Add const-tensor checks #3141
Add const-tensor checks #3141
Conversation
@tensorflow/micro Add consistent const-tensor checks and error messages across these kernels: TRANSPOSE STRIDED_SLICE FILL BROADCAST_TO EXPAND_DIMS Modify associated kernel unit tests. bug=fixes tensorflow#3140
| if (dims_data != nullptr) { | ||
| // dims must be a const tensor | ||
| tensors[0].allocation_type = kTfLiteMmapRo; | ||
| } |
There was a problem hiding this comment.
Better to have this condition after all the tensor creations are done?
There was a problem hiding this comment.
The code only applies to the dims tensor, so it is colocated with the tensor creation.
| output_data); | ||
| tflite::micro::KernelRunner runner = | ||
| CreateFillTestRunner(dims_shape, dims_data, value_shape, value_data, | ||
| output_shape, output_data); |
There was a problem hiding this comment.
I just want to understand few things here.
- Is this test was successful previously?
- Do we want to make this test fail in the prepare stage?
- Why can't we use TestFill() API here aswell?
There was a problem hiding this comment.
The old kernel allowed the dims tensor to have changing values. This is incorrect and was corrected in the kernel such that the dims tensor must be a const-tensor.
As a result, we now want a negative test, and it should fail during the prepare phase.
We dont use TestFill because it doesn't support negative tests. We rarely rewrite old unit test code as it is not worth the effort (it would be better to plan an entirely new unit test code base).
@tensorflow/micro
Add consistent const-tensor checks and error messages across these kernels: TRANSPOSE
STRIDED_SLICE
FILL
BROADCAST_TO
EXPAND_DIMS
Modify associated kernel unit tests.
bug=fixes #3140