VectorBlox VNNX Conversion Issues: Removing Clip and ScatterND Operators

When converting AI models to VectorBlox’s .vnnx format, some operators¹ are not supported in VNNX, causing conversion errors. In such cases, these operators must be replaced with VNNX-supported operators.

1. Clip

Problem

The Clip operator generated during PyTorch to ONNX conversion is incompatible with VNNX.

Solution

• The Clip operator clips each element of the input tensor between a specified minimum (min) and maximum (max) value. The formula is:

• When converting from PyTorch to ONNX, find the part that creates the Clip operator, i.e., find torch.clamp in PyTorch. If found, replace it by combining min and max operations instead of torch.clamp in PyTorch.

Code Before/After Comparison

# Before (problem occurs)
bbox[:,:,0] = torch.clamp(x - w/2, 0, image_size[0] - 1)

# After
bbox[:,:,0] = torch.minimum(torch.maximum(x -  w/2, zero), x_max)

2. ScatterND

Problem

The ScatterND operator generated during ONNX conversion is incompatible with VNNX.

Solution

• ScatterND copies a given data tensor and overwrites values at positions specified by indices with updates to create a new tensor.
• However, ONNX does not have a general assignment operator like x[:,0] = 1 to modify existing values.
• Instead, it must create a new tensor by copying the original x and changing only the 0th column to 1. This is why the ScatterND operator appears.
• Instead of modifying specific positions, use torch.stack to stack and store operation results.
• The torch.stack operation is later converted to Concat or Unsqueeze + Concat operations.

Code Before/After Comparison

# Before (problem occurs)
bbox[:,:,0] = ...
bbox[:,:,1] = ...

# After
x1 = operation result of bbox[:,:,0]
y1 = operation result of bbox[:,:,1]
...
bbox = torch.stack((x1,y1,x2,y2), dim=-1)

Language: 한국어 (Korean)