Boolean as a model input

[salehisaeed]

Hi all,

I have a basic question regarding loading a trained model using CppFlow.

So, my multi-input model should be called through:
auto output = model ( { {"serving_default_args_0:0", states}, {"serving_default_deterministic:0", deterministic} }, {"StatefulPartitionedCall:0"} );
The states is my float inputs that can be created through cppflow::fill({1, 99}, 1.0f);. But the deterministic input is a boolean. I am not sure how to produce it. I tried many possibilities without success.

Any idea or help would be much appreciated.
Saeed

[serizba]

Hi @salehisaeed

Perhaps the best way to get a bool tensor is to use cast:

unsigned char mytensor = 0;
auto c = cppflow::tensor(mytensor);
cppflow::cast(mytensor, TF_UINT8, TF_BOOL);

[serizba]

Exactly! Glad that you fixed it!

[salehisaeed]

Hi @serizba

Thanks for your reply. The c tensor is not used in your proposed code:

unsigned char mytensor = 0;
auto c = cppflow::tensor(mytensor);
cppflow::cast(mytensor, TF_UINT8, TF_BOOL);

Did you mean?

 cppflow::cast(c, TF_UINT8, TF_BOOL);

Nevertheless, it still does not work for me. The c tensor still has a type of dtype=TF_UINT8 after casting and thus my model fails.

[salehisaeed]

I tried @tdakhran suggestion and added a specialized constructor to cppflow::tensor and it works OK.

Thank you all for the help.

[serizba]

Probably the @tdakhran solution is better.

What I meant was:

unsigned char c = 0;
auto mytensor = cppflow::tensor(c);
mytensor = cppflow::cast(mytensor, TF_UINT8, TF_BOOL);

Perhaps you didn't assign the casted tensor. cppflow::cast returns a tensor, so you need to assign it.

[salehisaeed]

Correct! Now, this solution also works.

Thanks!

[tdakhran]

@serizba, does it make sense to upstream the support of bool tensors with the overhead of extra copy? I can prepare PR.