data loaders
dataset = dataset if isinstance(dataset, (list, tuple)) else [dataset]
dataloader_setting = dict(
samples_per_gpu=cfg.data.get('samples_per_gpu', {}),
workers_per_gpu=cfg.data.get('workers_per_gpu', {}),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
seed=cfg.seed)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('train_dataloader', {}))
data_loaders = [
build_dataloader(ds, **dataloader_setting) for ds in dataset
]
# put model on gpus
if distributed:
find_unused_parameters = cfg.get('find_unused_parameters', True)
# Sets the `find_unused_parameters` parameter in
# torch.nn.parallel.DistributedDataParallel
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False,
find_unused_parameters=find_unused_parameters)
else:
model = MMDataParallel(model.cuda(cfg.gpu_ids[0]),
device_ids=cfg.gpu_ids)
# build runner
optimizer = build_optimizer(model, cfg.optimizer)
runner = EpochBasedRunner(model,
optimizer=optimizer,
work_dir=cfg.work_dir,
logger=logger,
meta=meta)
# an ugly workaround to make .log and .log.json filenames the same
runner.timestamp = timestamp
# fp16 setting
fp16_cfg = cfg.get('fp16', None)
if fp16_cfg is not None:
optimizer_config = Fp16OptimizerHook(**cfg.optimizer_config,
**fp16_cfg,
distributed=distributed)
elif distributed and 'type' not in cfg.optimizer_config:
optimizer_config = OptimizerHook(**cfg.optimizer_config)
else:
optimizer_config = cfg.optimizer_config
# register hooks
runner.register_training_hooks(cfg.lr_config, optimizer_config,
cfg.checkpoint_config, cfg.log_config,
cfg.get('momentum_config', None))
if distributed:
runner.register_hook(DistSamplerSeedHook())
# register eval hooks
if validate:
eval_cfg = cfg.get('evaluation', {})
val_dataset = build_dataset(cfg.data.val, dict(test_mode=True))
dataloader_setting = dict(
# samples_per_gpu=cfg.data.get('samples_per_gpu', {}),
samples_per_gpu=1,
workers_per_gpu=cfg.data.get('workers_per_gpu', {}),
# cfg.gpus will be ignored if distributed
num_gpus=len(cfg.gpu_ids),
dist=distributed,
shuffle=False)
dataloader_setting = dict(dataloader_setting,
**cfg.data.get('val_dataloader', {}))
val_dataloader = build_dataloader(val_dataset, **dataloader_setting)
eval_hook = DistEvalHook if distributed else EvalHook
runner.register_hook(eval_hook(val_dataloader, **eval_cfg))
if cfg.resume_from:
runner.resume(cfg.resume_from)
elif cfg.load_from:
runner.load_checkpoint(cfg.load_from)
runner.run(data_loaders, cfg.workflow, cfg.total_epochs)
