Closed JiacongSun closed 3 months ago
The onnx parser will generate broken graphs (for such as mobilenetv2.onnx
, resnet18.onnx
) and need to be fixed. The visualization function below is enabled for debugging.
https://github.com/KULeuven-MICAS/zigzag/blob/4e923c3b0c451de2b4afaf5f874664948d05fe0a/zigzag/parser/onnx/ONNXModelParser.py#L97
@asyms @JiacongSun As discussed, we would also like to rename these stages to something that describes better what it does. Are there any suggestions? What about RemoveLayerTransferOverhead
?
@asyms @JiacongSun As discussed, we would also like to rename these stages to something that describes better what it does. Are there any suggestions? What about
RemoveLayerTransferOverhead
?
I feel RemoveLayerTransferOverhead
or RemoveLayerTransferOverheadStage
is less intuitive compared to the original RemoveUnusedMemoryStage
. I prefer to keep the original name or stress more on the point of data locality (such as MemoryDataLocalityStage
?)
@asyms @JiacongSun As discussed, we would also like to rename these stages to something that describes better what it does. Are there any suggestions? What about
RemoveLayerTransferOverhead
?I feel
RemoveLayerTransferOverhead
orRemoveLayerTransferOverheadStage
is less intuitive compared to the originalRemoveUnusedMemoryStage
. I prefer to keep the original name or stress more on the point of data locality (such asMemoryDataLocalityStage
?)
What about ExploitInterLayerDataLocalityStage
?
@asyms @JiacongSun As discussed, we would also like to rename these stages to something that describes better what it does. Are there any suggestions? What about
RemoveLayerTransferOverhead
?I feel
RemoveLayerTransferOverhead
orRemoveLayerTransferOverheadStage
is less intuitive compared to the originalRemoveUnusedMemoryStage
. I prefer to keep the original name or stress more on the point of data locality (such asMemoryDataLocalityStage
?)What about
ExploitInterLayerDataLocalityStage
?
Good for me! Also rename the previous stage SearchUnsedMemoryStage
to SearchInterLayerDataLocalityStage
for consistency?
@asyms @JiacongSun As discussed, we would also like to rename these stages to something that describes better what it does. Are there any suggestions? What about
RemoveLayerTransferOverhead
?I feel
RemoveLayerTransferOverhead
orRemoveLayerTransferOverheadStage
is less intuitive compared to the originalRemoveUnusedMemoryStage
. I prefer to keep the original name or stress more on the point of data locality (such asMemoryDataLocalityStage
?)What about
ExploitInterLayerDataLocalityStage
?Good for me! Also rename the previous stage
SearchUnsedMemoryStage
toSearchInterLayerDataLocalityStage
for consistency?
I would prefer to merge the two stages, since they are completely dependent anyway. This would also solve the naming issue
@asyms @JiacongSun As discussed, we would also like to rename these stages to something that describes better what it does. Are there any suggestions? What about
RemoveLayerTransferOverhead
?I feel
RemoveLayerTransferOverhead
orRemoveLayerTransferOverheadStage
is less intuitive compared to the originalRemoveUnusedMemoryStage
. I prefer to keep the original name or stress more on the point of data locality (such asMemoryDataLocalityStage
?)What about
ExploitInterLayerDataLocalityStage
?Good for me! Also rename the previous stage
SearchUnsedMemoryStage
toSearchInterLayerDataLocalityStage
for consistency?I would prefer to merge the two stages, since they are completely dependent anyway. This would also solve the naming issue
OK! That would be great!
I thought the two stages required to be inserted at different points in the list of all stages? How do we plan on merging the two? I see two possibilities I would like your feedback on:
Fix
SearchUnusedMemoryStage
andRemoveUnusedMemoryStage