Error using tf.boolean_mask

[EnricoBos]

Ciao, I am working on my own version of ("Tiny") YOLO. These days I'm writing the object detection part but I have a problem using the tf.boolean_mask function. To figure out if there is a problem with my code (Tensor dimensions or other), I looked at the example from the official TensorFlow page expecting to get the same result.

Here is my code: float[,] jj = new float[3, 2]; jj[0, 0] = 1; jj[0, 1] = 2; jj[1, 0] = 3; jj[1, 1] = 4; jj[2, 0] = 5; jj[2, 1] = 6; Tensor tensor =new Tensor(jj);

bool[] jjj = new bool[3]; jjj[0] = true; jjj[1] = false; jjj[2] = true; Tensor mask = new Tensor(jjj);

Tensor Bool_mask =(tf.boolean_mask(tensor, mask));

Maybe I am using the function incorrectly, but I don't understand where the problem lies. Any suggestions? thank you Enrico

[Oceania2018]

Can you help PR an unit test for this case?

[EnricoBos]

Ciao Oceania, yes sure.

[teknikal-wizard]

I've just faced an issue with boolean mask in my project too. As a sanity check I tried using the existing repo test code and I see the same.

It is slightly different to @EnricoBos error.

Using

let tensor = [| 0; 1; 2; 3 |]
let mask = np.array([| true; false; true; false  |])
let masked = tf.boolean_mask(tensor, mask);

I get

Tensorflow.InvalidArgumentError: 'ConcatOp : Ranks of all input tensors should match: shape[0] = [0] vs. shape[1] = []'

[bemobolo]

I have experienced another error (TF.NET version 0.70.1) with boolean_mask:

tf.boolean_mask(new int[] { 0, 1, 2, 3 }, new bool[] { true, false, true, false })

Throws:

Tensorflow.InvalidArgumentError
  HResult=0x80131500
  Message=Shape must be rank 1 but is rank 0 for '{{node All/boolean_mask/concat}} = ConcatV2[N=3, T=DT_INT32, 
Tidx=DT_INT32](All/boolean_mask/strided_slice_1, All/boolean_mask/Prod, All/boolean_mask/strided_slice_2, 
All/boolean_mask/concat/axis)' with input shapes: [0], [], [0], [].
  Source=Tensorflow.Binding
  StackTrace:
   at Tensorflow.ops._create_c_op(Graph graph, NodeDef node_def, Tensor[] inputs, Operation[] control_inputs, OpDef op_def)
   at Tensorflow.Operation..ctor(NodeDef node_def, Graph g, Tensor[] inputs, TF_DataType[] output_types, ITensorOrOperation[] control_inputs, TF_DataType[] input_types, String original_op, OpDef op_def)
   at Tensorflow.Graph.create_op(String op_type, Tensor[] inputs, TF_DataType[] dtypes, TF_DataType[] input_types, String name, Dictionary`2 attrs, OpDef op_def, Boolean compute_device)
   at Tensorflow.OpDefLibrary._apply_op_helper(String op_type_name, String name, Dictionary`2 keywords)
   at Tensorflow.Contexts.Context.ExecGraphAction(String OpType, String Name, ExecuteOpArgs args)
   at Tensorflow.Contexts.Context.ExecuteOp(String opType, String name, ExecuteOpArgs args)
   at Tensorflow.gen_array_ops.concat_v2(Tensor[] values, Int32 axis, String name)
   at Tensorflow.array_ops.concat(Tensor[] values, Int32 axis, String name)
   at Tensorflow.array_ops.<>c__DisplayClass4_0`2.<boolean_mask>b__0(NameScope <p0>)
   at Tensorflow.Binding.tf_with[TIn,TOut](TIn py, Func`2 action)
   at Tensorflow.array_ops.boolean_mask[T1,T2](T1 tensor, T2 mask, String name, Int32 axis)
   at Tensorflow.tensorflow.boolean_mask[T1,T2](T1 tensor, T2 mask, String name, Int32 axis)

[Pigrecos]

different solutions, one could be this(in boolean_mask function): var _leading_size = gen_math_ops.prod(shape(tensor_tensor)[$"{axis}:{axis + ndims_mask}"], new[] { 0 }) ; var leading_size = array_ops._autopacking_conversion_function(new[] { _leading_size }, _leading_size.dtype, "");