NOT IMPLEMENTED: Gemm with channel first A input is not supported

[Nephalem]

Some errors occurred during the model conversion process. I have provided my model file below. 

Cube-AI report ：NOT IMPLEMENTED: Gemm with channel first A input is not supported

It seems to be caused by an inability to handle the transposed data, but how do I resolve this issue? Or is the transpose operator unusable?

[Julian E.]

Hello @Nephalem,

It seems there is also an issue with the Gemm layers in your model.

If we take a look at the first one:

With:

• A is (1, 144)
• B is (128, 144)
• C is (128) (broadcasted)
• Y is (1, 128)

And:

• transA=0 → A is not transposed (remains 1x144).
• transB=1 → B is transposed before multiplication (B changes from 128x144 to 144x128).
• alpha=1 → No scaling on the A * B^T product.
• beta=1 → The bias (C) is added as is.

So, your operation effectively becomes: 𝑌=(𝐴×𝐵𝑇)+𝐶

You can try to replace the Gemm layers by with basic layers such as transpose, matmul and add.
Something like this:

import torch

# Define the inputs
A = torch.randn(1, 144)    # Shape (1, 144)
B = torch.randn(128, 144)  # Shape (128, 144)
C = torch.randn(128)       # Shape (128,)

# Perform equivalent Gemm operation using matmul, transpose, and add
Y = torch.matmul(A, B.T) + C  # B.T gives shape (144, 128), matmul gives (1, 128)

# Print result shape to verify
print(Y.shape)  # Should be (1, 128)

Please double check with the Gemm pytorch documentation that this is correct.

Have a good day,

Julian

[Nephalem]

Hello @Julian E. 

Following your instructions, I solved that problem, but now I've encountered a new issue。

TOOL ERROR: list index out of range

I have a few other questions. For example, how can I make CUBE-AI print longer debug logs? It's really hard to tell exactly which array is causing the out-of-bounds error from just this single line of error message. Thank you, and I hope you have a great day

[Julian E. · Accepted solution]

Hello @Nephalem ,

Yes indeed, it is hard to tell.

In the X Cube AI menu, you can click setting and set --verbosity 3:

But I don't think it is enough.

What you should do is use stedgeai.exe manually.

When you do a validate in X Cube AI, it literally does:

stedgeai.exe generate --model your_model --target stm32n6 --st-neural-art

So you can open a bash terminal and do this manually but you also need to first do:

export _DEBUG=2

to have more logs.

Have a good day,

Julian

[Nephalem]

Hello @Julian E. 

I tried it, but didn’t get any output in the folder. Where should I add the export _DEBUG=2 parameter? My device uses the STM32F4 series, and the interface is slightly different from yours, though not significantly. If possible, could you help me analyze where the array out-of-bounds error mentioned in the model occurs? I would greatly appreciate it. Alternatively, could you share the CUBE-AI debug report for me to review?

 

Thank you

hope you have a nice day!