def _post_backward_callback(mixture, n_iter, log_probs, prior_crossent, q_entropy, temperature, writer):
if n_iter % 100 == 0:
densities = mixture.forward(torch.Tensor(z)).numpy()
f = plt.figure(figsize=(10, 10))
zz = np.argmax(densities, axis=1).reshape([1000, 1000])
plt.contourf(xx, yy, zz, 50, cmap="rainbow")
with torch.no_grad():
for i, component in enumerate(mixture.components):
X_k = component.sample(512)
plt.scatter(X_k[:, 0].numpy(), X_k[:, 1].numpy(),
c=[colors[i]],
s=5)
plt.xlim(-1.1, 1.1)
plt.ylim(-1.1, 1.1)
writer.add_image("distributions", figure2tensor(f), n_iter)
plt.close(f)
opt.step()
if it % 2 == 0:
writer.add_scalar("debug/loss", loss, it)
#writer.add_scalar("debug/log_probs", base_log_probs_, it)
#writer.add_scalar("debug/log_abs_det_jacobian", log_abs_det_jacob_, it)
if it % 5 == 0:
with torch.no_grad():
Xhat = flow.sample(1000)
f = plt.figure(figsize=(10, 10))
#plt.xlim(-1.5, 1.5)
#plt.ylim(-1.5, 1.5)
plt.title(f"{it} iterations")
plt.scatter(X[:, 0], X[:, 1], s=5, c="blue", alpha=0.5)
plt.scatter(Xhat[:, 0], Xhat[:, 1], s=5, c="red", alpha=0.5)
writer.add_image("debug/samples", figure2tensor(f), it)
plt.close(f)
f = plt.figure(figsize=(10, 10))
plt.title(f"{it} iterations")
#plt.xlim(-1.5, 1.5)
#plt.ylim(-1.5, 1.5)
z = flow.base_dist.sample((1000, ))
zhat, _ = flow.inverse(X)
plt.scatter(z[:, 0], z[:, 1], s=5, c="blue", alpha=0.5)
plt.scatter(zhat[:, 0], zhat[:, 1], s=5, c="red", alpha=0.5)
writer.add_image("debug/base_dist", figure2tensor(f), it)
plt.close(f)
# %%
loss
best_flow = deepcopy(flow)
loss.backward()
if it % 50 == 0:
writer.add_scalar("loss", loss, it)
if it % 5000 == 0:
with torch.no_grad():
xhat_samples = flow.final_density.sample((1000, ))
f = plt.figure(figsize=(10, 10))
plt.xlim(-30, 30)
plt.ylim(-20, 20)
plt.scatter(xhat_samples[:, 0], xhat_samples[:, 1], s=5, c="red", alpha=0.5)
plt.scatter(x_samples[:, 0], x_samples[:, 1], s=5, c="blue", alpha=0.5)
writer.add_image("distributions", figure2tensor(f), it)
plt.close(f)
opt.step()
# %%
flow = best_flow
# %%
xhat_samples = flow.final_density.sample((1000, ))
plt.scatter(xhat_samples[:, 0], xhat_samples[:, 1], s=5, c="red")
plt.scatter(x_samples[:, 0], x_samples[:, 1], s=5, c="blue")
#plt.xlim(0, 60)
#plt.ylim(-15, 15)
plt.show()
def fit(self,
X,
dataloader,
n_epochs=1,
opt=None,
temperature_schedule=None,
clip_grad=None,
verbose=False,
writer=None):
best_loss = float("inf")
best_params = dict()
if temperature_schedule is None:
temperature_schedule = lambda t: 1
if verbose:
epochs = trange(n_epochs, desc="epoch")
else:
epochs = range(n_epochs)
for epoch in epochs:
for i, xb in enumerate(dataloader):
opt.zero_grad()
n_iter = epoch * (
(len(X) - 1) // dataloader.batch_size + 1) + i
log_probs, prior_crossent, q_entropy = self.elbo(
xb, temperature_schedule(n_iter))
loss = -(log_probs + prior_crossent + q_entropy)
if loss != loss:
continue
if loss <= best_loss:
best_loss = loss.item()
best_params = self.state_dict()
# if we're writing to tensorboard
if writer is not None:
if n_iter % 20 == 0:
writer.add_scalar('losses/log_probs', log_probs,
n_iter)
#writer.add_scalar('losses/prior_crossent', prior_crossent, n_iter)
writer.add_scalar('losses/q_entropy', q_entropy,
n_iter)
loss.backward()
if n_iter % 100 == 0:
with torch.no_grad():
densities = mixture.forward(torch.Tensor(z)).numpy()
f = plt.figure(figsize=(10, 10))
zz = np.argmax(densities, axis=1).reshape([1000, 1000])
plt.contourf(xx, yy, zz, 50, cmap="rainbow")
colors = ["yellow", "green", "black", "cyan"]
with torch.no_grad():
for i, component in enumerate(mixture.components):
X_k = component.sample(500)
plt.scatter(X_k[:, 0].numpy(),
X_k[:, 1].numpy(),
c=colors[i],
s=5)
plt.xlim(-1.1, 1.1)
plt.ylim(-1.1, 1.1)
writer.add_image("distributions", figure2tensor(f),
n_iter)
plt.close(f)
#if clip_grad is not None:
# torch.nn.utils.clip_grad_norm_(self.parameters(), clip_grad)
opt.step()
#if n_iter == 2500:
#print("changing learning rates")
#for param_group in opt.param_groups:
#if param_group["label"] == "remaining":
# param_group["lr"] = 6e-2
#if param_group["label"] == "encoder":
# param_group["lr"] = 1e-3
return best_loss, best_params