Custom model optimizing is failing

[d.gentner]

Hey everyone,

I’ve been working on converting a custom dataset-trained PyTorch model (SMP DeepLabV3+) to run on Hailo using the Dataflow Compiler, and I’ve hit a roadblock. Hoping someone here might have some advice!

Here’s what I’ve done so far:

• Converted the model to ONNX format.
• Followed the tutorial Jupyter notebooks:
  • Created the calibration .npy file.
  • Generated the .alls file with the correct normalization settings.

But when I got to the optimization step, I hit this error:
ValueError: Inputs and input nodes not the same length in layer combustion_model/concat1 - inputs: 4, nodes: 5

I reviewed the .hn file using Netron and noticed that the resize and concat38 layers both feed into the same node

I figured it was an issue with how the resize layer interacts with the 4 convolution layers. So, I manually added another concat layer to split things up.

But… the error shifted to the new layer:
ValueError: Inputs and input nodes not the same length in layer combustion_model/concat11 - inputs: 3, nodes: 4

After that, I thought maybe the problem was my calibration set. I suspected it might be because some images didn’t have all 4 segmentation classes present, so I created a subset where every image has all 4 classes.

Error still persists.

Does anyone have experience with this type of issue or know what might be going wrong? Any ideas would be super appreciated!

[d.gentner]

this is the screenshot of the added concat.

[d.gentner]

Hello together,

does anyone have an idea? Maybe the admins?
Thanks!

[omria]

Hey @d.gentner ,

Welcome to the Hailo Community!

Addressing Concatenation Layer Issues in Model Optimization

Let’s break down the problem and solution:

Understanding the Issue

The error ValueError: Inputs and input nodes not the same length in layer ... indicates a structural misalignment in your model’s concatenation layers. This typically occurs when:

• The ONNX graph has mismatched inputs for Concat operations
• Intermediate layer outputs aren’t properly connected to the concatenation nodes

Recommended Solution

1. Model Analysis

First, examine your model structure using Netron, Focus on the concat1 and concat11 layers to verify their input connections.

2. Model Architecture Correction

If you find input mismatches, update your PyTorch model:

# Example of proper concatenation layer setup
class YourModel(nn.Module):
    def forward(self, x):
        # Ensure all branch outputs are properly defined
        branch1 = self.conv1(x)
        branch2 = self.conv2(x)
        branch3 = self.resize(x)
        
        # Verify all inputs are included
        concat_output = torch.cat([branch1, branch2, branch3], dim=1)
        return concat_output

3. ONNX Export Configuration

Use these settings for reliable ONNX export:

torch.onnx.export(
    model, 
    dummy_input, 
    "model.onnx",
    opset_version=11,
    input_names=['input'],
    output_names=['output'],
    dynamic_axes={'input': {0: 'batch_size'},
                 'output': {0: 'batch_size'}}
)

Feel free to share your model or specific error messages if you need more detailed guidance!

[d.gentner]

Hi Omria,

Thank you very mcuh for your answer!

I do not belive that the ONNX file has an error because I can Infere the model and the netron visualization does not show any issues like the .har file.

[d.gentner]