Hello @sandeepchawla74,
Sorry for the late answer.
We found that your issue is a bug due to a shape computed wrongly.
In your model, seemingly starting at the conv2d_12, the output that should be of shape (1x512x8x8) is seen by the st edge ai core as (1x512x18x18) which cause the concat below to try concatenating a shape (1x512x18x18 with 1x512x8x8 (right arrow)) which does not pass.
The issue seems triggered by the dilatation and the lack of an attribute "kernel_shape" not specified which is legal in ONNX.
By editing the model and adding the "kernel_shape" to every conv2d, it solves the issue.
It also save the model using ONNX ir version 9 as the version 10 is not supported by the core.
Here is the code:
import onnx
from onnx import helper, numpy_helper, checker
# Load your model
model = onnx.load("modified_model.onnx")
graph = model.graph
for node in graph.node:
if node.op_type == "Conv":
# Check if kernel_shape attribute exists
if not any(attr.name == "kernel_shape" for attr in node.attribute):
# Get the weight tensor (2nd input)
weight_name = node.input[1]
weight_tensor = next((init for init in graph.initializer if init.name == weight_name), None)
if weight_tensor is not None:
W = numpy_helper.to_array(weight_tensor)
# W shape = (out_channels, in_channels/groups, kH, kW)
kH, kW = W.shape[2], W.shape[3]
# Add kernel_shape attribute
kernel_shape_attr = helper.make_attribute("kernel_shape", [kH, kW])
node.attribute.append(kernel_shape_attr)
print(f"Added kernel_shape {[kH, kW]} to Conv node {node.name or '[unnamed]'}")
else:
print(f"Could not find weights for Conv node {node.name}, skipping.")
# Force IR version = 9
model.ir_version = 9
# (optional) Run checker to confirm validity
checker.check_model(model)
# Save fixed model
onnx.save(model, "model_fixed_ir9.onnx")
Have a good day,
Julian
