help="Use cuda?")
parser.add_argument("--gpus", default=1, type=int, help="nums of gpu to use")
parser.add_argument("--threads",
default=8,
type=int,
help="Number of threads for data loader to use.")
parser.add_argument("--fixrandomseed",
default=False,
help="train with fix random seed")
# Pretrain model setting
parser.add_argument("--resume", type=str, help="Path to checkpoint.")
# Dataloader setting
parser.add_argument("--train",
default=["../dataset/NTIRE2018"],
type=str,
nargs='+',
help="path of training dataset")
parser.add_argument("--val",
default=["../dataset/NTIRE2018_VAL"],
type=str,
nargs='+',
help="path of validation dataset")
if __name__ == "__main__":
import utils
opt = parser.parse_args()
utils.details(opt, None)
ihazeparser.add_argument('--gt', help='Input directory for Clear Image', type=str,
default="/media/disk1/EdwardLee/dataset/IndoorTrainGT")
ihazeparser.add_argument('--output', help='Output directory', type=str,
default="/media/disk1/EdwardLee/IndoorTrain")
# O-Haze
ohazeparser = domainparser.add_parser("O-Haze")
ohazeparser.add_argument('--hazy', help='Input directory for Hazy Image', type=str,
default="/media/disk1/EdwardLee/dataset/OutdoorTrainHazy")
ohazeparser.add_argument('--gt', help='Input directory for Clear Image', type=str,
default="/media/disk1/EdwardLee/dataset/OutdoorTrainGT")
ohazeparser.add_argument('--output', help='Output directory', type=str,
default="/media/disk1/EdwardLee/OutdoorTrain")
args = parser.parse_args()
utils.details(args)
# --------------------------------------------------------- #
# Check folders here, make the directories if don't exist. #
# 0. The image #
# 1. Root folder #
# 1.1 gt #
# 1.2 hazy #
# --------------------------------------------------------- #
if not os.path.exists(args.hazy):
raise IOError("File doesn't not exist: {}".format(args.hazy))
if not os.path.exists(args.gt):
raise IOError("File doesn't not exist: {}".format(args.gt))
# 1 Root Folder
for path in opt.train:
if not os.path.exists(path):
raise ValueError("{} doesn't exist".format(path))
# Check validation dataset directory
for path in opt.val:
if not os.path.exists(path):
raise ValueError("{} doesn't exist".format(path))
# Make checkpoint storage directory
name = "{}_{}".format(opt.tag, date.today().strftime("%Y%m%d"))
os.makedirs(os.path.join(opt.checkpoints, name), exist_ok=True)
# Copy the code of model to logging file
if os.path.exists(os.path.join(opt.detail, name, 'model')):
shutil.rmtree(os.path.join(opt.detail, name, 'model'))
if os.path.exists(os.path.join(opt.checkpoints, name, 'model')):
shutil.rmtree(os.path.join(opt.checkpoints, name, 'model'))
shutil.copytree('./model', os.path.join(opt.detail, name, 'model'))
shutil.copytree('./model', os.path.join(opt.checkpoints, name, 'model'))
shutil.copyfile(__file__, os.path.join(opt.detail, name, os.path.basename(__file__)))
# Show Detail
print('==========> Training setting')
utils.details(opt, os.path.join(opt.detail, name, 'args.txt'))
# Execute main process
main(opt)
#loss
labels = concat_labels.to(DEVICE)
loss = criterion(log_probs, labels, tuple(lengths), tuple(len_of_labels))
#count accuracy
nl = loss.item()
print(" ["+str(epoch)+","+str(vi)+"] loss: "+str(nl))
if not math.isnan(nl):
vloss += nl
#vavdis += predictor.evaluateError(output, labels, len_of_labels)
vtotal += output.shape[0]
vi += 1
_end_time = time.time()
details((_end_time - _begin_time), ave_loss/i, avg_dis/total, vloss/vi, vavdis/vtotal)
_begin_time = time.time()
torch.save(model.state_dict(),"models/model-"+str(int(_begin_time))+"-"+str(int(vavdis/vtotal))+".pt")
result = []
model.eval()
test_loader = loader("data/transformed_test_data.npy")
for voice,lengths in test_loader:
#evaluate
output = model(voice)
preds = predictor.predict(output)
result = result + preds