Spline and a few other transforms don't work for density estimation...

[stefanwebb]

I've discovered there is some bug that means that .backwards() throws an error when you call it the second time during the learning loop for density estimation. I've created an extra test for this and am in the process of debugging (see #2544)

Here's an example of the bug:

import torch
import pyro
import pyro.distributions as dist
import pyro.distributions.transforms as T

import numpy as np
from sklearn import datasets
from sklearn.preprocessing import StandardScaler

n_samples = 1000
X, y = datasets.make_circles(n_samples=n_samples, factor=0.5, noise=0.05)
X = StandardScaler().fit_transform(X)

base_dist = dist.Normal(torch.zeros(2), torch.ones(2))
spline_transform = T.spline(2)
flow_dist = dist.TransformedDistribution(base_dist, [spline_transform])

torch.autograd.set_detect_anomaly(True)

steps = 1000
dataset = torch.tensor(X, dtype=torch.float, requires_grad=False)
optimizer = torch.optim.Adam(spline_transform.parameters(), lr=1e-2)
for step in range(steps):
    optimizer.zero_grad()
    loss = -flow_dist.log_prob(dataset.detach()).mean()
    loss.backward()
    optimizer.step()
    
    if step % 100 == 0:
        print(f'step: {step}, loss: {loss.item()}')

[stefanwebb]

#2544 has merged so I'm closing this