Adds a new custom op and transformation to partition graphs into multiple ONNX models. Supports non-linear models and partitions with multiple inputs/outputs.
There are two ways to define a partitioning:
PartitionFromLambda(): Takes a function that maps node -> partition_id.
PartitionFromDict(): Wrapper that takes a dictionary to partition nodes based on their graph index.
Other changes:
Mount build directory within docker container. This is the default storage location for new partitions.
Remove StreamingDataflowPartition from finn-base. It will be moved to the fpgadataflow domain.
maltanar
commented
3 years ago
Adds a new custom op and transformation to partition graphs into multiple ONNX models. Supports non-linear models and partitions with multiple inputs/outputs.
There are two ways to define a partitioning:
Other changes: