Open seanshpark opened 4 years ago
Related #964
While in GraphBuild
step, output tensor dtype
is set to the IR Node
.
We can simply return dtype
of the Node
in type inference.
Question; do we still need TypeInference?
Question; do we still need TypeInference?
IMHO we don't, but it could be transformed into TypeChecker.
There was another opinion, that TypeInference could be helpful after optimization passes which do not set dtype
properly. IMHO those optimizations should be fixed to set dtype
(and verify that fact with TypeChecker).
TypeInference could be helpful after optimization passes which do not set dtype properly
Ah.. thanks for reminding me :)
From https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/zeros_like,
Lets find a way to support this in luci
ZerosLikeOptions
don't have this attribute