""" Dependency-Graph representation of Pipeline API terms. """ from networkx import ( DiGraph, topological_sort, ) from six import iteritems, itervalues from zipline.utils.memoize import lazyval from zipline.pipeline.visualize import display_graph from .term import LoadableTerm class CyclicDependency(Exception): pass class TermGraph(DiGraph): """ An abstract representation of Pipeline Term dependencies. This class does not keep any additional metadata about any term relations other than dependency ordering. As such it is only useful in contexts where you care exclusively about order properties (for example, when drawing visualizations of execution order). Parameters ---------- terms : dict A dict mapping names to final output terms. Attributes ---------- outputs Methods ------- ordered() Return a topologically-sorted iterator over the terms in self. See Also -------- ExecutionPlan """ def __init__(self, terms): super(TermGraph, self).__init__() self._frozen = False parents = set() for term in itervalues(terms): self._add_to_graph(term, parents) # No parents should be left between top-level terms. assert not parents self._outputs = terms self._ordered = topological_sort(self) # Mark that no more terms should be added to the graph. self._frozen = True def _add_to_graph(self, term, parents): """ Add a term and all its children to ``graph``. ``parents`` is the set of all the parents of ``term` that we've added so far. It is only used to detect dependency cycles. """ if self._frozen: raise ValueError( "Can't mutate %s after construction." % type(self).__name__ ) # If we've seen this node already as a parent of the current traversal, # it means we have an unsatisifiable dependency. This should only be # possible if the term's inputs are mutated after construction. if term in parents: raise CyclicDependency(term) parents.add(term) self.add_node(term) for dependency in term.dependencies: self._add_to_graph(dependency, parents) self.add_edge(dependency, term) parents.remove(term) @property def outputs(self): """ Dict mapping names to designated output terms. """ return self._outputs def ordered(self): """ Return a topologically-sorted iterator over the terms in `self`. """ return iter(self._ordered) @lazyval def loadable_terms(self): return tuple(term for term in self if isinstance(term, LoadableTerm)) @lazyval def jpeg(self): return display_graph(self, 'jpeg') @lazyval def png(self): return display_graph(self, 'png') @lazyval def svg(self): return display_graph(self, 'svg') def _repr_png_(self): return self.png.data def initial_refcounts(self, initial_terms): """ Calculate initial refcounts for execution of this graph. Parameters ---------- initial_terms : iterable[Term] An iterable of terms that were pre-computed before graph execution. Each node starts with a refcount equal to its outdegree, and output nodes get one extra reference to ensure that they're still in the graph at the end of execution. """ refcounts = self.out_degree() for t in self.outputs.values(): refcounts[t] += 1 for t in initial_terms: self.decref_dependencies(t, refcounts) return refcounts def decref_dependencies(self, term, refcounts): """ Decrement in-edges for ``term`` after computation. Parameters ---------- term : zipline.pipeline.Term The term whose parents should be decref'ed. refcounts : dict[Term -> int] Dictionary of refcounts. Return ------ garbage : set[Term] Terms whose refcounts hit zero after decrefing. """ garbage = set() # Edges are tuple of (from, to). for parent, _ in self.in_edges([term]): refcounts[parent] -= 1 # No one else depends on this term. Remove it from the # workspace to conserve memory. if refcounts[parent] == 0: garbage.add(parent) return garbage class ExecutionPlan(TermGraph): """ Graph represention of Pipeline Term dependencies that includes metadata about extra rows required to perform computations. Each node in the graph has an `extra_rows` attribute, indicating how many, if any, extra rows we should compute for the node. Extra rows are most often needed when a term is an input to a rolling window computation. For example, if we compute a 30 day moving average of price from day X to day Y, we need to load price data for the range from day (X - 29) to day Y. Parameters ---------- terms : dict A dict mapping names to final output terms. all_dates : pd.DatetimeIndex An index of all known trading days for which ``terms`` will be computed. start_date : pd.Timestamp The first date for which output is requested for ``terms``. end_date : pd.Timestamp The last date for which output is requested for ``terms``. Attributes ---------- outputs offset extra_rows Methods ------- ordered() Return a topologically-sorted iterator over the terms in self. """ def __init__(self, terms, all_dates, start_date, end_date, min_extra_rows=0): super(ExecutionPlan, self).__init__(terms) for term in terms.values(): self.set_extra_rows( term, all_dates, start_date, end_date, min_extra_rows=min_extra_rows, ) def set_extra_rows(self, term, all_dates, start_date, end_date, min_extra_rows): """ Compute ``extra_rows`` for transitive dependencies of ``root_terms`` """ # A term can require that additional extra rows beyond the minimum be # computed. This is most often used with downsampled terms, which need # to ensure that the first date is a computation date. extra_rows_for_term = term.compute_extra_rows( all_dates, start_date, end_date, min_extra_rows, ) if extra_rows_for_term < min_extra_rows: raise ValueError( "term %s requested fewer rows than the minimum of %d" % ( term, min_extra_rows, ) ) self._ensure_extra_rows(term, extra_rows_for_term) for dependency, additional_extra_rows in term.dependencies.items(): self.set_extra_rows( dependency, all_dates, start_date, end_date, min_extra_rows=extra_rows_for_term + additional_extra_rows, ) @lazyval def offset(self): """ For all pairs (term, input) such that `input` is an input to `term`, compute a mapping:: (term, input) -> offset(term, input) where ``offset(term, input)`` is the number of rows that ``term`` should truncate off the raw array produced for ``input`` before using it. We compute this value as follows:: offset(term, input) = (extra_rows_computed(input) - extra_rows_computed(term) - requested_extra_rows(term, input)) Examples -------- Case 1 ~~~~~~ Factor A needs 5 extra rows of USEquityPricing.close, and Factor B needs 3 extra rows of the same. Factor A also requires 5 extra rows of USEquityPricing.high, which no other Factor uses. We don't require any extra rows of Factor A or Factor B We load 5 extra rows of both `price` and `high` to ensure we can service Factor A, and the following offsets get computed:: offset[Factor A, USEquityPricing.close] == (5 - 0) - 5 == 0 offset[Factor A, USEquityPricing.high] == (5 - 0) - 5 == 0 offset[Factor B, USEquityPricing.close] == (5 - 0) - 3 == 2 offset[Factor B, USEquityPricing.high] raises KeyError. Case 2 ~~~~~~ Factor A needs 5 extra rows of USEquityPricing.close, and Factor B needs 3 extra rows of Factor A, and Factor B needs 2 extra rows of USEquityPricing.close. We load 8 extra rows of USEquityPricing.close (enough to load 5 extra rows of Factor A), and the following offsets get computed:: offset[Factor A, USEquityPricing.close] == (8 - 3) - 5 == 0 offset[Factor B, USEquityPricing.close] == (8 - 0) - 2 == 6 offset[Factor B, Factor A] == (3 - 0) - 3 == 0 Notes ----- `offset(term, input) >= 0` for all valid pairs, since `input` must be an input to `term` if the pair appears in the mapping. This value is useful because we load enough rows of each input to serve all possible dependencies. However, for any given dependency, we only want to compute using the actual number of required extra rows for that dependency. We can do so by truncating off the first `offset` rows of the loaded data for `input`. See Also -------- zipline.pipeline.graph.TermGraph.offset zipline.pipeline.engine.SimplePipelineEngine._inputs_for_term zipline.pipeline.engine.SimplePipelineEngine._mask_and_dates_for_term """ extra = self.extra_rows return { # Another way of thinking about this is: # How much bigger is the array for ``dep`` compared to ``term``? # How much of that difference did I ask for. (term, dep): (extra[dep] - extra[term]) - requested_extra_rows for term in self for dep, requested_extra_rows in term.dependencies.items() } @lazyval def extra_rows(self): """ A dict mapping `term` -> `# of extra rows to load/compute of `term`. Notes ---- This value depends on the other terms in the graph that require `term` **as an input**. This is not to be confused with `term.dependencies`, which describes how many additional rows of `term`'s inputs we need to load, and which is determined entirely by `Term` itself. Example ------- Our graph contains the following terms: A = SimpleMovingAverage([USEquityPricing.high], window_length=5) B = SimpleMovingAverage([USEquityPricing.high], window_length=10) C = SimpleMovingAverage([USEquityPricing.low], window_length=8) To compute N rows of A, we need N + 4 extra rows of `high`. To compute N rows of B, we need N + 9 extra rows of `high`. To compute N rows of C, we need N + 7 extra rows of `low`. We store the following extra_row requirements: self.extra_rows[high] = 9 # Ensures that we can service B. self.extra_rows[low] = 7 See Also -------- zipline.pipeline.graph.TermGraph.offset zipline.pipeline.term.Term.dependencies """ return { term: attrs['extra_rows'] for term, attrs in iteritems(self.node) } def _ensure_extra_rows(self, term, N): """ Ensure that we're going to compute at least N extra rows of `term`. """ attrs = self.node[term] attrs['extra_rows'] = max(N, attrs.get('extra_rows', 0))