[error] Failed to produce compiled graph

General
1

I’m using the Hailo tutorial to convert my custom model.
I’ve successfully completed the steps from ONNX → HAR → Quantization HAR, but I’m getting the following error during the hef = runner.compile() step:

It seems that the input sizes for ew_mult1 are mismatched, which might be causing the issue.

How can I fix this?

[error] Failed to produce compiled graph
BackendAllocatorException: Compilation failed: The graph is invalid. The input shapes of node ew_mult1 are not aligned (1, 16384, 24) != (128, 128, 24)

image
image1139×887 107 KB

2

Is there any guide for this issue?

3

Hey @sslee1 ,

The issue arises from a tensor shape mismatch during an ElementWise operation—specifically at the node ew_mult1 during compilation.

From your model graph, I can see that ew_mult1 receives two inputs:

  • One input has the shape [-1, 1, 16384, 24]
  • The other, coming from a resize operation, has shape [-1, 1, 1, 24]

This smaller tensor is likely being broadcasted implicitly in your original framework (TensorFlow or PyTorch), which is supported by those runtimes. However:

The DFC does not support automatic broadcasting in all scenarios, particularly when spatial dimensions differ. It requires explicit shape alignment between tensors in elementwise operations.

To resolve this, you should explicitly reshape or tile the [1, 1, 24] tensor to match [1, 16384, 24] before the ew_mult1 operation.

Here’s how to fix it in your model before exporting to ONNX:

For TensorFlow:

# If tensor shape is [B, 1, 1, 24]
x = tf.tile(x, [1, 1, 16384, 1])  # Shape becomes [B, 1, 16384, 24]

For PyTorch:

# If tensor shape is [B, 1, 1, 24]
x = x.expand(-1, 1, 16384, -1)    # Or use repeat: x = x.repeat(1, 1, 16384, 1)

Make sure this reshape aligns with your model’s logic (e.g. applying a per-channel scale across spatial positions).

You should re exporting the ONNX and re Parsing you’re model.
This should allow the compiler to proceed without triggering shape mismatch errors.