[onert] Introduce BiasInsertionPass #13842
Conversation
This commit introduces BiasInsertionPass to insert bias for training in case bias is optional input. ONE-DCO-1.0-Signed-off-by: ragmani <[email protected]>
| void run() final; | ||
|
|
||
| public: | ||
| void visit(const ir::operation::FullyConnected &node) override; |
There was a problem hiding this comment.
Should we have to insert bias into each node?
I'd like to review this PR, But I feel hard to understand the background 🥲
There was a problem hiding this comment.
Sometimes TFLite converter may generate a inference model that does not have bias input(i.e. optional input) when the bias data has all zero values. It's reasonable in inference because there is no reason to add zero values. But, in training, bias is necessary because bias needs to be trained and updated.
As a result, we need to insert bias input if a model to be trained does have bias input.
There was a problem hiding this comment.
@zetwhite
Thanks for your advise offline. I made a note about that at #13808 (comment).
ragmani
added
PR/NO MERGE
|
We concluded that we assumed onert only supports training for pre-trained models. So this PR is not necessary anymore. |
This commit introduces BiasInsertionPass to insert bias for training in case bias is optional input.
ONE-DCO-1.0-Signed-off-by: ragmani [email protected]