tensor.basic_opt
– Tensor Optimizations¶
Tensor optimizations addressing the ops in basic.py.
-
class
aesara.tensor.basic_opt.
FusionOptimizer
(local_optimizer)[source]¶ Graph optimizer that simply runs local fusion operations.
TODO: This is basically a
EquilibriumOptimizer
; we should just use that.-
apply
(fgraph)[source]¶ Apply the optimization to a
FunctionGraph
.It may use all the methods defined by the
FunctionGraph
. If theGlobalOptimizer
needs to use a certain tool, such as anInstanceFinder
, it can do so in itsadd_requirements
method.
-
-
class
aesara.tensor.basic_opt.
InplaceElemwiseOptimizer
(OP)[source]¶ This is parameterized so that it works for
Elemwise
Op
s.-
apply
(fgraph)[source]¶ Usage: InplaceElemwiseOptimizer(op).optimize(fgraph)
Attempts to replace all Broadcast ops by versions of them that operate inplace. It operates greedily: for each Broadcast Op that is encountered, for each output, tries each input to see if it can operate inplace on that input. If so, makes the change and go to the next output or Broadcast Op.
Examples
x + y + z -> x += y += z
(x + y) * (x * y) -> (x += y) *= (x * y) or (x + y) *= (x *= y)
-
-
class
aesara.tensor.basic_opt.
ShapeFeature
[source]¶ Graph optimizer for removing all calls to shape().
This optimizer replaces all Shapes and Subtensors of Shapes with Shape_i and MakeVector Ops.
This optimizer has several goals:
- to ‘lift’ Shapes to as close to the inputs as possible.
- to infer the shape of every node in the graph in terms of the input shapes.
- remove all fills
(at.second, at.fill)
from the graph
Lifting shapes as close to the inputs as possible is important for canonicalization because it is very bad form to have to compute something just to know how big it will be. Firstly, it is a waste of time to compute such outputs. But it is important to get rid of these outputs as early as possible in the compilation process because the extra computations make it appear as if many internal graph nodes have multiple clients. Many optimizations refuse to work on nodes with multiple clients.
Lifting is done by using an
<Op>.infer_shape
function if one is present, or else using a conservative default. An Op that supports shape-lifting should define a infer_shape(self, fgraph, node, input_shapes) function. The argument input_shapes is a tuple of tuples… there is an interior tuple for each input to the node. The tuple has as many elements as dimensions. The element in position i of tuple j represents the i’th shape component of the j’th input. The function should return a tuple of tuples. One output tuple for each node.output. Again, the i’th element of the j’th output tuple represents the output[j].shape[i] of the function. If an output is not a TensorType, then None should be returned instead of a tuple for that output.For example the infer_shape for a matrix-matrix product would accept input_shapes=((x0,x1), (y0,y1)) and return ((x0, y1),).
Inferring the shape of internal nodes in the graph is important for doing size-driven optimizations. If we know how big various intermediate results will be, we can estimate the cost of many Ops accurately, and generate c-code that is specific [e.g. unrolled] to particular sizes.
In cases where you cannot figure out the shape, raise a ShapeError.
Notes
Right now there is only the ConvOp that could really take advantage of this shape inference, but it is worth it even just for the ConvOp. All that’s necessary to do shape inference is 1) to mark shared inputs as having a particular shape, either via a .tag or some similar hacking; and 2) to add an optional In() argument to promise that inputs will have a certain shape (or even to have certain shapes in certain dimensions). We can’t automatically infer the shape of shared variables as they can change of shape during the execution by default. (NOT IMPLEMENTED YET, BUT IS IN TRAC)
Using Shape information in Optimizations
To use this shape information in OPTIMIZATIONS, use the
shape_of
dictionary.For example:
try: shape_of = fgraph.shape_feature.shape_of except AttributeError: # This can happen when the mode doesn't include the ShapeFeature. return shape_of_output_zero = shape_of[node.output[0]]
The
shape_of_output_zero
symbol will contain a tuple, whose elements are either integers or symbolic integers.TODO: check to see if the symbols are necessarily non-constant… or are integer literals sometimes Aesara constants?? That would be confusing.
-
clone
()[source]¶ Create a clone that can be attached to a new
FunctionGraph
.This default implementation returns
self
, which carries the assumption that theFeature
is essentially stateless. If a subclass has state of its own that is in any way relative to a givenFunctionGraph
, this method should be overridden with an implementation that actually creates a fresh copy.
-
default_infer_shape
(fgraph, node, i_shapes)[source]¶ Return a list of shape tuple or None for the outputs of node.
This function is used for Ops that don’t implement infer_shape. Ops that do implement infer_shape should use the i_shapes parameter, but this default implementation ignores it.
-
get_shape
(var, idx)[source]¶ Optimization can call this to get the current shape_i
It is better to call this then use directly shape_of[var][idx] as this method should update shape_of if needed.
TODO: Up to now, we don’t update it in all cases. Update in all cases.
-
on_attach
(fgraph)[source]¶ Called by
FunctionGraph.attach_feature
, the method that attaches the feature to theFunctionGraph
. Since this is called after theFunctionGraph
is initially populated, this is where you should run checks on the initial contents of theFunctionGraph
.The on_attach method may raise the
AlreadyThere
exception to cancel the attach operation if it detects that another Feature instance implementing the same functionality is already attached to theFunctionGraph
.The feature has great freedom in what it can do with the
fgraph
: it may, for example, add methods to it dynamically.
-
on_change_input
(fgraph, node, i, r, new_r, reason)[source]¶ Called whenever
node.inputs[i]
is changed fromvar
tonew_var
. At the moment the callback is done, the change has already taken place.If you raise an exception in this function, the state of the graph might be broken for all intents and purposes.
-
on_detach
(fgraph)[source]¶ Called by
FunctionGraph.remove_feature
. Should remove any dynamically-added functionality that it installed into the fgraph.
-
on_import
(fgraph, node, reason)[source]¶ Called whenever a node is imported into
fgraph
, which is just before the node is actually connected to the graph.Note: this is not called when the graph is created. If you want to detect the first nodes to be implemented to the graph, you should do this by implementing
on_attach
.
-
same_shape
(x: aesara.graph.basic.Variable, y: aesara.graph.basic.Variable, dim_x: Optional[int] = None, dim_y: Optional[int] = None) bool [source]¶ Return
True
ifx
andy
have the same shape.Parameters: - x – The
Variable
for which its shape is to be compared withy
’s shape. - y – The
Variable
for which its shape is to be compared withx
’s shape. - dim_x – If non
None
, compare only the dimension ofx
equal todim_x
. - dim_y – If non
None
, compare only the dimension ofy
equal todim_y
.
- x – The
-
set_shape
(r, s, override=False)[source]¶ Assign the shape
s
to previously un-shaped variabler
.Parameters: - r (a variable) –
- s (None or a tuple of symbolic integers) –
- override (If False, it mean r is a new object in the fgraph.) – If True, it mean r is already in the fgraph and we want to override its shape.
-
class
aesara.tensor.basic_opt.
ShapeOptimizer
[source]¶ Optimizer that adds
ShapeFeature
as a feature.-
apply
(fgraph)[source]¶ Apply the optimization to a
FunctionGraph
.It may use all the methods defined by the
FunctionGraph
. If theGlobalOptimizer
needs to use a certain tool, such as anInstanceFinder
, it can do so in itsadd_requirements
method.
-
-
class
aesara.tensor.basic_opt.
UnShapeOptimizer
[source]¶ Optimizer that removes
ShapeFeature
as a feature.-
apply
(fgraph)[source]¶ Apply the optimization to a
FunctionGraph
.It may use all the methods defined by the
FunctionGraph
. If theGlobalOptimizer
needs to use a certain tool, such as anInstanceFinder
, it can do so in itsadd_requirements
method.
-
-
aesara.tensor.basic_opt.
apply_rebroadcast_opt
(rval)[source]¶ Apply as many times as required the optimization local_useless_rebroadcast and local_rebroadcast_lift.
Parameters: rval (a Variable) – Return type: A Variable (the same if no optimization can be applied)
-
aesara.tensor.basic_opt.
broadcast_like
(value, template, fgraph, dtype=None)[source]¶ Return a Variable with the same shape and dtype as the template, filled by broadcasting value through it.
value
will be cast as necessary.
-
aesara.tensor.basic_opt.
encompasses_broadcastable
(b1, b2)[source]¶ Parameters: - b1 – The broadcastable attribute of a tensor type.
- b2 – The broadcastable attribute of a tensor type.
Returns: True if the broadcastable patterns b1 and b2 are such that b2 is broadcasted to b1’s shape and not the opposite.
Return type: bool
-
aesara.tensor.basic_opt.
is_an_upcast
(type1, type2)[source]¶ Given two data types (as strings), check if converting to type2 from type1 constitutes an upcast. Differs from aesara.scalar.upcast
-
aesara.tensor.basic_opt.
is_dimshuffle_useless
(new_order, input)[source]¶ - Checks for two types of useless dimshuffles:
- 1 - dimshuffle all dimensions in order. 2 - dimshuffle a broadcastable dimension.
-
aesara.tensor.basic_opt.
local_elemwise_fusion
(fgraph, node)[source]¶ Fuse
Elemwise
Op
s in a node.As part of specialization, we fuse two consecutive
Elemwise
Op
s of the same shape.For mixed dtype, we let the
Composite
Op
do the cast. It lets the C compiler do the cast.The number of dimensions is validated at call time by Aesara itself.
-
aesara.tensor.basic_opt.
local_elemwise_fusion_op
(op_class, max_input_fct=<function <lambda>>, maker=None)[source]¶ Create a recursive function that fuses
Elemwise
Op
s.The basic idea is that we loop through an
Elemwise
node’s inputs, find otherElemwise
nodes, determine the scalars input types for all of theElemwise
Op
s, construct a new scalarOp
using the scalar input types and eachElemwise
’s scalarOp
, and use the composite scalarOp
in a new “fused”Elemwise
.It’s parameterized in order to work for
Elemwise
Op
s.Parameters: - op_class (type) –
Elemwise
class (the one that we want to fuse) - max_input_fct (callable) – A function that returns the maximum number of inputs that this
Elemwise
can take. On the CPU we limit to 32 input variables since that is the maximum NumPy support. - maker (callable) – A function with the signature
(node, *args)
that constructs anop_class
instance (e.g.op_class(*args)
).
- op_class (type) –