def normalize_uint8_image(image, config='default'):
from normalize_config import get_normalize_config
mean, std = get_normalize_config(config)
image = image - mean
if std is not None:
image /= std
return image
def get_val_data_iterator(args,
comm,
channels,
spatial_size=(224, 224),
norm_config='default'):
# Pipelines and Iterators for validation
mean, std = get_normalize_config(norm_config)
if std is None:
std = [1., 1., 1.]
pad_output = get_pad_output_by_channels(channels)
val_pipe = ValPipeline(args.batch_size,
args.dali_num_threads,
comm.rank,
args.val_dir,
args.val_list,
args.dali_nvjpeg_memory_padding,
seed=comm.rank + 1,
device_id=int(comm.ctx.device_id),
num_shards=comm.n_procs,
channel_last=args.channel_last,
spatial_size=spatial_size,
dtype=args.type_config,
mean=list(mean),
std=list(std),
pad_output=pad_output)
vdata = dali_iterator.DaliIterator(val_pipe)
vdata.size = int_div_ceil(val_pipe.epoch_size("Reader"), comm.n_procs)
return vdata
def get_train_data_iterator(args, comm, channels, norm_config='default'):
# Pipelines and Iterators for training
mean, std = get_normalize_config(norm_config)
if std is None:
std = [1., 1., 1.]
pad_output = get_pad_output_by_channels(channels)
train_pipe = TrainPipeline(args.batch_size,
args.dali_num_threads,
comm.rank,
args.train_dir,
args.train_list,
args.dali_nvjpeg_memory_padding,
seed=comm.rank + 1,
num_shards=comm.n_procs,
channel_last=args.channel_last,
dtype=args.type_config,
mean=list(mean),
std=list(std),
pad_output=pad_output)
data = dali_iterator.DaliIterator(train_pipe)
data.size = int_div_ceil(train_pipe.epoch_size("Reader"), comm.n_procs)
return data