def build_model():
model = Wavenet(out_channels=2,
num_blocks=args.num_blocks,
num_layers=args.num_layers,
residual_channels=args.residual_channels,
gate_channels=args.gate_channels,
skip_channels=args.skip_channels,
kernel_size=args.kernel_size,
cin_channels=args.cin_channels,
upsample_scales=[16, 16])
print(model)
n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('number of teacher parameters:', n_params)
return model
device=device)
loadtr = data.DataLoader(training_set,
batch_size=1,
shuffle=True,
num_workers=0,
worker_init_fn=np.random.seed)
loadval = data.DataLoader(validation_set, batch_size=1, num_workers=0)
# In[6]:
#model = Unet(skipDim, quantization_channels, residualDim,device)
model = Wavenet(field, skipDim, residualDim, dilations0, dilations1)
#model = nn.DataParallel(model)
model = model.cuda()
criterion = nn.CrossEntropyLoss()
# in wavenet paper, they said crossentropyloss is far better than MSELoss
optimizer = optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-5)
# use adam to train
maxloss = np.zeros(50) + 100
# In[7]:
start_epoch = 0
if continueTrain: # if continueTrain, the program will find the checkpoints
if os.path.isfile(resumefile):
print("=> loading checkpoint '{}'".format(resumefile))
checkpoint = torch.load(resumefile)
start_epoch = checkpoint['epoch']
# best_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
