For some Const node: "initializer shape is inconsistent"

[pengwa]

Traceback (most recent call last):
  File "/usr/lib/python3.5/runpy.py", line 184, in _run_module_as_main
    "__main__", mod_spec)
  File "/usr/lib/python3.5/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/home/pengwang/community/tensorflow-onnx/tf2onnx/convert.py", line 100, in <module>
    main()
  File "/home/pengwang/community/tensorflow-onnx/tf2onnx/convert.py", line 92, in main
    optimize=not args.continue_on_error)
  File "/home/pengwang/community/tensorflow-onnx/tf2onnx/graph.py", line 427, in make_model
    raise ValueError("initializer shape is inconsistent")
ValueError: initializer shape is inconsistent
python3 -m tf2onnx.convert --input /tmp/frozen/dcgan_2.pb --inputs y:0,z:1 --outputs generator/Sigmoid:0 --output dcgan_2.pb.onnx --verbose --continue_on_error
generated onnx is located/home/pengwang/dcgan_2.pb.onnx
------ switch back to original directory: /home/pengwang/community/learning/onnx

The problematic node looks as below:

node {
  name: "generator/g_bn2/Const_1"
  op: "Const"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_FLOAT
        tensor_shape {
          dim {
          }
        }
      }
    }
  }
}

https://github.com/onnx/tensorflow-onnx/blob/master/tf2onnx/graph.py#L421, in runtime, the values are:
list(shape) is [0]
initializer.dims is [1]

Since the above constant is a scalar, I think initializer.dims might get wrong?
initializer.dims is created in add_initializer, which is called by https://github.com/onnx/tensorflow-onnx/blob/master/tf2onnx/tfonnx.py#L206, where when i print(tensor), and print(type(tensor)), it shows

name: "value"
t {
  dims: 1
  data_type: FLOAT
  float_data: 0.0
  name: "generator/g_bn2/Const:1"
}
type: TENSOR

<class 'onnx_pb2.AttributeProto'>

Any thoughts? @guschmue

-------------------paste the good const node as compassion in below ------------------------------

The other hand, example const nodes that works well

node {
  name: "generator/g_bn2/AssignMovingAvg/decay"
  op: "Const"
  attr {
    key: "_class"
    value {
      list {
        s: "loc:@generator/g_bn2/moving_mean"
      }
    }
  }
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_FLOAT
        tensor_shape {
        }
        float_val: 0.10000000149011612
      }
    }
  }
}


node {
  name: "generator/g_h3/biases"
  op: "Const"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_FLOAT
        tensor_shape {
          dim {
            size: 1
          }
        }
        float_val: 0.0900091901421547
      }
    }
  }
}


node {
  name: "generator/ones_1/shape_as_tensor"
  op: "Const"
  attr {
    key: "dtype"
    value {
      type: DT_INT32
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_INT32
        tensor_shape {
          dim {
            size: 4
          }
        }
        tensor_content: "@\000\000\000\016\000\000\000\016\000\000\000\n\000\000\000"
      }
    }
  }
}

node {
  name: "generator/g_h3/w"
  op: "Const"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_FLOAT
        tensor_shape {
          dim {
            size: 5
          }
          dim {
            size: 5
          }
          dim {
            size: 1
          }
          dim {
            size: 138
          }
        }
		tensor_content: ""
      }
    }
  }
}

[guschmue]

I remember that there was some issue with caffe2 and winml that some tensor ops don't work if you use a scalar but they are perfectly happy if you pass [scalar] and I think we intentionally made a change to use [scalar] for that reason. Let me look at this again, it is some time ago and I don't know if caffe2 still needs this. We could maybe ifdef this for the caffe2 runtime ... this target thing was meant to allow runtime specific workarounds.

if ctx.is_target(TARGET_CAFFE2):

[pengwa]

Thanks @guschmue for the reply, while this issue should be in tf-graph->onnx conversion stage, not yet hit the caffe2 runtime.

[pengwa]

Do you mean let me comment the https://github.com/onnx/tensorflow-onnx/blob/master/tf2onnx/graph.py#L421 with check "if ctx.is_target(TARGET_CAFFE2):"?

[guschmue]

void my comment ,,, not the same. I'll look at it today.

[pengwa]

I just comment the L421 check to unblock the model conversion, so maybe we can consider to loose the shape==dim check for some special case. anyway, we can discuss once you get better context later. :)

[guschmue]

sure

[pengwa]

not having this issue recently, close this unless we got this failure again.