src/beanmachine/ppl/compiler/runtime.py (45 lines):
	- line 6: # TODO: Update this comment
	- line 210: # TODO: Support kwargs
	- line 366: # TODO: Add a note here describing why it is important that the function
	- line 370: # TODO: Verify that they do match.
	- line 591: # TODO: We probably need to make a distinction between torch.mm and
	- line 733: # TODO: Unfortunately, the second rule does not match the actual behavior of
	- line 770: # TODO: What if we have a non-tensor indexed with a stochastic value?
	- line 813: # TODO: Remove this. We should insert TO_REAL nodes when necessary
	- line 934: # TODO: We have a problem that we need to resolve regarding compilation of models
	- line 969: # TODO: Should we produce some sort of warning here telling the user that
	- line 983: # TODO: This scenario is another opportunity for a warning or error, since
	- line 1077: # TODO: Better error
	- line 1082: # TODO: Better error
	- line 1164: # TODO: Note that we only memoize calls to RVs when the arguments
	- line 1190: # TODO: Better error
	- line 1206: # TODO: Better exception
	- line 1214: # TODO: Better exception
	- line 1220: # TODO: Better exception
	- line 1224: # TODO: Note that this can be a considerable overestimate. For example, if we
	- line 1229: # TODO: Make max_possibilities a global tweakable setting of the accumulator.
	- line 1235: # TODO: Better exception
	- line 1245: # TODO: Better error
	- line 1248: # TODO: What happens if we have a @functional that does not return
	- line 1249: # TODO: a graph node? A functional that returns a constant is
	- line 1250: # TODO: weird, but it should be legal.  Figure out what the
	- line 1251: # TODO: right thing to do is for this scenario.
	- line 1336: # TODO: Do we need to consider the scenario where we do not have
	- line 1342: # TODO: Handle kwargs
	- line 1422: # TODO: Do a sanity check that the arguments match and give
	- line 1423: # TODO: a good error if they do not. Alternatively, catch
	- line 1424: # TODO: the exception if the call fails and replace it with
	- line 1425: # TODO: a more informative error.
	- line 1436: # TODO: What happens if the function is a member of a class?
	- line 1437: # TODO: Do we recompile it for different instances of the
	- line 1438: # TODO: receiver? Do we recompile it on every call? Check this.
	- line 1447: # TODO: Better error message
	- line 1527: # TODO: Better error
	- line 1562: # TODO: Get this into alpha order
	- line 1569: # TODO: Better error
	- line 1584: # TODO: There are a great many more pure instance functions on tensors;
	- line 1585: # TODO: which do we wish to support?
	- line 1622: # TODO: Better error
	- line 1627: # TODO: Better error
	- line 1633: # TODO: Better error
	- line 1639: # TODO: Better error


src/beanmachine/ppl/compiler/single_assignment.py (31 lines):
	- line 9: # TODO: This module is badly named; it really should be "python simplifier" or some
	- line 25: #   * id = simple_expression  # TODO: List the simple expressions.
	- line 30: #   * [id] = id  # TODO: What about tuples?
	- line 48: #   * dict() is allowed. (TODO: this could be rewritten to {})
	- line 51: #   * TODO: There are similar exceptions for set and list; say what they are.
	- line 58: # TODO: assert statements are removed in bm_to_bmg; move that functionality here.
	- line 59: # TODO: We can reduce "del" statements to one of three forms: "del x", "del x.y", "del x[y]"
	- line 61: # TODO: Figure out how to desugar try: body except expr as bar: body else: body finally: body
	- line 63: # TODO: Figure out how to desugar with expr as target : block
	- line 65: # TODO: say something about global / nonlocal
	- line 66: # TODO: say something about yield
	- line 67: # TODO: say something about classes
	- line 68: # TODO: say something about type annotations
	- line 69: # TODO: say something about async
	- line 70: # TODO: say something about conditional expressions
	- line 71: # TODO: say something about formatted strings
	- line 147: # TODO: The identifier "dict" should be made global unique in target name space
	- line 180: # TODO: Better naming convention. In particular, _rules is the reflexive
	- line 215: # TODO: In the type this function, both instances of the type Any should
	- line 359: # TODO: The identifier "dict" should be made global unique in target name space
	- line 573: # TODO: Should we also eliminate plain returns? We could rewrite
	- line 611: # TODO: We can go further than this and eliminate all else clauses
	- line 654: # TODO: the "for" loop in Python supports an "else" clause which is only activated
	- line 657: # TODO: the "for" loop could be rewritten as fetching an iterator and iterating
	- line 1285: # TODO: Ideally, would like to merge [1].ctx with the [0].ctx below
	- line 1322: # TODO: Ideally, would like to merge [1].ctx with the [0].ctx below
	- line 1323: # TODO: The identifier "dict" should be made global unique in target name space
	- line 1389: # TODO: The identifier "dict" should be made global unique in target name space
	- line 1445: # TODO: The identifier "dict" should be made global unique in target name space
	- line 1540: # TODO: Note that this rule combined with the previous rule
	- line 2079: # TODO: We need to find a good way to order things in this file


src/beanmachine/ppl/compiler/bm_graph_builder.py (19 lines):
	- line 42: # TODO: The original idea was to use these integers when generating code
	- line 106: # TODO: The code in the "handle" methods which folds operations
	- line 113: # TODO: This should be private
	- line 124: # TODO: This is only used to remove an observation; restrict it
	- line 141: # TODO: This is only used to remove observation sample node that
	- line 350: # TODO: This code is mostly but not entirely in alpha order
	- line 469: # TODO: This code is not very well organized; consider sorting it
	- line 610: # TODO: We don't have a subtraction node; we render this as
	- line 723: # TODO: We could optimize -(-x) to x here.
	- line 744: # TODO: What should the result of NOT on a tensor be?
	- line 745: # TODO: Should it be legal at all in the graph?
	- line 746: # TODO: In Python, (not tensor(x)) is equal to (not x).
	- line 747: # TODO: It is NOT equal to (tensor(not x)), which is what
	- line 748: # TODO: you might expect.
	- line 909: # TODO: Verify that the list is well-formed.
	- line 920: # TODO: Should this be idempotent?
	- line 921: # TODO: Should it be an error to add two unequal observations to one node?
	- line 929: # TODO: BMG requires that the target of a query be classified
	- line 976: # TODO: The graph should be acyclic by construction;


src/beanmachine/graph/graph.cpp (19 lines):
	- line 213: // TODO: the following is used in beta.cpp only. Does it really need to be here?
	- line 308: // TODO: Investigate the semantics of _backward(skip_observed),
	- line 330: // TODO: used for testing only, should integrate it with
	- line 411: // TODO: As of May 2021, this method is being used for testing only.
	- line 439: // TODO: remove this error message and leave it to compute_gradients
	- line 446: // TODO: can we make signature use a default generator then?
	- line 457: // TODO clarify why we need to reset gradients
	- line 480: // TODO: also used in tests only
	- line 502: // TODO: this is the one actually used in code (as opposed to full_log_prob used
	- line 520: // TODO: rename log_abs_jacobian_determinant
	- line 537: /* TODO: used in testing only; it looks like there has not been a need for it in
	- line 549: // TODO: from now on, we have methods for adding nodes, checking validity,
	- line 555: // TODO: this does not have to apply to parents only; make it a more general
	- line 804: // TODO: Add checks that the observed value is in range.
	- line 833: // TODO: Add checks that the observed values are in range.
	- line 834: // TODO: Check that an observed simplex is given a simplex.
	- line 967: // TODO: clarify the meaning of log_prob_vals and log_prob_allchains
	- line 1044: // TODO: why don't the initialization below to be done for _infer?
	- line 1187: return variational_params; // TODO: this should have been defined as a field,


src/beanmachine/ppl/compiler/bmg_nodes.py (15 lines):
	- line 375: # TODO: We do not yet have a BMG node for Binomial
	- line 391: # TODO: we may wish to add bounded integers to the BMG type system.
	- line 488: # TODO: Add support for the Cauchy distribution as well.
	- line 571: # TODO: We do not yet have an implementation of the uniform
	- line 948: # TODO: Should we implement a general switch node in BMG?
	- line 956: # TODO: Check that cases are all constant nodes.
	- line 957: # TODO: Check that there is one value for each case.
	- line 1004: # TODO: We now have matrix multiplication in BMG; finish this implementation
	- line 1064: # TODO: If we have exp(x) - 1 in the graph and x is known to be of type
	- line 1099: # TODO: replace "log" with "log1mexp" as needed below and update defs
	- line 1330: # TODO: Here we treat an observation as node which takes input
	- line 1338: # TODO: **Observations are logically distinct from models.**
	- line 1350: # TODO: Consider adding a verification pass which ensures
	- line 1368: # TODO: BMG requires that the target of a query be classified
	- line 1374: # TODO: As with observations, properly speaking there is no


src/beanmachine/ppl/compiler/bm_to_bmg.py (11 lines):
	- line 53: # TODO: Detect unsupported operators
	- line 54: # TODO: Detect unsupported control flow
	- line 55: # TODO: Would be helpful if we could track original source code locations.
	- line 405: # TODO: Add classes, lambdas, and so on
	- line 440: # TODO: Handle the situation if there are tabs
	- line 503: # TODO: Consider producing a warning or error if we cannot determine which lambda
	- line 506: # TODO: return None if we are unable to get the source
	- line 564: # TODO: return None if we are unable to get the source
	- line 687: # TODO: Consider detecting jitted functions which use the nonlocal statement, and
	- line 690: # TODO: rename bmg outer variable to something less likely
	- line 841: # TODO: This method is only here for testing purposes.  Get rid of it.


src/beanmachine/ppl/compiler/fix_unsupported.py (11 lines):
	- line 47: # TODO: Suppose we have something like Uniform(1.0, 2.0).  Can we replace that
	- line 120: # TODO: If left is a ToMatrixNode then we can construct a second
	- line 208: # TODO: If we have a contiguous set of cases, say {2, 3, 4}, then
	- line 212: # TODO: If we have a slightly noncontiguous set of cases, say {0, 1, 3},
	- line 216: # TODO: If we have arbitrary natural cases, say 1, 10, 101, then we could
	- line 257: # TODO: Consider producing a warning for this situation, because
	- line 276: # TODO: Generate a better error message for switches that we cannot yet
	- line 295: # TODO:
	- line 345: # TODO: We should make a rewriter that detects stochastic index
	- line 357: # TODO: We have an ordering problem in the fixers:
	- line 394: # TODO: The edge labels used to visualize the graph in DOT


src/beanmachine/ppl/compiler/fix_vectorized_models.py (9 lines):
	- line 14: # TODO Move this to a utils module
	- line 38: # TODO: Consider optimizing distributions where the tensor elements are all
	- line 62: # TODO: categorical
	- line 63: # TODO: categorical logit
	- line 64: # TODO: dirichlet
	- line 81: # TODO: Do addition and multiply need to be the multi- versions?
	- line 92: # TODO soon: LogSumExp, ExpM1, Logm1exp
	- line 93: # TODO later: all comparisons, all bitwise, floordiv,
	- line 235: # TODO: What if the observation is of a different size than the


src/beanmachine/ppl/compiler/bmg_requirements.py (8 lines):
	- line 130: # TODO: BMG requires it to be a *broadcast* matrix; what happens
	- line 141: # TODO: This will produce a bad error message experience. What we want to
	- line 184: # TODO: The current code produces a bad error message experience because at
	- line 231: # TODO: Consider what to do if the node type is Tensor.
	- line 237: # TODO: Consider what to do if the node type is Tensor.
	- line 286: # TODO: We might consider an optimization pass which does so.
	- line 334: # TODO: We need a better error message for that scenario.
	- line 370: # TODO: What about simplex matrices? It should be legal to do


src/beanmachine/ppl/compiler/patterns.py (7 lines):
	- line 30: # TODO: Tensor comprehension patterns
	- line 62: # TODO: Display as a tree and graph also
	- line 68: # TODO: If we save the patterns that failed as well, then we can build a
	- line 69: # TODO: diagnostic engine that describes why a value failed to match against
	- line 70: # TODO: a complex pattern.
	- line 592: # TODO: We could bail out after the first success.
	- line 615: # TODO: We could bail out after the first failure.


src/beanmachine/graph/graph.h (6 lines):
	- line 473: // TODO: shouldn't we then restrict them to those classes? See below.
	- line 481: // TODO: shouldn't we then restrict them to those classes? See above.
	- line 795: // TODO: This public method returns a pointer to an internal data structure
	- line 802: // TODO: create Samples class and remove the following friend classes
	- line 812: // TODO: a lot of members used to be private, but we need access to them
	- line 871: // TODO: Review what members of this class can be made static.


src/beanmachine/ppl/experimental/vi/variational_infer.py (6 lines):
	- line 68: # TODO: reinterpret batch dimension?
	- line 144: # TODO: caused by e.g. negative scales in `dist.Normal`;
	- line 291: # TODO: add new `self.world_.params_` not already in optimizer
	- line 319: # TODO: more general enumeration
	- line 326: # TODO: downstream observation likelihoods p(obs | rvid)
	- line 345: # TODO: caused by e.g. negative scales in `dist.Normal`;


src/beanmachine/ppl/compiler/fix_requirements.py (5 lines):
	- line 174: # TODO: We do not yet handle the case where we are converting from, say,
	- line 184: # TODO: We do not yet handle the case where we are converting from
	- line 211: # TODO: We might want to restrict this. For example, if we have
	- line 219: # TODO: We might want to build a warning mechanism that informs the developer
	- line 404: # TODO: The edge labels used to visualize the graph in DOT


src/beanmachine/ppl/compiler/bmg_types.py (5 lines):
	- line 60: # TODO: We might also need:
	- line 656: # TODO: add a range check to make sure it fits into the integer
	- line 688: # TODO: Move this to bmg_requirements.py
	- line 730: # TODO: Mark this as abstract
	- line 740: # TODO: Memoize these, remove memoization of construction functions below.


src/beanmachine/ppl/compiler/lattice_typer.py (5 lines):
	- line 231: # TODO: Consider adding a pass which optimizes away IF(X, Y, Y) to
	- line 233: # TODO: What if we have an atomic type on one side and a 1x1 matrix
	- line 238: # TODO: We need to consider what happens if the consequence and alternative
	- line 251: # TODO: We need to consider what happens if the value's types
	- line 394: # TODO: Consider asserting that the node is unsupported by BMG.


src/beanmachine/graph/to_dot.cpp (4 lines):
	- line 57: // TODO: Better display of matrices
	- line 62: // TODO: Better display of simplexes
	- line 202: // TODO: Include value of observation
	- line 236: // TODO: Better display of factors.


src/beanmachine/ppl/compiler/sizer.py (4 lines):
	- line 150: # TODO: Is there a better way to do this other than try it and see what happens?
	- line 151: # TODO: Try torch.distributions.utils.broadcast_all
	- line 202: # TODO:
	- line 234: # TODO: Torch supports both broadcasting and non-broadcasting versions


src/beanmachine/ppl/compiler/support.py (4 lines):
	- line 81: # TODO: We could do better for the comparison operators because we known
	- line 120: # TODO:
	- line 241: # TODO: This approximation is an over-estimate; for instance, when
	- line 302: # TODO: Move this prod helper function out of bmg_nodes.py


src/beanmachine/ppl/inference/bmg_inference.py (4 lines):
	- line 73: # TODO: Figure out a better way to pass this flag around
	- line 224: # TODO: Make _rv_to_query public. Add it to BMGraphBuilder?
	- line 261: # TODO: Add verbose level
	- line 262: # TODO: Add logging


src/beanmachine/ppl/compiler/error_report.py (4 lines):
	- line 21: # TODO: We use the DOT label of the node in the error messages here,
	- line 55: # TODO: Fix this error message for the case where we require
	- line 98: # TODO: Improve wording and diagnosis.
	- line 118: # TODO: Better error


src/beanmachine/ppl/compiler/fix_observations.py (4 lines):
	- line 52: # TODO: How should we deal with observations of
	- line 53: # TODO: matrix-valued samples?
	- line 55: # TODO: Handle the case where there are two inconsistent
	- line 56: # TODO: observations of the same sample


src/beanmachine/ppl/compiler/ast_patterns.py (3 lines):
	- line 148: # TODO: what should we do about AnnAssign?
	- line 478: # TODO: Matches "tensor" and "foo.tensor"
	- line 479: # TODO: Do we need to specifically match just torch? What if there is an alias?


src/beanmachine/ppl/compiler/bmg_node_types.py (3 lines):
	- line 6: # TODO: For reasons unknown, Pyre is unable to find type information about
	- line 7: # TODO: beanmachine.graph from beanmachine.ppl.  I'll figure out why later;
	- line 8: # TODO: for now, we'll just turn off error checking in this module.


src/beanmachine/ppl/compiler/profiler.py (3 lines):
	- line 50: # TODO: Sort by total time of children - WARNING: Important not to sort by
	- line 52: # TODO: compute unattributed via property
	- line 58: # TODO: There are two shortcomings to the current solution. First, it prints


src/beanmachine/ppl/experimental/inference_compilation/ic_infer.py (3 lines):
	- line 390: # TODO: bottlenecking?
	- line 449: # TODO: try pooling rather than just slicing out last hidden
	- line 509: # TODO: non-diagonal covariance


src/beanmachine/ppl/experimental/vi/mean_field_variational_approximation.py (3 lines):
	- line 90: # TODO: remove once fixed upstream
	- line 93: # TODO: This needs to be now set after self.new_dist has been instantiated
	- line 165: # TODO: do we need to a _sum_rightmost here?


src/beanmachine/graph/global/global_state.cpp (2 lines):
	- line 29: // TODO: rename to initialize_unconstrained_value_types
	- line 73: sto_node->get_unconstrained_value(true); // TODO: rename this function


website/documentation.toml (2 lines):
	- line 13: # TODO: Base path for Docusaurus 2 content
	- line 16: # TODO: regular expression for sections that are visible initially: unfurled = "..."


src/beanmachine/ppl/compiler/fix_beta_conjugate_prior.py (2 lines):
	- line 26: # TODO: For now we support conjugate prior transformation on
	- line 36: # TODO: This check can be removed if it is not a necessary condition.


src/beanmachine/ppl/legacy/inference/abstract_mh_infer.py (2 lines):
	- line 323: # TODO adjust log density on numerical runtime errors
	- line 387: # TODO: Consider allowing values that can be converted to tensors.


src/beanmachine/ppl/compiler/fix_normal_conjugate_prior.py (2 lines):
	- line 29: # TODO: For now we support conjugate prior transformation on
	- line 40: # TODO: This check can be removed if it is not a necessary condition.


src/beanmachine/graph/global/util.cpp (2 lines):
	- line 19: // TODO: add simplex transform
	- line 29: // TODO: rename method to be more clear


src/beanmachine/ppl/diagnostics/common_statistics.py (2 lines):
	- line 41: """TODO have separate diagnostics for discrete variables.
	- line 79: """TODO have separate diagnostics for discrete variables.


src/beanmachine/ppl/testlib/hypothesis_testing.py (2 lines):
	- line 66: # TODO: Consider refactoring the common input checks for both methods
	- line 78: # TODO: We use z_{1-alpha} instead of -z_alpha for compatibility


src/beanmachine/ppl/compiler/rules.py (2 lines):
	- line 292: # TODO: Consider implementing some sort of warning for this case?
	- line 686: # TODO: always_succeeds will be wrong for the returned object.


src/beanmachine/graph/operator/linalgop.cpp (2 lines):
	- line 180: // TODO: change type to ranged natural
	- line 226: // TODO: change type to ranged natural


src/beanmachine/graph/mh.h (2 lines):
	- line 90: // TODO: this should be delegated to steppers, since that's
	- line 94: // TODO: review what really needs to be private or public in MH's API.


src/beanmachine/applications/hme/abstract_model.py (2 lines):
	- line 131: # TODO: add support for data pre-processing on random effects
	- line 144: # TODO: add support for random slope


src/beanmachine/ppl/utils/memoize.py (2 lines):
	- line 174: # TODO: We do not collect statistics on memoization use here.
	- line 175: # TODO: We do not canonicalize arguments as the memoizer does above.


src/beanmachine/ppl/compiler/fix_bool_arithmetic.py (1 line):
	- line 36: # TODO: We could support b ** n where b is bool and n is a natural


src/beanmachine/ppl/compiler/internal_error.py (1 line):
	- line 84: # TODO: Consider adding a compiler version number, hash


src/beanmachine/graph/stepper/single_site/nmc_scalar_single_site_stepping_method.cpp (1 line):
	- line 54: // TODO: generalize so it works with any proposer, not just nmc_proposer:


src/beanmachine/graph/operator/operator.h (1 line):
	- line 38: // TODO: should we just remove it from Operator then?


src/beanmachine/ppl/compiler/gen_bmg_graph.py (1 line):
	- line 107: # TODO: We could consider traversing only nodes reachable from


src/beanmachine/graph/distribution/categorical.cpp (1 line):
	- line 84: // TODO: Yeah, how?


src/beanmachine/graph/global/global_mh.cpp (1 line):
	- line 22: // TODO: tie samples directly to inference


src/beanmachine/ppl/compiler/beanstalk_common.py (1 line):
	- line 10: # TODO: Allowing these constructions raises additional problems that


src/beanmachine/applications/hme/null_mixture_model.py (1 line):
	- line 232: # FIXME: update the method to support customized prediction goal


src/beanmachine/graph/proposer/proposer.h (1 line):
	- line 11: // TODO: should change name to "Proposal" to better match the literature.


src/beanmachine/ppl/compiler/fix_problems.py (1 line):
	- line 45: # TODO: Add other notes on ordering constraints here.


src/beanmachine/ppl/compiler/fix_multiary_ops.py (1 line):
	- line 33: # TODO: This is a bit opaque. Add a helper method for this.


src/beanmachine/ppl/compiler/fix_bool_comparisons.py (1 line):
	- line 134: # TODO: Should we treat "x is y" the same as "x == y" when they are


scripts/convert_ipynb_to_mdx.py (1 line):
	- line 210: # TODO: Handle svg images.


src/beanmachine/ppl/experimental/gp/models.py (1 line):
	- line 118: # TODO: Make sure observation noise is transformed correctly


src/beanmachine/ppl/legacy/world/world.py (1 line):
	- line 772: # TODO: messy, consider strategy pattern


src/beanmachine/ppl/inference/proposer/single_site_uniform_proposer.py (1 line):
	- line 24: # TODO: we should sample from a transformed dist


src/beanmachine/graph/stepper/single_site/default_single_site_stepping_method.cpp (1 line):
	- line 89: // TODO: the reason for that is not clear;


sphinx/source/docs.py (1 line):
	- line 199: # TODO: Would be slightly faster if we applied module filtering inside walk_packages


src/beanmachine/graph/mh.cpp (1 line):
	- line 227: // TODO: eventually we want to have different classes of Node


src/beanmachine/ppl/compiler/typer_base.py (1 line):
	- line 38: #   type information has already been cached. (TODO: Is this useful? Maybe remove it.)


src/beanmachine/graph/stepper/single_site/nmc_dirichlet_gamma_single_site_stepping_method.cpp (1 line):
	- line 107: // TODO: identify code that depends on this, let it zero gradients


src/beanmachine/graph/proposer/default_initializer.cpp (1 line):
	- line 25: // TODO: it seems like the initializer is needing to know


src/beanmachine/ppl/compiler/gen_dot.py (1 line):
	- line 41: # TODO: It is strange having a visualizer that edits the graph