IndexError with Gemm and Tanh

Oops there’s a typo in the title. Originally the error occurs with Gemm+Tanh, and I still get the error with Gemm+ReLU, as seen in the attached model.

Yo ​@yan12125 ! Yeah, as you say, this sounds really similar to the issue ​@MArio was having.

I just took a quick glance at the log, and before we dig any deeper, I saw something that might help us out here - the compiler expects 4D tensors, but I think the model is feeding it a 2D tensor directly into the Gemm. As I recall, this was a similar problem in the other post.

Reshaping the input into 4D ((1, 10, 1, 1) instead of (1, 10), perhaps?) could be worth a shot 👍

Let me know!

Thanks. I tried some ways to use 4D input, but neither works

• Use (1, 10, 1, 1) as input and put Reshape layer before Gemm to turn 4D data back to 2D for Gemm. The error looks the same as the original model - IndexError in axelera.compiler.frontend.passes.pass_legalize_flatten_linear. The tested model is model-with-reshape.onnx
• Use (1, 1, 1, 10) and use 4D pytorch linear layer. The model is generated from the following script, and Axelera compiler results into another error - AssertionError in axelera.compiler.frontend.passes.pass_rewrite_dense_to_conv2d. Detailed logs and the tested model (model-4d-linear) are attached below.

from torch import nn import torch.onnx

class Model(nn.Module):     def __init__(self, input_dim, hidden_size, output_dim):         super().__init__()         self.layers = nn.Sequential(             nn.Linear(input_dim, hidden_size, bias=False),             nn.ReLU(),         )

    def forward(self, x):         x = x.view(1, 10)         return self.layers(x)

model = Model(input_dim=10, hidden_size=10, output_dim=10) torch.onnx.export(model, (torch.rand(1, 1, 1, 10),), 'model.onnx')

Thanks! Indeed prepending Identity Conv & Flatten layers to the ONNX model allows compilation to work.

Excellent news, and good to know at this end too, so thanks for the update! That could be useful info if anyone else has this 👍