training_data = DataLoader(piano, batch_size=2, shuffle=True)
model = WaveNet().cuda()
optimizer = optim.Adam(model.parameters(), lr=1e-3)
scheduler = optim.lr_scheduler.StepLR(optimizer, 500, gamma=0.5)
for epoch in range(20000):
running_loss = 0.0
for index, (data, target, _) in enumerate(training_data):
data = Variable(data.type(torch.FloatTensor)).cuda()
logits = model(data)
logits = logits[:, :, :-1]
y = target[:, :, recp_field:].squeeze(1).cuda()
loss = F.cross_entropy(logits, y).cuda()
optimizer.zero_grad()
loss.backward()
optimizer.step()
scheduler.step()
running_loss += loss.item()
# print("[%d %.3f]" % (epoch + 1, running_loss / (index+1)))
if epoch % 100 == 99:
print("[%d %.3f]" % (epoch + 1, running_loss / (index + 1)))
torch.save(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}, 'checkpoint.pth')
print('finished')
train_loader = DataLoader(train_data,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers)
print("Number of training inputs:", len(train_data))
print('Bach training')
losses, predictions = train(model, optimizer, train_loader, criterion,
args.epochs, False)
else:
print("Invalid dataset.")
quit()
if args.save_path:
model.to(torch.device('cpu'))
state = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict()
}
pth = args.save_path + '/' + args.model_name + '.pt'
print(pth)
torch.save(state, pth)
torch.cuda.empty_cache()
if args.save_path:
save_wav = args.save_path + '/' + args.model_name
D.generation(model,
device,
filename=save_wav,
seconds=args.gen_len,
dataset=args.dataset)
mel = mel.to(device)
wav = wav.to(device)
optimizer.zero_grad()
output = model(mel, wav[:, :-1])
output = output.transpose(-1, -2)
loss = criterion(output.contiguous().view(-1, 256),
wav[:, 1:].contiguous().view(-1).type(torch.LongTensor).to(device))
running_loss += loss.item()
loss.backward()
optimizer.step()
ctr += 1
pbar.update(1)
torch.save(model.state_dict(), 'weights'+str(epoch) + '.pt')
for data in val_loader:
mel, wav = data
mel = mel.to(device)
wav = wav.to(device)
model.eval()
with torch.no_grad():
output = model(mel, wav[:, :-1])
output = output.transpose(-1, -2)
loss = criterion(output.contiguous().view(-1, 256),
wav[:, 1:].contiguous().view(-1).type(torch.LongTensor).to(device))
val_loss += loss.item()
val_ctr += 1