lifejacket 1.0.0__tar.gz → 1.0.2__tar.gz

{lifejacket-1.0.0 → lifejacket-1.0.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lifejacket
-Version: 1.0.0
+Version: 1.0.2
 Summary: Consistent standard errors for longitudinal data collected under pooling online decision policies.
 Author-email: Nowell Closser <nowellclosser@gmail.com>
 Requires-Python: >=3.10

{lifejacket-1.0.0 → lifejacket-1.0.2}/lifejacket/calculate_derivatives.py

@@ -18,8 +18,6 @@ logging.basicConfig(
     level=logging.INFO,
 )
 
-# TODO: Consolidate function loading logic
-
 
 def get_batched_arg_lists_and_involved_user_ids(func, sorted_user_ids, args_by_user_id):
     """

lifejacket-1.0.2/lifejacket/constants.py

@@ -0,0 +1,16 @@
+class SmallSampleCorrections:
+    NONE = "none"
+    Z1theta = "Z1theta"
+    Z2theta = "Z2theta"
+    Z3theta = "Z3theta"
+
+
+class FunctionTypes:
+    LOSS = "loss"
+    ESTIMATING = "estimating"
+
+
+class SandwichFormationMethods:
+    BREAD_T_QR = "bread_T_qr"
+    MEAT_SVD_SOLVE = "meat_svd_solve"
+    NAIVE = "naive"

{lifejacket-1.0.0 → lifejacket-1.0.2}/lifejacket/deployment_conditioning_monitor.py

@@ -35,6 +35,12 @@ logging.basicConfig(
 
 
 class DeploymentConditioningMonitor:
+    """
+    Experimental feature.  Monitors the conditioning of the RL portion of the bread matrix.
+    Repeats more logic from post_deployment_analysis.py than is ideal, but this is for experimental use only.
+    Unit tests should be unskipped and expanded if this is to be used more broadly.
+    """
+
     whole_RL_block_conditioning_threshold = None
     diagonal_RL_block_conditioning_threshold = None
 
@@ -76,7 +82,7 @@ class DeploymentConditioningMonitor:
         incremental: bool = True,
     ) -> None:
         """
-        Analyzes a dataset to estimate parameters and variance using adaptive and classical sandwich estimators.
+        Analyzes a dataset to estimate parameters and variance using adjusted and classical sandwich estimators.
 
         Parameters:
         proposed_policy_num (int | float):
@@ -124,13 +130,13 @@ class DeploymentConditioningMonitor:
         small_sample_correction (str):
             Type of small sample correction to apply.
         collect_data_for_blowup_supervised_learning (bool):
-            Whether to collect data for doing supervised learning about adaptive sandwich blowup.
+            Whether to collect data for doing supervised learning about adjusted sandwich blowup.
         form_adjusted_meat_adjustments_explicitly (bool):
-            If True, explicitly forms the per-subject meat adjustments that differentiate the adaptive
+            If True, explicitly forms the per-subject meat adjustments that differentiate the adjusted
             sandwich from the classical sandwich. This is for diagnostic purposes, as the
-            adaptive sandwich is formed without doing this.
-        stabilize_joint_adjusted_bread_inverse (bool):
-            If True, stabilizes the joint adaptive bread inverse matrix if it does not meet conditioning
+            adjusted sandwich is formed without doing this.
+        stabilize_joint_bread (bool):
+            If True, stabilizes the joint bread matrix if it does not meet conditioning
             thresholds.
 
         Returns:
@@ -157,6 +163,7 @@ class DeploymentConditioningMonitor:
                 alg_update_func_args_beta_index,
                 alg_update_func_args_action_prob_index,
                 alg_update_func_args_action_prob_times_index,
+                alg_update_func_args_previous_betas_index,
                 suppress_interactive_data_checks,
             )
 
@@ -230,7 +237,7 @@ class DeploymentConditioningMonitor:
 
         if whole_RL_block_condition_number > self.whole_RL_block_conditioning_threshold:
             logger.warning(
-                "The RL portion of the bread inverse up to this point exceeds the threshold set (condition number: %s, threshold: %s). Consider an alternative update strategy which produces less dependence on previous RL parameters (via the data they produced) and/or improves the conditioning of each update itself.  Regularization may help with both of these.",
+                "The RL portion of the bread up to this point exceeds the threshold set (condition number: %s, threshold: %s). Consider an alternative update strategy which produces less dependence on previous RL parameters (via the data they produced) and/or improves the conditioning of each update itself.  Regularization may help with both of these.",
                 whole_RL_block_condition_number,
                 self.whole_RL_block_conditioning_threshold,
             )
@@ -241,7 +248,7 @@ class DeploymentConditioningMonitor:
             > self.diagonal_RL_block_conditioning_threshold
         ):
             logger.warning(
-                "The diagonal RL block of the bread inverse up to this point exceeds the threshold set (condition number: %s, threshold: %s). This may illustrate a fundamental problem with the conditioning of the RL update procedure.",
+                "The diagonal RL block of the bread up to this point exceeds the threshold set (condition number: %s, threshold: %s). This may illustrate a fundamental problem with the conditioning of the RL update procedure.",
                 new_diagonal_RL_block_condition_number,
                 self.diagonal_RL_block_conditioning_threshold,
             )
@@ -295,11 +302,11 @@ class DeploymentConditioningMonitor:
         jnp.ndarray[jnp.float32],
     ]:
         """
-        Constructs the classical and adaptive inverse bread and meat matrices, as well as the average
+        Constructs the classical and bread and meat matrices, as well as the average
         estimating function stack and some other intermediate pieces.
 
         This is done by computing and differentiating the average weighted estimating function stack
-        with respect to the betas and theta, using the resulting Jacobian to compute the inverse bread
+        with respect to the betas and theta, using the resulting Jacobian to compute the bread
         and meat matrices, and then stably computing sandwiches.
 
         Args:
@@ -471,7 +478,7 @@ class DeploymentConditioningMonitor:
     ]:
         """
         Computes the average weighted estimating function stack across all subjects, along with
-        auxiliary values used to construct the adaptive and classical sandwich variances.
+        auxiliary values used to construct the adjusted and classical sandwich variances.
 
         If only_latest_block is True, only uses data from the most recent update.
 
@@ -614,7 +621,7 @@ class DeploymentConditioningMonitor:
             )
 
         # 5. Now we can compute the weighted estimating function stacks for all subjects
-        # as well as collect related values used to construct the adaptive and classical
+        # as well as collect related values used to construct the adjusted and classical
         # sandwich variances.
         RL_stacks = jnp.array(
             [

{lifejacket-1.0.0 → lifejacket-1.0.2}/lifejacket/form_adjusted_meat_adjustments_directly.py

@@ -21,7 +21,7 @@ logging.basicConfig(
 def form_adjusted_meat_adjustments_directly(
     theta_dim: int,
     beta_dim: int,
-    joint_adaptive_bread_inverse_matrix: jnp.ndarray,
+    joint_bread_matrix: jnp.ndarray,
     per_user_estimating_function_stacks: jnp.ndarray,
     study_df: pd.DataFrame,
     active_col_name: str,
@@ -38,18 +38,16 @@ def form_adjusted_meat_adjustments_directly(
     action_prob_col_name: str,
 ) -> jnp.ndarray:
     logger.info(
-        "Explicitly forming the per-user meat adjustments that differentiate the adaptive sandwich from the classical sandwich."
+        "Explicitly forming the per-user meat adjustments that differentiate the adjusted sandwich from the classical sandwich."
     )
 
     # 1. Form the M-matrices, which are shared across users.
     # This is not quite the paper definition of the M-matrices, which
     # includes multiplication by the classical bread.  We don't care about
     # that here, since in forming the adjustments there is a multiplication
-    # by the classical bread inverse that cancels it out.
-    V_blocks_together = joint_adaptive_bread_inverse_matrix[-theta_dim:, :-theta_dim]
-    RL_stack_beta_derivatives_block = joint_adaptive_bread_inverse_matrix[
-        :-theta_dim, :-theta_dim
-    ]
+    # by the classical bread that cancels it out.
+    V_blocks_together = joint_bread_matrix[-theta_dim:, :-theta_dim]
+    RL_stack_beta_derivatives_block = joint_bread_matrix[:-theta_dim, :-theta_dim]
     effective_M_blocks_together = np.linalg.solve(
         RL_stack_beta_derivatives_block.T, V_blocks_together.T
     ).T
@@ -89,17 +87,17 @@ def form_adjusted_meat_adjustments_directly(
     # be added to each users inference estimating function before an outer product is taken with
     # itself to get each users's contributioan theta-only meat matrix).
     # Result is shape (num_users, theta_dim).
-    # Form the per-user adaptive meat adjustments explicitly for diagnostic purposes.
+    # Form the per-user adjusted meat adjustments explicitly for diagnostic purposes.
     per_user_meat_adjustments_stacked = np.einsum(
         "utb,nub->nt", M_blocks_stacked, per_user_RL_only_est_fns_stacked
     )
 
-    # Log some diagnostics about the pieces going into the adaptive meat adjustments
+    # Log some diagnostics about the pieces going into the adjusted meat adjustments
     # and the adjustments themselves.
     V_blocks = np.split(
         V_blocks_together, V_blocks_together.shape[1] // beta_dim, axis=1
     )
-    logger.info("Examining adaptive meat adjustments.")
+    logger.info("Examining adjusted meat adjustments.")
     # No scientific notation
     np.set_printoptions(suppress=True)
 
@@ -118,11 +116,11 @@ def form_adjusted_meat_adjustments_directly(
     )
 
     logger.debug(
-        "Per-user adaptive meat adjustments, to be added to inference estimating functions before forming the meat. Formed from the sum of the products of the M-blocks and the corresponding RL update estimating functions for each user: %s",
+        "Per-user adjusted meat adjustments, to be added to inference estimating functions before forming the meat. Formed from the sum of the products of the M-blocks and the corresponding RL update estimating functions for each user: %s",
         per_user_meat_adjustments_stacked,
     )
     logger.debug(
-        "Norms of per-user adaptive meat adjustments: %s",
+        "Norms of per-user adjusted meat adjustments: %s",
         np.linalg.norm(per_user_meat_adjustments_stacked, axis=1),
     )
 
@@ -148,10 +146,10 @@ def form_adjusted_meat_adjustments_directly(
 
     logger.debug(
         "M_blocks, one per update, each shape theta_dim x beta_dim. The sum of the products "
-        "of each of these times a user's corresponding RL estimating function forms their adaptive "
+        "of each of these times a user's corresponding RL estimating function forms their adjusted "
         "adjustment. The M's are the blocks of the the product of the V's concatened and the inverse of "
-        "the RL-only upper-left block of the joint adaptive bread inverse. In other words, the lower "
-        "left block of the joint adaptive bread. Also note that the inference estimating function "
+        "the RL-only upper-left block of the joint bread. In other words, the lower "
+        "left block of the joint bread. Also note that the inference estimating function "
         "derivative inverse is omitted here despite the definition of the M's in the paper, because "
         "that factor simply cancels later: %s",
         M_blocks_stacked,
@@ -159,11 +157,11 @@ def form_adjusted_meat_adjustments_directly(
     logger.debug("Norms of M-blocks: %s", np.linalg.norm(M_blocks_stacked, axis=(1, 2)))
 
     logger.debug(
-        "RL block of joint adaptive bread inverse. The *inverse* of this goes into the M's: %s",
+        "RL block of joint bread. The *inverse* of this goes into the M's: %s",
         RL_stack_beta_derivatives_block,
     )
     logger.debug(
-        "Norm of RL block of joint adaptive bread inverse: %s",
+        "Norm of RL block of joint bread: %s",
         np.linalg.norm(RL_stack_beta_derivatives_block),
     )
 
@@ -171,11 +169,11 @@ def form_adjusted_meat_adjustments_directly(
         RL_stack_beta_derivatives_block
     )
     logger.debug(
-        "Inverse of RL block of joint adaptive bread inverse. This goes into the M's: %s",
+        "Inverse of RL block of joint bread. This goes into the M's: %s",
         inverse_RL_stack_beta_derivatives_block,
     )
     logger.debug(
-        "Norm of Inverse of RL block of joint adaptive bread inverse: %s",
+        "Norm of Inverse of RL block of joint bread: %s",
         np.linalg.norm(inverse_RL_stack_beta_derivatives_block),
     )
 
@@ -230,17 +228,17 @@ def form_adjusted_meat_adjustments_directly(
         per_user_meat_adjustments_stacked
         + per_user_inference_estimating_functions_stacked
     )
-    per_user_theta_only_adaptive_meat_contributions = jnp.einsum(
+    per_user_theta_only_adjusted_meat_contributions = jnp.einsum(
         "ni,nj->nij",
         per_user_adjusted_inference_estimating_functions_stacked,
         per_user_adjusted_inference_estimating_functions_stacked,
     )
-    adaptive_theta_only_meat_matrix = jnp.mean(
-        per_user_theta_only_adaptive_meat_contributions, axis=0
+    adjusted_theta_only_meat_matrix = jnp.mean(
+        per_user_theta_only_adjusted_meat_contributions, axis=0
     )
     logger.info(
-        "Theta-only adaptive meat matrix (no small sample corrections): %s",
-        adaptive_theta_only_meat_matrix,
+        "Theta-only adjusted meat matrix (no small sample corrections): %s",
+        adjusted_theta_only_meat_matrix,
     )
     classical_theta_only_meat_matrix = jnp.mean(
         jnp.einsum(
@@ -311,7 +309,7 @@ def form_adjusted_meat_adjustments_directly(
     # much or being too opinionated about what to log.
     breakpoint()
 
-    # # Visualize the inverse RL block of joint adaptive bread inverse using seaborn heatmap
+    # # Visualize the inverse RL block of joint bread using seaborn heatmap
     # pyplt.figure(figsize=(8, 6))
     # sns.heatmap(inverse_RL_stack_beta_derivatives_block, annot=False, cmap="viridis")
     # pyplt.title("Inverse RL Block of Joint Adaptive Bread Inverse")
@@ -320,7 +318,7 @@ def form_adjusted_meat_adjustments_directly(
     # pyplt.tight_layout()
     # pyplt.show()
 
-    # # # Visualize the RL block of joint adaptive bread inverse using seaborn heatmap
+    # # # Visualize the RL block of joint bread using seaborn heatmap
 
     # pyplt.figure(figsize=(8, 6))
     # sns.heatmap(RL_stack_beta_derivatives_block, annot=False, cmap="viridis")
@@ -330,4 +328,4 @@ def form_adjusted_meat_adjustments_directly(
     # pyplt.tight_layout()
     # pyplt.show()
 
-    return per_user_theta_only_adaptive_meat_contributions
+    return per_user_theta_only_adjusted_meat_contributions