Closed disktnk closed 5 years ago
[NOTE]
Chainer returns always int32 (source), but ONNX ArgMax(Min) constrains outputs as int64 (Operators.md). Other framework returns int64 on default (tf, pytorch), along with numpy
int32
int64
>>> np.argmax(np.arange(6).reshape(2,3).astype(np.float32)).dtype dtype('int64')
Current ONNX-Chainer does not provide explicit dtype assignment, cannot support them simply.
[NOTE]
Chainer returns always
int32(source), but ONNX ArgMax(Min) constrains outputs asint64(Operators.md). Other framework returnsint64on default (tf, pytorch), along with numpyCurrent ONNX-Chainer does not provide explicit dtype assignment, cannot support them simply.