Onnxruntime: when there are few cast connected, the result might be wrong

[pengwa]

Describe the bug

look at the graph below. The input is bool[1, 100], we for example feed it with all True.

Then Cast (to 1) change it to all 1.0.

Then ReduceSum(axes=1) generate results [100.0]

Then Cast(to 9) change it to [True]

Then Cast(to 1) change it to [100.0]

So the last step should not expected, because user would expect: when they want to cast a True to float, the value should be 1.0.

I guess this is caused when Cast is computed, the value is not changed actually, maybe just a cast.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 10
• ONNX Runtime installed from (source or binary): master
• ONNX Runtime version: master
• Python version: 3.6
• GCC/Compiler version (if compiling from source): VS 2017
• CUDA/cuDNN version: N/A
• GPU model and memory: N/A

To Reproduce
Describe steps/code to reproduce the behavior: see above

Expected behavior

whenever cast a bool True to float, the value should be 1.0

Screenshots
If applicable, add screenshots to help explain your problem.

Additional context
Add any other context about the problem here.

[yufenglee]

@raymondxyang could you fix this issue with your cast update?

[raymondxyang]

Just verified the eigen (used in cast op kernel) will not produce the described result, when casting float->bool->float. in other word, 100.0 will be cast into 1.0 in this case. Could u send me the model for more debugging processes? @pengwa Thanks! 😃

[raymondxyang]

followed up offline. can re-produce the error. looking into it..