You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.
support such an operator in CachedOp. This step requires writing a new executor that runs operators in a computation graph one by one without shape inference and memory planning. We can start with executing operators synchronously, but eventually, we should execute them asynchronously.
an optimization is to use the subgraph API to split the computation graph into subgraphs so that a subgraph either contain only operators allowing static shape inference or operators with dynamic shape.
The text was updated successfully, but these errors were encountered:
I think that a good solution to support pseudo-dynamic (in most cases it's enough) shape is to support infer_shape_partial (through defer shape inference) in HybridBlock.
MXNet doesn't support operators that don't allow static shape inference, such as unique and boolean_index. One possible solution is proposed in https://cwiki.apache.org/confluence/display/MXNET/Dynamic+shape
This task can be broken into multiple steps:
The text was updated successfully, but these errors were encountered: