Runtime error with an empty input tensor in 1.17 not in 1.15 about onnxruntime HOT 8 OPEN

You are dealing with a model that has dynamic shapes. The error above usually occurs, when the dynamic shape does not propagate properly downstream and there is a mismatch in some part. For example, it might have an initializer that requires only a fixed shape, so the model would work with one shape, but not another.

from onnxruntime.

Is there a workaround we can implement to fix the issue for the given model in release 1.17.3?

from onnxruntime.

The issue is likely with the model, but it might be a converter issue as well. What is the source of your model? Is it a pytorch converted model or it is created in house?

from onnxruntime.

If you sess_opts.enable_mem_reuse = False then the model runs, but does not produce expected shapes. This is again because the dynamic shapes are not propagated properly.

2024-04-26 11:51:37.4997890 [W:onnxruntime:, execution_frame.cc:879 onnxruntime::ExecutionFrame::VerifyOutputSizes] Expected shape from model of {-1} does not match actual shape of {0,8} for output /Squeeze_4_output_0
2024-04-26 11:51:37.5124220 [W:onnxruntime:, execution_frame.cc:879 onnxruntime::ExecutionFrame::VerifyOutputSizes] Expected shape from model of {-1} does not match actual shape of {0,8} for output GN2v01_TrackOrigin

from onnxruntime.

Let me dig some more

from onnxruntime.

The issue is likely with the model, but it might be a converter issue as well. What is the source of your model? Is it a pytorch converted model or it is created in house?

It is converted from Pytorch.

from onnxruntime.

The following node has a hardcoded output shape to 'n_tracks'. When you run the model with the given inputs Squeeze gets a shape of {1, 0, 8} and outputs {0,8} which cannot match a 1-D shape.

Similar story with another node. Dynamic shapes are tricky. However, that node certainly cannot hardcode 1-D shape for its output.

Alternatively, you may need to look upstream, perhaps, that 0 would have been Squeezed out if it was 1.

node {
    input: "/Equal_output_0"
    output: "/If_output_0"
    name: "/If"
    op_type: "If"
    attribute {
      name: "then_branch"
      g {
        node {
          input: "/ArgMax_output_0"
          input: "/Constant_5_output_0"
          output: "/Squeeze_4_output_0"
          name: "/Squeeze_4"
          op_type: "Squeeze"
        }
        name: "torch_jit1"
        initializer {
          dims: 1
          data_type: 7
          name: "/Constant_5_output_0"
          raw_data: "\000\000\000\000\000\000\000\000"
        }
        doc_string: ""
        output {
          name: "/Squeeze_4_output_0"
          type {
            tensor_type {
              elem_type: 7
              shape {
                dim {
                  dim_param: "n_tracks"
                }
              }
            }
          }
        }
      }
      type: GRAPH
    }
    attribute {
      name: "else_branch"
      g {
        node {
          input: "/ArgMax_output_0"
          output: "/Identity_output_0"
          name: "/Identity"
          op_type: "Identity"
        }
        name: "torch_jit2"
        doc_string: ""
        output {
          name: "/Identity_output_0"
          type {
            tensor_type {
              elem_type: 7
              shape {
                dim {
                  dim_param: "Identity/Identity_output_0_dim_0"
                }
                dim {
                  dim_param: "n_tracks"
                }
              }
            }
          }
        }
      }
      type: GRAPH
    }
  }

from onnxruntime.

Thanks for looking into this in depth. We will try to improve the model architecture next time to avoid hardcoded dimensions.

I tried with the option sess_opts.enable_mem_reuse = False and observed the same error. Do you know why it runs well in release 1.15 but not in 1.17?

from onnxruntime.