Fix failing tutorials, change MNLE default for log_transform to False (…

…#1367)

* fix fmpe typo, add nle seed

* fix: change default log transfrom to False.

This was set to True for positive reaction times but does not hold in general.

* small nb fixes

sbi/neural_nets/net_builders/mnle.py

@@ -64,7 +64,7 @@ def build_mnle(
     hidden_features: int = 50,
     hidden_layers: int = 2,
     tail_bound: float = 10.0,
-    log_transform_x: bool = True,
+    log_transform_x: bool = False,
     **kwargs,
 ):
     """Returns a density estimator for mixed data types.

tests/mnle_test.py

@@ -162,7 +162,9 @@ def test_mnle_accuracy_with_different_samplers_and_trials(
     x = mixed_simulator(theta, stimulus_condition=1.0)
 
     # MNLE
-    density_estimator = likelihood_nn(model="mnle", flow_model=flow_model)
+    density_estimator = likelihood_nn(
+        model="mnle", flow_model=flow_model, log_transform_x=True
+    )
     trainer = MNLE(prior, density_estimator=density_estimator)
     trainer.append_simulations(theta, x).train(training_batch_size=200)
     posterior = trainer.build_posterior()
@@ -294,7 +296,7 @@ def test_mnle_with_experimental_conditions(mcmc_params_accurate: dict):
     )
 
     # MNLE
-    estimator_fun = likelihood_nn(model="mnle", z_score_x=None)
+    estimator_fun = likelihood_nn(model="mnle", log_transform_x=True)
     trainer = MNLE(proposal, estimator_fun)
     estimator = trainer.append_simulations(theta, x).train()
 
@@ -362,9 +364,7 @@ def test_log_likelihood_over_local_iid_theta(
     """
 
     # train mnle on mixed data
-    trainer = MNLE(
-        density_estimator=likelihood_nn(model="mnle", z_score_x=None),
-    )
+    trainer = MNLE()
     proposal = MultipleIndependent(
         [
             Gamma(torch.tensor([1.0]), torch.tensor([0.5])),

tests/tutorials_test.py

@@ -22,7 +22,7 @@ def test_tutorials(notebook_path):
     """Test that all notebooks in the tutorials directory can be executed."""
     with open(notebook_path) as f:
         nb = nbformat.read(f, as_version=4)
-        ep = ExecutePreprocessor(timeout=1200, kernel_name='python3')
+        ep = ExecutePreprocessor(timeout=600, kernel_name='python3')
         print(f"Executing notebook {notebook_path}")
         try:
             ep.preprocess(nb, {'metadata': {'path': os.path.dirname(notebook_path)}})

tutorials/16_implemented_methods.ipynb

@@ -187,7 +187,7 @@
     "from sbi.inference import FMPE\n",
     "\n",
     "inference = FMPE(prior)\n",
-    "# FMPE does support multiple rounds of inference\n",
+    "# FMPE does not support multiple rounds of inference\n",
     "theta = prior.sample((num_sims,))\n",
     "x = simulator(theta)\n",
     "inference.append_simulations(theta, x).train()\n",
@@ -310,7 +310,8 @@
     "\n",
     "inference = MNLE(prior)\n",
     "theta = prior.sample((num_sims,))\n",
-    "x = simulator(theta)\n",
+    "# add a column of discrete data to x.\n",
+    "x = torch.cat((simulator(theta), torch.bernoulli(theta[:, :1])), dim=1)\n",
     "_ = inference.append_simulations(theta, x).train()\n",
     "posterior = inference.build_posterior().set_default_x(x_o)"
    ]

tutorials/Example_01_DecisionMakingModel.ipynb

@@ -129,7 +129,8 @@
     "        Beta(torch.tensor([2.0]), torch.tensor([2.0])),\n",
     "    ],\n",
     "    validate_args=False,\n",
-    ")"
+    ")\n",
+    "prior_transform = mcmc_transform(prior)"
    ]
   },
   {
@@ -184,7 +185,7 @@
     "true_posterior = MCMCPosterior(\n",
     "    potential_fn=BinomialGammaPotential(prior, x_o),\n",
     "    proposal=prior,\n",
-    "    theta_transform=mcmc_transform(prior, enable_transform=True),\n",
+    "    theta_transform=prior_transform,\n",
     "    **mcmc_kwargs,\n",
     ")\n",
     "true_samples = true_posterior.sample((num_samples,))"
@@ -228,7 +229,8 @@
     "x = mixed_simulator(theta)\n",
     "\n",
     "# Train MNLE and obtain MCMC-based posterior.\n",
-    "trainer = MNLE()\n",
+    "estimator_builder = likelihood_nn(model=\"mnle\", log_transform_x=True)\n",
+    "trainer = MNLE(proposal, estimator_builder)\n",
     "estimator = trainer.append_simulations(theta, x).train()"
    ]
   },
@@ -610,7 +612,7 @@
     }
    ],
    "source": [
-    "estimator_builder = likelihood_nn(model=\"mnle\", z_score_x=None)  # we don't want to z-score the binary data.\n",
+    "estimator_builder = likelihood_nn(model=\"mnle\", log_transform_x=True)\n",
     "trainer = MNLE(proposal, estimator_builder)\n",
     "estimator = trainer.append_simulations(theta, x).train()"
    ]
@@ -847,7 +849,6 @@
     "\n",
     "fig, ax = pairplot(\n",
     "    [prior.sample((1000,))] + posterior_samples,\n",
-    "    # points=theta_o,\n",
     "    diag=\"kde\",\n",
     "    upper=\"contour\",\n",
     "    diag_kwargs=dict(bins=100),\n",