Graph manipulation issue: Expand

What commit version of aidge do you use

• aidge_core: branch dev / commit 22f04038

Problem description

After importing and quantizing my model, I create custom metaops. When expanding existing metaops before wreating new ones, it seems the graph is getting corrupted. The mermaid file saved before expand_metaops function shows connexion of input0 to node PaddedConv2D_0 (1st node of the model).

input0((in#0)):::inputCls--->|" [1, 3, 384, 512] Float32<br/>&darr;<br/>0"|PaddedConv2D_0

The mermaid file saved after expand_metaops function shows connexion of input0 to node Resize_2.

input0((in#0)):::inputCls-.->|"&darr;<br/>1"|Resize_2

Reproducible example

pneuro_export.py.txt model.onnx

[ERROR] - Missing mandatory input#0 for node [_0_MetaConv_0 - (MetaConv)]
[FATAL] - Pad2D: input #0 should be associated with a Tensor
Traceback (most recent call last):
  File "/local2/is157416/rm254443/git_repo/workspace/seg_mobv1_3x384x512_alph_1/pneuro_export.py", line 200, in <module>
    scheduler.forward(True, [input_tensor])
RuntimeError: Pad2D: input #0 should be associated with a Tensor

changed the description

After a call with @raphaelmillet, adding model.set_ordered_inputs([(model.get_node("_0_MetaConv_0"), 0)]) fFixed the forward issue.

It looks like the forward function does not set the input tensor to the right inputs, despite the input being avialable (see error message: RuntimeError: Pad2D: input #0 should be associated with a Tensor.

Further investigation need to be done to understand why the graph is corrupted.

added Bug 🐛 label

The issue here is that GraphView::replace() does not preserve graph inputs/outputs order, because currently the node is first removed, then a new one is inserted.

This has caused issues in the past and should be fixed in the future, altough the fix is not trivial.