Problems with TF_GraphGetTensorNumDims when applied to "Variable" operations

[juraj-bicikl]

For some reason, when applying TF_GraphGetTensorNumDims to a "Variable" operation,
the functions returns -1, even though the shape and number of dimensions are
well defined.

A simple example:

#include <string>
#include <iostream>
#include "tensorflow/c/c_api.h"

int main() {

  // creating a TF_Graph
  TF_Graph* graph = TF_NewGraph();

  // creating a "Variable" operation
  TF_OperationDescription* opDesc = TF_NewOperation(graph, "Variable", "w");
  const long long dims[2] = {2, 2};
  TF_SetAttrShape(opDesc, "shape", dims, 2);
  TF_SetAttrType(opDesc, "dtype", TF_DOUBLE);

  TF_Status* status = TF_NewStatus();
  TF_Operation* w = TF_FinishOperation(opDesc, status);
  std::string finish_message = std::string(TF_Message(status));

  TF_Port w_port = {w, 0};
  int w_num_dims = TF_GraphGetTensorNumDims(graph, w_port, status);
  std::string num_dims_message = std::string(TF_Message(status));

  std::cout << "finish_message: " << finish_message << '\n';
  std::cout << "w_num_dims: " << w_num_dims << '\n';
  std::cout << "num_dims_message: " << num_dims_message << '\n';

  TF_DeleteGraph(graph);
  TF_DeleteStatus(status);

  return 0;
}

The program returns:

finish_message:
w_num_dims: -1
num_dims_message:

However, if I do the same for a "Placeholder" operation (by simply replacing
"Variable" with "Placeholder), the returned number of dimensions is 2 (as expected).

It seems to me like a bug. It might be related to the previous issue that
I reported (#5059), since the inability to determine the shape of a tensor at a particular node propagates through the graph.

[asimshankar]

Thanks for the detailed report @juraj-bicikl . This particular case is because the C++ function for shape inference for the Variable op is explicitly returning an unknown shape (state_ops.cc). That does seem fishy to me, I'll investigate further and get back to you.

[asimshankar]

Unfortunately, for historical reasons, the "Variable" op's shape function will not be able to distinguish between an unknown shape or a scalar shape (similar to the Placeholder op). Before the 1.0 release, we are going to try and fix this up. However, in the mean time:

• I'm going to try and fix this up for non-scalar shapes
• You might consider the workaround for Variable ops where you explicitly set the shape using something like:

TF_GraphSetTensorShape(graph, w_port, dims, 2, status);

[juraj-bicikl]

Thanks a lot!
I just tried using TF_GraphSetTensorShape, and it seems like a nice workaround.