def plot_clustering(self, X, params, labels):
B = X.shape[0]
K = len(params)
nx, ny = 50, 50
if B > 1:
fig, axes = plt.subplots(2, B // 2, figsize=(2.5 * B, 10))
for b, ax in enumerate(axes.flatten()):
ulabels, colors = scatter(X[b], labels=labels[b], ax=ax)
for l, c in zip(ulabels, colors):
Xbl = X[b][labels[b] == l]
Z, x, y = meshgrid_around(Xbl, nx, ny, margin=0.1)
ll = self.net.flow.log_prob(Z,
context=params[l][b]).reshape(
nx, ny)
ax.contour(to_numpy(x),
to_numpy(y),
to_numpy(ll.exp()),
zorder=10,
alpha=0.3)
else:
ulabels, colors = scatter(X[0], labels=labels[0])
for l, c in zip(ulabels, colors):
Xbl = X[0][labels[0] == l]
Z, x, y = meshgrid_around(Xbl, nx, ny, margin=0.1)
ll = self.net.flow.log_prob(Z, context=params[l][0]).reshape(
nx, ny)
plt.contour(to_numpy(x),
to_numpy(y),
to_numpy(ll.exp()),
zorder=10,
alpha=0.3)
def plot_filtering(self, batch):
X = batch['X']
B = X.shape[0]
with torch.no_grad():
outputs = self.compute_loss(batch, train=False)
mu, cov = self.net.mvn.stats(outputs['theta'])
labels = (outputs['logits'] > 0.0).long()
fig, axes = plt.subplots(min(B, 2),
max(B // 2, 1),
figsize=(2.5 * B if B > 1 else 10, 10))
axes = [axes] if B == 1 else axes.flatten()
for b, ax in enumerate(axes):
scatter(X[b], labels=labels[b], ax=ax)
draw_ellipse(mu[b], cov[b], ax=ax)
def plot_step(self, X):
B = X.shape[0]
self.net.eval()
params, _, logits = self.net(X)
mu, cov = self.net.mvn.stats(params)
labels = (logits > 0.0).int().squeeze(-1)
if B == 1:
scatter(X[0], labels=labels[0])
draw_ellipse(mu[0][0], cov[0][0])
else:
fig, axes = plt.subplots(2, B // 2, figsize=(2.5 * B, 10))
for b, ax in enumerate(axes.flatten()):
scatter(X[b], labels=labels[b], ax=ax)
draw_ellipse(mu[b][0], cov[b][0], ax=ax)
def plot_step(self, X):
B = X.shape[0]
self.net.eval()
params, logits = self.net(X)
labels = (logits > 0.0).int().squeeze(-1)
nx, ny = 50, 50
fig, axes = plt.subplots(2, B // 2, figsize=(7 * B / 5, 5))
for b, ax in enumerate(axes.flatten()):
scatter(X[b], labels=labels[b], ax=ax)
Z, x, y = meshgrid_around(X[b], nx, ny, margin=0.1)
ll = self.net.flow.log_prob(Z, context=params[b]).reshape(nx, ny)
ax.contour(to_numpy(x),
to_numpy(y),
to_numpy(ll.exp()),
zorder=10,
alpha=0.3)
def plot_clustering(self, X, results):
labels = results.get('labels')
theta = results.get('theta')
B = X.shape[0]
K = theta.shape[1]
nx, ny = 50, 50
fig, axes = plt.subplots(min(B, 2),
max(B // 2, 1),
figsize=(2.5 * B if B > 1 else 10, 10))
axes = [axes] if B == 1 else axes.flatten()
for b, ax in enumerate(axes):
ulabels, colors = scatter(X[b], labels=labels[b], ax=ax)
for l, c in zip(ulabels, colors):
Xbl = X[b][labels[b] == l]
Z, x, y = meshgrid_around(Xbl, nx, ny, margin=0.1)
ll = self.net.flow.log_prob(Z,
context=theta[b,
l]).reshape(nx, ny)
ax.contour(to_numpy(x),
to_numpy(y),
to_numpy(ll.exp()),
zorder=10,
alpha=0.5)
ax.set_xticks([])
ax.set_yticks([])
def plot_filtering(self, batch):
X = batch['X']
B, N = X.shape[0], X.shape[1]
with torch.no_grad():
outputs = self.compute_loss(batch, train=False)
labels = (outputs['logits'] > 0.0).long()
theta = outputs['theta']
nx, ny = 50, 50
fig, axes = plt.subplots(min(B, 2),
max(B // 2, 1),
figsize=(2.5 * B if B > 1 else 10, 10))
axes = [axes] if B == 1 else axes.flatten()
for b, ax in enumerate(axes):
scatter(X[b], labels=labels[b], ax=ax)
Z, x, y = meshgrid_around(X[b], nx, ny, margin=0.1)
ll = self.net.flow.log_prob(Z, context=theta[b]).reshape(nx, ny)
ax.contour(to_numpy(x),
to_numpy(y),
to_numpy(ll.exp()),
zorder=10,
alpha=0.3)