Minor updates: improve docstring to correlated MCMC and remove some deprecated parameters

abcpy/NN_utilities/utilities.py

@@ -79,9 +79,9 @@ def jacobian(input, output, diffable=True):
         g = torch.zeros(output.shape).to(input)
         g[:, i] = 1
         if diffable:
-            J[:, i] = torch.autograd.grad(output, input, g, only_inputs=True, retain_graph=True, create_graph=True)[0]
+            J[:, i] = torch.autograd.grad(output, input, g, retain_graph=True, create_graph=True)[0]
         else:
-            J[:, i] = torch.autograd.grad(output, input, g, only_inputs=True, retain_graph=True)[0]
+            J[:, i] = torch.autograd.grad(output, input, g, retain_graph=True)[0]
     J = J.reshape(output.shape[0], output.shape[1], in_size)
     return J.transpose(2, 1)
 
@@ -107,18 +107,18 @@ def jacobian_second_order(input, output, diffable=True):
     for i in range(output.shape[1]):
         g = torch.zeros(output.shape).to(input)
         g[:, i] = 1
-        J[:, i] = torch.autograd.grad(output, input, g, only_inputs=True, retain_graph=True, create_graph=True)[0]
+        J[:, i] = torch.autograd.grad(output, input, g, retain_graph=True, create_graph=True)[0]
     J = J.reshape(output.shape[0], output.shape[1], in_size)
 
     for i in range(output.shape[1]):
         for j in range(input.shape[1]):
             g = torch.zeros(J.shape).to(input)
             g[:, i, j] = 1
             if diffable:
-                J2[:, i, j] = torch.autograd.grad(J, input, g, only_inputs=True, retain_graph=True, create_graph=True)[
+                J2[:, i, j] = torch.autograd.grad(J, input, g, retain_graph=True, create_graph=True)[
                                   0][:, j]
             else:
-                J2[:, i, j] = torch.autograd.grad(J, input, g, only_inputs=True, retain_graph=True)[0][:, j]
+                J2[:, i, j] = torch.autograd.grad(J, input, g, retain_graph=True)[0][:, j]
 
     J2 = J2.reshape(output.shape[0], output.shape[1], in_size)
 

abcpy/inferences.py

@@ -4072,9 +4072,11 @@ def _compute_accepted_cov_mats(self, covFactor, new_cov_mats):
 
 class MCMCMetropoliHastings(BaseLikelihood, InferenceMethod):
     """
-    Simple Metropolis-Hastings MCMC working with the approximate likelihood functions Approx_likelihood, with
+    Metropolis-Hastings MCMC working with the approximate likelihood functions Approx_likelihood, with
     multivariate normal proposals.
 
+    As the Approximate Likelihood is estimated from simulations from the data, this is a pseudo-marginal MCMC approach.
+
     Parameters
     ----------
     root_models : list
@@ -4150,6 +4152,12 @@ def sample(self, observations, n_samples, n_samples_per_param=100, burnin=1000,
         use MCMC with transformed space, you need to specify lower and upper bounds in the corresponding parameters (see
         details in the description of `bounds`).
 
+        As the Approximate Likelihood is estimated from simulations from the data, this is a pseudo-marginal MCMC
+        approach. As pseudo-marginal can lead to sticky ``chains'', this method also implements correlated
+        pseudo-marginal MCMC [2], where the noise used to estimate the target at subsequent steps is correlated.
+        See the argument `n_groups_correlated_randomness` to learn more.
+
+
         The returned journal file contains also information on acceptance rates (in the configuration dictionary).
 
         [1] Haario, H., Saksman, E., & Tamminen, J. (2001). An adaptive Metropolis algorithm. Bernoulli, 7(2), 223-242.
@@ -4223,7 +4231,8 @@ def sample(self, observations, n_samples, n_samples_per_param=100, burnin=1000,
             and was discussed in [2] for the Bayesian Synthetic Likelihood framework. Notice that, when
             n_groups_correlated_randomness > 0 and speedup_dummy is True, you obtain different results for different
             values of n_groups_correlated_randomness due to different ways of handling random seeds.
-            When None, we do not keep track of the random seeds. Default value is None.
+            When None, we do not keep track of the random seeds (i.e. it is a vanilla pseudo-marginal); the same is true
+            when n_groups_correlated_randomness=1. Default value is None.
         use_tqdm : boolean, optional
             Whether using tqdm or not to display progress. Defaults to True.
         journal_file: str, optional