Zha0q1 opened a new issue #19753:
URL: https://github.com/apache/incubator-mxnet/issues/19753
It looks like regardless of the input type the indices output `B[1]` is
always of type float32. I checked the c++ code and I think it's supposed to be
int32/64? Is this behavior by design?
``` python
>>> A = np.array([1,2,3], dtype='float16')
>>> B = npx.topk(A, k=3, ret_typ='both')
>>> B[0].dtype
dtype('float16')
>>> B[1].dtype
dtype('float32')
```
```c++
inline bool TopKType(const nnvm::NodeAttrs& attrs,
std::vector<int> *in_attrs,
std::vector<int> *out_attrs) {
const TopKParam& param = nnvm::get<TopKParam>(attrs.parsed);
size_t in_size = in_attrs->size();
size_t out_size = out_attrs->size();
CHECK_EQ(in_size, 1);
CHECK(out_size == 1 || out_size == 2);
// out_attr[0] -> stores value
// out_attr[1] -> stores indices
if (out_size > 1) {
if (param.ret_typ == topk_enum::kReturnValue) {
#if MXNET_USE_INT64_TENSOR_SIZE == 1
CHECK(type_assign(&(*out_attrs)[1], mshadow::kInt64))
#else
CHECK(type_assign(&(*out_attrs)[1], mshadow::kInt32))
#endif
<< "Failed to set the type of ret_indices.";
} else {
CHECK(type_assign(&(*out_attrs)[1], param.dtype))
<< "Failed to set the type of ret_indices.";
}
}
if (param.ret_typ == topk_enum::kReturnIndices) {
CHECK(type_assign(&(*out_attrs)[0], param.dtype))
<< "Failed to set the type of ret_indices.";
} else {
TYPE_ASSIGN_CHECK(*out_attrs, 0, in_attrs->at(0));
TYPE_ASSIGN_CHECK(*in_attrs, 0, out_attrs->at(0));
return out_attrs->at(0) != -1;
}
return true;
}
```
@access2rohit @leezu @sxjscience @szha
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
