search

KULeuven-MICAS / stream

Multi-core HW accelerator mapping optimization framework for layer-fused ML workloads.
https://kuleuven-micas.github.io/stream/
MIT License
38 stars 18 forks source link

Is this necessary? #37

Closed Andful closed 1 month ago

Andful commented 2 months ago

https://github.com/KULeuven-MICAS/stream/blob/6ac88776d1b9e5698f39014500f3d0ae7cde20a7/stream/classes/io/onnx/default.py#L21

Is this necessary? This breaks the CONCAT DNN operations, such as the one in SqueezeNet. Was it working before this change was introduces?

asyms commented 1 month ago

This is intended as the DummyNode should only exist for 'simple' operators where the dependencies can just be propagated. A concat operator requires more complex dependency propagation. This has since been changed. See: