calculating Gauss density one px at a time

nerf/utils.py

@@ -324,17 +324,14 @@ def nll(truths, preds, vars):
  truths = truths.flatten(end_dim=-2).cpu().numpy()
  preds = preds.flatten(end_dim=-2).cpu().numpy()
  vars = vars.flatten(end_dim=-2).cpu().numpy()
- covs = []
- for px in vars:
- covs.append(np.diag(px))
- covs = np.array(covs)
- print(truths.shape)
- print(preds.shape)
- print(vars.shape)
- print(covs.shape)
- print(multivariate_normal(truths, preds, covs).shape)
- print(torch.log(multivariate_normal(truths, preds, covs)).mean().shape)
- return torch.log(multivariate_normal(truths, preds, covs)).mean()
+ log_pdf_vals = []
+ for px_ind in range(vars.shape[0]):
+ log_pdf_vals.append(
+ np.log(
+ multivariate_normal(truths[px_ind], preds[px_ind], np.diag(vars[px_ind]))
+ )
+ )
+ return np.mean(log_pdf_vals)
 
  self.fn = nll