# define model
mt = MusicTransformer(embedding_dim=config.embedding_dim,
vocab_size=config.vocab_size,
num_layer=config.num_layers,
max_seq=config.max_seq,
dropout=config.dropout,
debug=config.debug,
loader_path=config.load_path)
if config.init_ckpt is not None:
mt.load_state_dict(torch.load(config.init_ckpt))
print("Weights from %s loaded" % config.init_ckpt)
mt.to(config.device)
opt = optim.Adam(mt.parameters(), lr=0, betas=(0.9, 0.98),
eps=1e-9) # setting rate inside
scheduler = CustomSchedule(
config.embedding_dim, optimizer=opt) # custom implementation of rate decay
# multi-GPU set
if torch.cuda.device_count() > 1:
single_mt = mt
mt = torch.nn.DataParallel(mt, output_device=torch.cuda.device_count() - 1)
else:
single_mt = mt
# init metric set
metric_set = MetricsSet({
'accuracy':
CategoricalAccuracy(),
'loss':
SmoothCrossEntropyLoss(config.label_smooth, config.vocab_size,
config.pad_token),
print(dataset)
# load model
learning_rate = config.l_r
# define model
mt = MusicTransformer(embedding_dim=config.embedding_dim,
vocab_size=config.vocab_size,
num_layer=config.num_layers,
max_seq=config.max_seq,
dropout=config.dropout,
debug=config.debug,
loader_path=config.load_path)
mt.to(config.device)
opt = optim.Adam(mt.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9)
scheduler = CustomSchedule(config.embedding_dim, optimizer=opt)
# multi-GPU set
if torch.cuda.device_count() > 1:
single_mt = mt
mt = torch.nn.DataParallel(mt, output_device=torch.cuda.device_count() - 1)
else:
single_mt = mt
# init metric set
metric_set = MetricsSet({
'accuracy':
CategoricalAccuracy(),
'loss':
SmoothCrossEntropyLoss(config.label_smooth, config.vocab_size,
config.pad_token),
# load model
learning_rate = config.l_r
# define model
mt = MusicTransformer(
embedding_dim=config.embedding_dim,
vocab_size=config.vocab_size,
num_layer=config.num_layers,
max_seq=config.max_seq,
dropout=config.dropout,
debug=config.debug, loader_path=config.load_path
)
mt.to(config.device)
opt = optim.Adam(mt.parameters(), lr=0, betas=(0.9, 0.98), eps=1e-9)
scheduler = CustomSchedule(config.embedding_dim, optimizer=opt)
# multi-GPU set
if torch.cuda.device_count() > 1:
single_mt = mt
mt = torch.nn.DataParallel(mt, output_device=torch.cuda.device_count()-1)
else:
single_mt = mt
# init metric set
metric_set = MetricsSet({
'accuracy': CategoricalAccuracy(),
'loss': SmoothCrossEntropyLoss(config.label_smooth, config.vocab_size, config.pad_token),
'bucket': LogitsBucketting(config.vocab_size)
})