WeGO-PyTorch Limitations - 3.5 English

Vitis AI User Guide (UG1414)

Document ID
UG1414
Release Date
2023-09-28
Version
3.5 English

The WeGO-Torch project is currently in an early access state and might encounter a few known usage issues. Here are the details of these issues and the necessary steps to address them:

  1. WeGO-Torch cannot support RCNN models (with control-flow) because:
    1. RCNN Model Support: WeGO-Torch does not currently support RCNN models with control-flow due to dynamic shape problems. In RCNN models, the shape of tensors can change during runtime when different images are provided as inputs. This poses challenges for deployment in WeGO. To make RCNN models compatible with WeGO, manual modifications are required to eliminate this constraint.
    2. Input Type Compatibility: RCNN models often accept Tensor[] as an input type, which WeGO's compile API does not support. Additionally, using Tensor [] as the input type implies that the float model itself is batch-sensitive, and the quantized models obtained through tracing differ based on the batch size used during the TorchScript tracing phase. To deploy these models in WeGO, the following steps are recommended:
      1. Replace Tensor [] with Tensor or Tensor, Tensor, ... (when the number of inputs is known) as the input type in the original float model.
      2. The batch size used for inference in WeGO must be the same as the one used in the export phase during quantization.
  2. WeGO-Torch currently covers only a subset of operators that data center DPUs can support. Consequently, certain operators might be dispatched to the CPU for execution, even if data center DPUs can support them.