def parse_args():
parser = argparse.ArgumentParser(description='Train the SZO nowcasting model')
parser.add_argument('--batch_size', dest='batch_size', help="batchsize of the training process",
default=None, type=int)
parser.add_argument('--cfg', dest='cfg_file', help='Optional configuration file', default=None, type=str)
parser.add_argument('--save_dir', help='The saving directory', required=True, type=str)
parser.add_argument('--ctx', dest='ctx', help='Running Context. E.g `--ctx gpu` or `--ctx gpu0,gpu1` or `--ctx cpu`',
type=str, default='gpu')
parser.add_argument('--resume', dest='resume', action='store_true', default=False)
parser.add_argument('--resume_param_only', dest='resume_param_only', action='store_true', default=False)
parser.add_argument('--lr', dest='lr', help='learning rate', default=None, type=float)
parser.add_argument('--wd', dest='wd', help='weight decay', default=None, type=float)
parser.add_argument('--grad_clip', dest='grad_clip', help='gradient clipping threshold',
default=None, type=float)
args = parser.parse_args()
args.ctx = parse_ctx(args.ctx)
if args.cfg_file is not None:
cfg_from_file(args.cfg_file, target=cfg.MODEL)
if args.batch_size is not None:
cfg.MODEL.TRAIN.BATCH_SIZE = args.batch_size
if args.lr is not None:
cfg.MODEL.TRAIN.LR = args.lr
if args.wd is not None:
cfg.MODEL.TRAIN.WD = args.wd
if args.grad_clip is not None:
cfg.MODEL.TRAIN.GRAD_CLIP = args.grad_clip
if args.wd is not None:
cfg.MODEL.TRAIN.WD = args.wd
cfg.MODEL.SAVE_DIR = args.save_dir
logging.info(args)
return args
def parse_args():
parser = argparse.ArgumentParser(
description='Train the Meteotn nowcasting model')
parser.add_argument('--batch_size',
dest='batch_size',
help="batchsize of the training process",
default=None,
type=int)
parser.add_argument('--cfg',
dest='cfg_file',
help='Configuration file',
required=True,
type=str)
parser.add_argument('--save_dir',
help='The saving directory',
required=True,
type=str)
parser.add_argument(
'--data_dir',
help=
'The data directory with hdf_archives folder, hdf_metadata.csv and mask.png',
required=True,
type=str)
parser.add_argument(
'--data_csv',
help=
'alternate metadata CSV file (default: [data_dir]/hdf_metadata.csv)',
default=None,
type=str)
parser.add_argument(
'--date_start',
help='Start date to filter the sequences (e.g. 2010-12-31)',
default=None,
type=lambda s: datetime.strptime(s, '%Y-%m-%d'))
parser.add_argument(
'--date_end',
help='End date to filter the sequences (e.g. 2016-12-31)',
default=None,
type=lambda s: datetime.strptime(s, '%Y-%m-%d'))
parser.add_argument(
'--ctx',
default='cpu',
help='Running Context. (default: %(default)s): `--ctx gpu` '
'or `--ctx gpu0,gpu1` for GPU(s). `--ctx cpu` for CPU')
parser.add_argument('--threshold',
dest='threshold',
help='rainfall filter threshold (default 0.28)',
default=None,
type=float)
args = parser.parse_args()
args.ctx = parse_ctx(args.ctx)
if args.cfg_file is not None:
cfg_from_file(args.cfg_file, target=cfg.MODEL)
if args.batch_size is not None:
cfg.MODEL.TRAIN.BATCH_SIZE = args.batch_size
if args.threshold is not None:
cfg.HKO.ITERATOR.FILTER_RAINFALL_THRESHOLD = float(args.threshold)
cfg.MODEL.SAVE_DIR = args.save_dir
logging.info(args)
return args
def __init__(self, model_dir: str, model_iter: int, model_cfg: str,
batch_size: int, ctx: list):
self.batch_size = batch_size
self.ctx = ctx
self.model_dir = model_dir
self.model_iter = model_iter
cfg_from_file(model_cfg, target=cfg.MODEL)
assert cfg.MODEL.FRAME_STACK == 1 and cfg.MODEL.FRAME_SKIP == 1
assert len(self.model_dir) > 0
hko_factory = RadarNowcastingFactory(batch_size=self.batch_size,
in_seq_len=cfg.MODEL.IN_LEN,
out_seq_len=cfg.MODEL.OUT_LEN)
self.encoder_net, self.forecaster_net, t_loss_net = encoder_forecaster_build_networks(
factory=hko_factory, context=self.ctx, for_finetune=True)
self.encoder_net.summary()
self.forecaster_net.summary()
t_loss_net.summary()
load_encoder_forecaster_params(load_dir=self.model_dir,
load_iter=self.model_iter,
encoder_net=self.encoder_net,
forecaster_net=self.forecaster_net)
self.states = EncoderForecasterStates(factory=hko_factory,
ctx=self.ctx[0])
def parse_mode_args(args):
args.ctx = parse_ctx(args.ctx)
if args.cfg_file:
cfg_from_file(args.cfg_file, target=cfg)
# Parameter loading
if args.test or cfg.MODEL.TESTING:
cfg.MODEL.TESTING = True
args.resume = True
if args.resume:
cfg.MODEL.RESUME = True
if args.load_dir:
cfg.MODEL.LOAD_DIR = args.load_dir
if args.load_iter:
cfg.MODEL.LOAD_ITER = args.load_iter
def parse_args():
parser = argparse.ArgumentParser(
description='Test the MovingMNIST++ dataset')
parser.add_argument('--batch_size',
dest='batch_size',
help="batchsize of the testing process",
default=None,
type=int)
parser.add_argument('--cfg',
dest='cfg_file',
help='Testing configuration file',
default=None,
type=str)
parser.add_argument('--load_dir',
help='The loading directory',
default=None,
type=str)
parser.add_argument('--load_iter',
help='The loading iterator',
default=None,
type=int)
parser.add_argument('--save_dir',
help='The saving directory',
required=True,
type=str)
parser.add_argument(
'--ctx',
dest='ctx',
help=
'Running Context. E.g `--ctx gpu` or `--ctx gpu0,gpu1` or `--ctx cpu`',
type=str,
default='gpu')
args = parser.parse_args()
args.ctx = parse_ctx(args.ctx)
if args.cfg_file is not None:
cfg_from_file(args.cfg_file, target=cfg)
if args.load_dir is not None:
cfg.MODEL.LOAD_DIR = args.load_dir
if args.load_iter is not None:
cfg.MODEL.LOAD_ITER = args.load_iter
cfg.MODEL.SAVE_DIR = args.save_dir
logging.info(args)
return args
def parse_args():
parser = argparse.ArgumentParser(description='Test the HKO nowcasting model')
parser.add_argument('--cfg', dest='cfg_file', help='Optional configuration file', type=str)
parser.add_argument('--load_dir', help='The directory to load the model', default=None, type=str)
parser.add_argument('--load_iter', help='The iterator to load', default=-1, type=int)
parser.add_argument('--save_dir', help='The saving directory', required=True, type=str)
parser.add_argument('--ctx', dest='ctx', help='Running Context. E.g `--ctx gpu` or `--ctx gpu0,gpu1` or `--ctx cpu`',
type=str, default='gpu')
parser.add_argument('--finetune', dest='finetune', help='Whether to do online finetuning',
default=None, type=int)
parser.add_argument('--finetune_min_mse', dest='finetune_min_mse', help='Minimum error for finetuning',
default=None, type=float)
parser.add_argument('--mode', dest='mode', help='Whether to used fixed setting or online setting',
required=True, type=str)
parser.add_argument('--dataset', dest='dataset', help='Whether to used the test set or the validation set',
default="test", type=str)
parser.add_argument('--lr', dest='lr', help='learning rate', default=None, type=float)
parser.add_argument('--wd', dest='wd', help='weight decay', default=None, type=float)
parser.add_argument('--grad_clip', dest='grad_clip', help='gradient clipping threshold',
default=None, type=float)
args = parser.parse_args()
args.ctx = parse_ctx(args.ctx)
if args.cfg_file is not None:
cfg_from_file(args.cfg_file, target=cfg.MODEL)
if args.load_dir is not None:
cfg.MODEL.LOAD_DIR = args.load_dir
if args.load_iter != -1:
cfg.MODEL.LOAD_ITER = args.load_iter
if args.lr is not None:
cfg.MODEL.TEST.ONLINE.LR = args.lr
if args.wd is not None:
cfg.MODEL.TEST.ONLINE.WD = args.wd
if args.grad_clip is not None:
cfg.MODEL.TEST.ONLINE.GRAD_CLIP = args.grad_clip
if args.mode is not None:
cfg.MODEL.TEST.MODE = args.mode
if args.finetune is not None:
cfg.MODEL.TEST.FINETUNE = (args.finetune != 0)
if args.finetune_min_mse is not None:
cfg.MODEL.TEST.ONLINE.FINETUNE_MIN_MSE = args.finetune_min_mse
cfg.MODEL.SAVE_DIR = args.save_dir
logging.info(args)
return args
def parse_args():
parser = argparse.ArgumentParser(
description="Test the SST nowcasting model")
parser.add_argument("--cfg",
dest="cfg_file",
help="Optional configuration file",
type=str)
parser.add_argument("--load_dir",
help="The directory to load the model",
default=None,
type=str)
parser.add_argument("--load_iter",
help="The iterator to load",
default=-1,
type=int)
parser.add_argument("--save_dir",
help="The saving directory",
required=True,
type=str)
parser.add_argument(
"--ctx",
dest="ctx",
help=
"Running Context. E.g `--ctx gpu` or `--ctx gpu0,gpu1` or `--ctx cpu`",
type=str,
default="gpu",
)
parser.add_argument(
"--finetune",
dest="finetune",
help="Whether to do online finetuning",
default=None,
type=int,
)
parser.add_argument(
"--finetune_min_mse",
dest="finetune_min_mse",
help="Minimum error for finetuning",
default=None,
type=float,
)
parser.add_argument(
"--mode",
dest="mode",
help="Whether to used fixed setting or online setting",
required=True,
type=str,
)
parser.add_argument(
"--dataset",
dest="dataset",
help="Whether to used the test set or the validation set",
default="test",
type=str,
)
parser.add_argument("--lr",
dest="lr",
help="learning rate",
default=None,
type=float)
parser.add_argument("--wd",
dest="wd",
help="weight decay",
default=None,
type=float)
parser.add_argument(
"--grad_clip",
dest="grad_clip",
help="gradient clipping threshold",
default=None,
type=float,
)
args = parser.parse_args()
args.ctx = parse_ctx(args.ctx)
if args.cfg_file is not None:
cfg_from_file(args.cfg_file, target=cfg.MODEL)
if args.load_dir is not None:
cfg.MODEL.LOAD_DIR = args.load_dir
if args.load_iter != -1:
cfg.MODEL.LOAD_ITER = args.load_iter
if args.lr is not None:
cfg.MODEL.TEST.ONLINE.LR = args.lr
if args.wd is not None:
cfg.MODEL.TEST.ONLINE.WD = args.wd
if args.grad_clip is not None:
cfg.MODEL.TEST.ONLINE.GRAD_CLIP = args.grad_clip
if args.mode is not None:
cfg.MODEL.TEST.MODE = args.mode
if args.finetune is not None:
cfg.MODEL.TEST.FINETUNE = args.finetune != 0
if args.finetune_min_mse is not None:
cfg.MODEL.TEST.ONLINE.FINETUNE_MIN_MSE = args.finetune_min_mse
cfg.MODEL.SAVE_DIR = args.save_dir
logging.info(args)
return args
def parse_args():
parser = argparse.ArgumentParser(
description="Train the SST nowcasting model")
parser.add_argument(
"--batch_size",
dest="batch_size",
help="batchsize of the training process",
default=None,
type=int,
)
parser.add_argument(
"--cfg",
dest="cfg_file",
help="Optional configuration file",
default=None,
type=str,
)
parser.add_argument("--save_dir",
help="The saving directory",
required=True,
type=str)
parser.add_argument(
"--ctx",
dest="ctx",
help=
"Running Context. E.g `--ctx gpu` or `--ctx gpu0,gpu1` or `--ctx cpu`",
type=str,
default="gpu",
)
parser.add_argument("--lr",
dest="lr",
help="learning rate",
default=None,
type=float)
parser.add_argument("--wd",
dest="wd",
help="weight decay",
default=None,
type=float)
parser.add_argument(
"--grad_clip",
dest="grad_clip",
help="gradient clipping threshold",
default=None,
type=float,
)
args = parser.parse_args()
args.ctx = parse_ctx(args.ctx)
if args.cfg_file is not None:
cfg_from_file(args.cfg_file, target=cfg.MODEL)
if args.batch_size is not None:
cfg.MODEL.TRAIN.BATCH_SIZE = args.batch_size
if args.lr is not None:
cfg.MODEL.TRAIN.LR = args.lr
if args.wd is not None:
cfg.MODEL.TRAIN.WD = args.wd
if args.grad_clip is not None:
cfg.MODEL.TRAIN.GRAD_CLIP = args.grad_clip
if args.wd is not None:
cfg.MODEL.TRAIN.WD = args.wd
cfg.MODEL.SAVE_DIR = args.save_dir
logging.info(args)
return args