dataset_names=(args.dataset3, ),
preprocessing=False,
train=False,
kfold=args.kfold)
valid_dataset = valid_dataset1.__add__(valid_dataset2).__add__(valid_dataset3)
train_dataloader = AudioDataLoader(dataset=train_dataset,
batch_size=config.experiment['batch_size'],
drop_last=False,
shuffle=True)
valid_dataloader = AudioDataLoader(dataset=valid_dataset,
batch_size=config.experiment['batch_size'],
drop_last=False)
# Model and Optimizer
if args.model == 'cnn':
model = CNN(config=config.model).to(device)
elif args.model == 'crnn':
model = CRNN(config=config.model).to(device)
elif args.model == 'btc':
model = BTC_model(config=config.model).to(device)
else:
raise NotImplementedError
optimizer = optim.Adam(model.parameters(),
lr=config.experiment['learning_rate'],
weight_decay=config.experiment['weight_decay'],
betas=(0.9, 0.98),
eps=1e-9)
# Make asset directory
if not os.path.exists(os.path.join(asset_path, ckpt_path)):
os.makedirs(os.path.join(asset_path, ckpt_path))
# Data loader
train_dataset1 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset1,), num_workers=20, preprocessing=False, train=True, kfold=args.kfold)
train_dataset2 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset2,), num_workers=20, preprocessing=False, train=True, kfold=args.kfold)
train_dataset3 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset3,), num_workers=20, preprocessing=False, train=True, kfold=args.kfold)
train_dataset = train_dataset1.__add__(train_dataset2).__add__(train_dataset3)
valid_dataset1 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset1,), preprocessing=False, train=False, kfold=args.kfold)
valid_dataset2 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset2,), preprocessing=False, train=False, kfold=args.kfold)
valid_dataset3 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset3,), preprocessing=False, train=False, kfold=args.kfold)
valid_dataset = valid_dataset1.__add__(valid_dataset2).__add__(valid_dataset3)
train_dataloader = AudioDataLoader(dataset=train_dataset, batch_size=config.experiment['batch_size'], drop_last=False, shuffle=True)
valid_dataloader = AudioDataLoader(dataset=valid_dataset, batch_size=config.experiment['batch_size'], drop_last=False)
# Model and Optimizer
if args.pre_model == 'cnn':
pre_model = CNN(config=config.model).to(device)
elif args.pre_model == 'crnn':
pre_model = CRNN(config=config.model).to(device)
elif args.pre_model == 'btc':
pre_model = BTC_model(config=config.model).to(device)
else: raise NotImplementedError
if args.pre_model == 'cnn':
if args.voca == False:
if args.kfold == 0:
load_ckpt_file_name = 'idx_0_%03d.pth.tar'
load_restore_epoch = 10
else:
if args.kfold == 0:
load_ckpt_file_name = 'idx_1_%03d.pth.tar'
load_restore_epoch = 10
valid_dataset = valid_dataset.__add__(valid_dataset2)
if args.dataset3 != 'NA':
train_dataset3 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset3,), \
num_workers=20, preprocessing=False, train=True, kfold=args.kfold)
valid_dataset3 = AudioDataset(config, root_dir=config.path['root_path'], dataset_names=(args.dataset3,), preprocessing=False, train=False, kfold=args.kfold)
train_dataset = train_dataset.__add__(train_dataset3)
valid_dataset = valid_dataset.__add__(valid_dataset3)
train_dataloader = AudioDataLoader(dataset=train_dataset, batch_size=config.experiment['batch_size'], drop_last=False, shuffle=True)
valid_dataloader = AudioDataLoader(dataset=valid_dataset, batch_size=config.experiment['batch_size'], drop_last=False)
# Model and Optimizer
if args.model == 'cnn':
model = CNN(config=config.model).to(device)
elif args.model == 'crnn':
model = CRNN(config=config.model).to(device)
elif args.model == 'btc':
model = BTC_model(config=config.model).to(device)
else: raise NotImplementedError
optimizer = optim.Adam(model.parameters(), lr=config.experiment['learning_rate'], weight_decay=config.experiment['weight_decay'], betas=(0.9, 0.98), eps=1e-9)
# Make asset directory
if not os.path.exists(os.path.join(asset_path, ckpt_path)):
os.makedirs(os.path.join(asset_path, ckpt_path))
os.makedirs(os.path.join(asset_path, result_path))
# Load model
if os.path.isfile(os.path.join(asset_path, ckpt_path, ckpt_file_name % restore_epoch)):
checkpoint = torch.load(os.path.join(asset_path, ckpt_path, ckpt_file_name % restore_epoch))