X-CUBE-AI failed to convert my model and reported the following error.

[Nephalem]

I have provided my ONNX model, which has been tested.

INTERNAL ERROR: H not found in shape with shape map (BATCH, CH, W)

[Julian E.]

Hello @Nephalem ,

There is probably something wrong happening during the conversion of the model because of the original shape of your input.

I'll take a look an update you once I know more.

Have a good day,

Julian

[Julian E. · Accepted solution]

Hello @Nephalem,

Your issue here comes from Einsum layer.

In your case, the operation that does not work is einsum("bhqk,bkhd->bqhd")

It seems that we don't fully support Einsum currently, so you need to replace the einsum layers (you have multiple ones in your model by simple matrixes operation instead). for example:

Equivalent Operations:
Instead of einsum("bhqk,bkhd->bqhd", A, B), use:

import torch

# Example tensors
A = torch.randn(batch, heads, query, key) # (b, h, q, k)
B = torch.randn(batch, key, heads, dim) # (b, k, h, d)

# Transpose B to (b, h, k, d) so that k aligns for matmul
B_transposed = B.permute(0, 2, 1, 3) # (b, h, k, d)

# Perform batched matrix multiplication
result = torch.matmul(A, B_transposed) # (b, h, q, d)

# Swap axes to match expected output shape (b, q, h, d)
result = result.permute(0, 2, 1, 3) # (b, q, h, d)

Have a good day,

Julian