Kernel crashes with forward_dims method - Which argument in the set_compute_output_dims method ?

I am working on aidge with a YOLOv2 model. There is one Generic Operator : Reshape

My goal here is to implement this Reshape operator by following the tutorial about implementations.

Here is my code : aidge_yolov2.py.

This is the output log : log

Sorry, the onnx model file is too heavy, I can't share it :/

[UPDATED] : see the comments below

changed the description

My bad, I updated the code while there are multiple Reshape GenericOperators :

for node in graph.get_nodes(): 
   print(node.type())
   if node.type() == "Reshape":
        reshape_op =  node.get_operator()
        reshape_op.set_compute_output_dims("""????""") 
        reshape_op.set_impl(ReshapeImpl(reshape_op))

The thing is, I don't know what to write in argument in the set_compute_output_dims method for a reshape op.

At the : graph.forward_dims() command, the kernel crashes

Thanks

By the way, this is the mermaid file :yolov2_mermaid.mmd

The new code :aidge_yolov2.ipynb

changed the description

changed title from Error while implementing Reshape op : `false && "GenericOperator cannot forward dims"' to Kernel crashes with forward_dims method - Which argument in the set_compute_output_dims method ?

Indeed, the output dims computation is missing, but since the output dims are actually computed in the forward(), there is a problem! In the case of this operator, the static computeOutputDims() mechanim won't work.

We are currently working on a small refactoring to specifically address this issue...

Related to #95 (closed)

You can check branch forwarddims, which should solve your issue:

• graph.forward_dims() becomes optional, and it won't fail if no compute output dims is provided for Generic operator.
  • In this case, you should be aware that token-based scheduling will be used from this point instead of consumed-produced data quantities. If your are only using operators with full tensor consumption-production implementation (which is currently the case in Aidge), it does not change anything for you.
• You can therefore directly set the output tensor dims in your implementation!

mentioned in merge request !104 (merged)

assigned to @olivierbichler

added Bug 🐛 Help Wanted labels

changed milestone to %aidge - April 2024

changed health status to on track

Update Reshape support:

• implementation and core: !111 (merged)
• onnx import: aidge_onnx!38 (merged)

To answer the original question, you should provide a lambda function, like lambda x: x for the identity.

With the new native support of Reshape, do you still have an issue @diegob ?

Hi @pineapple ,

On which branch do I have to be for this update ?

Hi @diegob, you should be on the dev branch of aidge_core and aidge_onnx for now. It will be merged with main in about a week

Alright, after testing my code it seems to be working. The replace of the Reshape works.

However, I have another error but on the export. I will open a new issue in the aidge_onnx section. Thanks !

closed