def make_loss(self, pi, sigma, mu, labels=None, warmup=None, warmup_threshold=10):
"""
The special loss for mdn
:param logit: The output of the network
:param labels: The ground truth labels
:param warmup: The warmup process for the mean to get the range faster
:return: the total loss
"""
#return mdn.new_mdn_loss(pi, sigma, mu, labels)
if warmup is None or warmup> warmup_threshold: # If no warmup is being done
return mdn.mdn_loss(pi, sigma, mu, labels)
else:
print('warmup mode on')
# Make sigma to 1 (wide)
loss = 10*torch.mean(torch.pow(sigma-1, 2))
# Make the mu converge to labels
loss += 10*torch.mean(torch.pow(torch.mean(mu, 1) - labels, 2))
return loss
def train(model, train_loader, num_ep, lear_rate):
"""
Train MDN model using train dataset
Args:
model: pytorch MDN model to be trained
train_loader: loader for the train dataset
num_ep: number for epochs for the training
Returns:
model: pytorch MDN trained model
"""
optimizer = optim.Adam(model.parameters(),lr=lear_rate)
# train the model
for epoch in range(num_ep):
for batch_idx, (minibatch, labels) in enumerate(train_loader):
# rearrange data shape
minibatch = minibatch.reshape(batch_size, 1, 784)
labels = labels.reshape(batch_size, 1)
model.zero_grad()
pi, sigma, mu = model(minibatch)
loss = mdn.mdn_loss(pi, sigma, mu, labels)
loss.backward()
optimizer.step()
print('Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(minibatch), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
return model
# train the model
for epoch in range(num_epochs):
for batch_idx, (labels, minibatch) in enumerate(train_loader):
gt = labels[0:3]
minibatch = [[numb == minibatch[i] for numb in range(10)]
for i in range(batch_size)]
minibatch = torch.FloatTensor(minibatch).unsqueeze(1)
labels = labels.reshape(batch_size, 784)
model.zero_grad()
pi, sigma, mu = model(minibatch)
loss = mdn.mdn_loss(pi, sigma, mu, labels)
loss.backward()
optimizer.step()
print('Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(minibatch), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
# Visualize
if batch_idx == 0:
samples = mdn.sample(pi, sigma, mu).int()
images = samples[0:3]
plot(gt, images, epoch)
print(
"The training is compelete! \nVisualization results have been saved in the folder 'figures'"
training_set = torch.cat([cluster1, cluster2, cluster3])
print('Done')
print("Initializing model... ", end='')
model = nn.Sequential(nn.Linear(input_dims, 5), nn.Tanh(),
mdn.MDN(5, output_dims, num_gaussians))
optimizer = optim.Adam(model.parameters())
print('Done')
print('Training model... ', end='')
sys.stdout.flush()
for epoch in range(1000):
model.zero_grad()
pi, sigma, mu = model(training_set[:, 0:input_dims])
loss = mdn.mdn_loss(pi, sigma, mu, training_set[:, input_dims:])
loss.backward()
optimizer.step()
if epoch % 100 == 99:
print(f' {round(epoch/10)}%', end='')
sys.stdout.flush()
print(' Done')
print('Generating samples... ', end='')
pi, sigma, mu = model(training_set[:, 0:input_dims])
samples = mdn.sample(pi, sigma, mu)
print('Done')
print('Saving samples.png... ', end='')
fig = plt.figure()
ax = fig.add_subplot(projection='3d')