def train(Xtrain, Xvalid, centers=2, regression=False):
X = np.concatenate([x for (x, *_) in Xtrain])
gm = []
for X, *_ in Xtrain:
gm.append(X)
gm = np.concatenate(gm)
model = sklearn.mixture.GaussianMixture(centers, "diag")
gm = model.fit(gm)
component = [
Gaussian(torch.Tensor(gm.means_[i, :]),
torch.Tensor(1. / gm.covariances_[i, :]))
for i in range(centers)
]
mixture = Mixture(component, gm.weights_)
classifier = DensityClassifier(mixture, centers, 2)
X = torch.cat([
mixture(torch.Tensor(X)).unsqueeze_(0).detach()
for (X, y, *_) in Xtrain
])
if regression:
y = torch.FloatTensor([y for (X, y, *_) in Xtrain])
else:
y = torch.LongTensor([y for (X, y, *_) in Xtrain])
Xv = torch.cat([
mixture(torch.Tensor(X)).unsqueeze_(0).detach()
for (X, y, *_) in Xvalid
])
if regression:
yv = torch.FloatTensor([y for (X, y, *_) in Xvalid])
else:
yv = torch.LongTensor([y for (X, y, *_) in Xvalid])
logger = logging.getLogger(__name__)
# Set weights of classifier
for lr in [1e2, 1e1, 1e0, 1e-1, 1e-2, 1e-3]:
optimizer = torch.optim.SGD(classifier.pl.parameters(),
lr=lr,
momentum=0.9)
if regression:
criterion = torch.nn.modules.MSELoss()
else:
criterion = torch.nn.modules.CrossEntropyLoss()
best_loss = float("inf")
best_model = copy.deepcopy(classifier.pl.state_dict())
logger.debug("Learning rate: " + str(lr))
for i in range(1000):
z = classifier.pl(X)
if regression:
z = z[:, 1]
loss = criterion(z, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
zv = classifier.pl(Xv)
if regression:
zv = zv[:, 1]
loss = criterion(zv, yv)
if i % 100 == 0:
logger.debug(str(loss.detach().numpy()))
if loss < best_loss:
best_loss = loss
best_model = copy.deepcopy(classifier.pl.state_dict())
classifier.pl.load_state_dict(best_model)
reg = None
return cloudpred.utils.train_classifier(Xtrain,
Xvalid, [],
classifier,
regularize=reg,
iterations=1000,
eta=1e-4,
stochastic=True,
regression=regression)
def get_ll(X, mu):
sm = 0
for x in X:
sm += log(mixture(x, [0.5, 0.5], mu))
return sm
def log_lh(X, mu):
s = 0
for i in X:
s += np.log(mixture(i, [0.5, 0.5], mu))
return s
def get_ll(X, mu):
sm = 0
for x in X:
sm += log(mixture(x, [0.5, 0.5], mu))
return sm