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
# update configs according to CLI args
if args.work_dir is not None:
cfg.work_dir = args.work_dir
cfg.model.pretrained = None # ensure to use checkpoint rather than pretraining
# check memcached package exists
if importlib.util.find_spec('mc') is None:
traverse_replace(cfg, 'memcached', False)
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
if args.launcher == 'slurm':
cfg.dist_params['port'] = args.port
init_dist(args.launcher, **cfg.dist_params)
# logger
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, 'test_{}.log'.format(timestamp))
logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
# build the dataloader
dataset = build_dataset(cfg.data.val)
data_loader = build_dataloader(
dataset,
imgs_per_gpu=cfg.data.imgs_per_gpu,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
# build the model and load checkpoint
model = build_model(cfg.model)
load_checkpoint(model, args.checkpoint, map_location='cpu')
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) # dict{key: np.ndarray}
rank, _ = get_dist_info()
if rank == 0:
for name, val in outputs.items():
dataset.evaluate(
torch.from_numpy(val), name, logger, topk=(1, 5))
def init_weights(self, pretrained=None):
if isinstance(pretrained, str):
logger = get_root_logger()
load_checkpoint(self, pretrained, strict=True, logger=logger)
elif pretrained is None:
pass
else:
raise TypeError('pretrained must be a str or None')
def train_model(model,
dataset,
cfg,
distributed=False,
timestamp=None,
meta=None):
logger = get_root_logger(cfg.log_level)
# start training
if distributed:
_dist_train(
model, dataset, cfg, logger=logger, timestamp=timestamp, meta=meta)
else:
_non_dist_train(
model, dataset, cfg, logger=logger, timestamp=timestamp, meta=meta)
def init_weights(self, pretrained=None):
if isinstance(pretrained, str):
logger = get_root_logger()
load_checkpoint(self, pretrained, strict=True, logger=logger)
elif pretrained is None:
for m in self.modules():
if isinstance(m, nn.Conv2d):
kaiming_init(m, mode='fan_in', nonlinearity='relu')
elif isinstance(m, (_BatchNorm, nn.GroupNorm)):
constant_init(m, 1)
if self.zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
constant_init(m.norm3, 0)
elif isinstance(m, BasicBlock):
constant_init(m.norm2, 0)
else:
raise TypeError('pretrained must be a str or None')
def train_model(model,
dataset,
cfg,
distributed=False,
timestamp=None,
meta=None,
debug=False):
logger = get_root_logger(cfg.log_level)
# multiply the number of iters if we have update intervals set
if (cfg.get('by_iter', False) and cfg.get('optimizer_config', False) and cfg.optimizer_config.get('update_interval', 1) > 1):
print_log("Iter based runner with update_interval set. Multiplying iters by update_interval")
print_log(f"Before iters: {cfg.total_iters}")
cfg.total_iters = cfg.optimizer_config.update_interval * cfg.total_iters
print_log(f"After iters: {cfg.total_iters}")
# start training
if distributed and not debug:
_dist_train(
model, dataset, cfg, logger=logger, timestamp=timestamp, meta=meta)
else:
_non_dist_train(
model, dataset, cfg, logger=logger, timestamp=timestamp, meta=meta)
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
# update configs according to CLI args
if args.work_dir is not None:
cfg.work_dir = args.work_dir
layer_ind = [int(idx) for idx in args.layer_ind.split(',')]
cfg.model.backbone.out_indices = layer_ind
# checkpoint and pretrained are exclusive
assert cfg.model.pretrained == "random" or args.checkpoint is None, \
"Checkpoint and pretrained are exclusive."
# check memcached package exists
if importlib.util.find_spec('mc') is None:
for field in ['train', 'val', 'test']:
if hasattr(cfg.data, field):
getattr(cfg.data, field).data_source.memcached = False
# 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))
# logger
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, 'extract_{}.log'.format(timestamp))
logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
# build the dataloader
dataset_cfg = mmcv.Config.fromfile(args.dataset_config)
dataset = build_dataset(dataset_cfg.data.extract)
data_loader = build_dataloader(
dataset,
imgs_per_gpu=dataset_cfg.data.imgs_per_gpu,
workers_per_gpu=dataset_cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
# build the model and load checkpoint
model = build_model(cfg.model)
if args.checkpoint is not None:
logger.info("Use checkpoint: {} to extract features".format(
args.checkpoint))
load_checkpoint(model, args.checkpoint, map_location='cpu')
elif args.pretrained != "random":
logger.info('Use pretrained model: {} to extract features'.format(
args.pretrained))
else:
logger.info('No checkpoint or pretrained is give, use random init.')
if not distributed:
model = MMDataParallel(model, device_ids=[0])
else:
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
# build extraction processor
extractor = ExtractProcess(pool_type='specified',
backbone='resnet50',
layer_indices=layer_ind)
# run
outputs = extractor.extract(model, data_loader, distributed=distributed)
rank, _ = get_dist_info()
mmcv.mkdir_or_exist("{}/features/".format(args.work_dir))
if rank == 0:
for key, val in outputs.items():
split_num = len(dataset_cfg.split_name)
split_at = dataset_cfg.split_at
for ss in range(split_num):
output_file = "{}/features/{}_{}.npy".format(
args.work_dir, dataset_cfg.split_name[ss], key)
if ss == 0:
np.save(output_file, val[:split_at[0]])
elif ss == split_num - 1:
np.save(output_file, val[split_at[-1]:])
else:
np.save(output_file, val[split_at[ss - 1]:split_at[ss]])
def main():
args = parse_args()
cfg = Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
# update configs according to CLI args
if args.work_dir is not None:
cfg.work_dir = args.work_dir
if args.resume_from is not None:
cfg.resume_from = args.resume_from
cfg.gpus = args.gpus
# my
if args.imgs_per_gpu is not None:
cfg.data.imgs_per_gpu = args.imgs_per_gpu
if args.val_imgs_per_gpu is not None:
cfg.custom_hooks[0].imgs_per_gpu = args.val_imgs_per_gpu
#
# check memcached package exists
if importlib.util.find_spec('mc') is None:
traverse_replace(cfg, 'memcached', False)
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
assert cfg.model.type not in \
['DeepCluster', 'MOCO', 'SimCLR', 'ODC', 'NPID'], \
"{} does not support non-dist training.".format(cfg.model.type)
else:
distributed = True
if args.launcher == 'slurm':
cfg.dist_params['port'] = args.port
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, 'train_{}.log'.format(timestamp))
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([('{}: {}'.format(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('Distributed training: {}'.format(distributed))
logger.info('Config:\n{}'.format(cfg.text))
# set random seeds
if args.seed is not None:
logger.info('Set random seed to {}, deterministic: {}'.format(
args.seed, args.deterministic))
set_random_seed(args.seed, deterministic=args.deterministic)
cfg.seed = args.seed
meta['seed'] = args.seed
if args.pretrained is not None:
assert isinstance(args.pretrained, str)
cfg.model.pretrained = args.pretrained
model = build_model(cfg.model)
datasets = [build_dataset(cfg.data.train)]
assert len(cfg.workflow) == 1, "Validation is called by hook."
if cfg.checkpoint_config is not None:
# save openselfsup version, config file content and class names in
# checkpoints as meta data
cfg.checkpoint_config.meta = dict(openselfsup_version=__version__,
config=cfg.text)
# add an attribute for visualization convenience
train_model(model,
datasets,
cfg,
distributed=distributed,
timestamp=timestamp,
meta=meta)
def main():
print(f"Using num gpus: {torch.cuda.device_count()}")
args = parse_args()
cfg = Config.fromfile(args.config)
if args.local_rank == 0:
wandb.init(config=cfg.model)
wandb.config.update(cfg.data)
wandb.config.update(
{"pipelines": ','.join([p.type for p in cfg.data.train.pipeline])})
wandb.config.update({"epochs": cfg.total_epochs})
wandb.config.update({"dataset_size": cfg.dataset_size})
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
# update configs according to CLI args
if wandb.run is not None:
# save to wandb run dir for tracking and saving the models
cfg.work_dir = wandb.run.dir
elif args.work_dir is not None:
cfg.work_dir = args.work_dir
if args.resume_from is not None:
cfg.resume_from = args.resume_from
cfg.gpus = args.gpus
# check memcached package exists
if importlib.util.find_spec('mc') is None:
traverse_replace(cfg, 'memcached', False)
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
if not args.debug:
assert cfg.model.type not in \
['DeepCluster', 'MOCO', 'SimCLR', 'ODC', 'NPID'], \
"{} does not support non-dist training unless debugging (use --debug flag).".format(
cfg.model.type)
else:
distributed = True
if args.launcher == 'slurm':
cfg.dist_params['port'] = args.port
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, 'train_{}.log'.format(timestamp))
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([('{}: {}'.format(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('Distributed training: {}'.format(distributed))
logger.info('Config:\n{}'.format(cfg.text))
# set random seeds
if args.seed is not None:
logger.info('Set random seed to {}, deterministic: {}'.format(
args.seed, args.deterministic))
set_random_seed(args.seed, deterministic=args.deterministic)
cfg.seed = args.seed
meta['seed'] = args.seed
if args.pretrained is not None:
assert isinstance(args.pretrained, str)
cfg.model.pretrained = args.pretrained
model = build_model(cfg.model)
if args.local_rank == 0:
print(model)
if args.debug:
logger.info(
"DEBUGGING enabled, setting batch size to 64 to allow 1 gpu debugging"
)
cfg.data.batch_size = 64
model.set_debug()
datasets = [build_dataset(cfg.data.train)]
assert len(cfg.workflow) == 1, "Validation is called by hook."
if cfg.checkpoint_config is not None:
# save openselfsup version, config file content and class names in
# checkpoints as meta data
cfg.checkpoint_config.meta = dict(openselfsup_version=__version__,
config=cfg.text)
if args.local_rank == 0:
wandb.watch(model)
# add an attribute for visualization convenience
train_model(model,
datasets,
cfg,
distributed=distributed,
timestamp=timestamp,
meta=meta,
debug=args.debug)
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
# update configs according to CLI args
if args.work_dir is not None:
cfg.work_dir = args.work_dir
cfg.model.pretrained = None # ensure to use checkpoint rather than pretraining
# check memcached package exists
if importlib.util.find_spec('mc') is None:
traverse_replace(cfg, 'memcached', False)
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
if args.launcher == 'slurm':
cfg.dist_params['port'] = args.port
init_dist(args.launcher, **cfg.dist_params)
# logger
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, 'test_{}.log'.format(timestamp))
logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
# build the dataloader
dataset = build_dataset(cfg.data.val)
data_loader = build_dataloader(dataset,
imgs_per_gpu=cfg.data.imgs_per_gpu,
workers_per_gpu=cfg.data.workers_per_gpu,
dist=distributed,
shuffle=False)
# build the model and load checkpoint
model = build_model(cfg.model)
activations = defaultdict(list)
#idea from gist.github.com/Tushar-N/680633ec18f5cb4b47933da7d10902af
if args.layer_type == nn.Linear: #can save all activations
def save_activation(name, mod, inp, out):
activations[name].append(out.cpu())
else:
def save_activation(name, mod, inp, out):
activations[name] = [out.cpu()]
load_checkpoint(model, args.checkpoint, map_location='cpu')
for name, m in model.named_modules():
if type(m) == args.layer_type:
m.register_forward_hook(partial(save_activation, name))
if not distributed:
model = MMDataParallel(model, device_ids=[0])
outputs = single_gpu_test(model, data_loader)
else:
raise NotImplementedError(
"Distributed Data Parallel does not register hooks.")
model = MMDistributedDataParallel(
model.cuda(),
device_ids=[torch.cuda.current_device()],
broadcast_buffers=False)
outputs = multi_gpu_test(model, data_loader) # dict{key: np.ndarray}
activations = {
name: torch.cat(outputs, 0)
for name, outputs in activations.items()
}
act_file = osp.join(cfg.work_dir, "model_acts")
np.savez(act_file, **activations)
