[RFC][DTypes] pooling and LeakyReLU #2280
Conversation
| for (unsigned i = 0; i < in_type->size(); ++i) { | ||
| CHECK(in_type->at(i) == mshadow::default_type_flag || | ||
| in_type->at(i) == -1) << "Unsupported data type " << in_type->at(i); | ||
| int dtype = -1; |
There was a problem hiding this comment.
The default infertype forces float32 because most ops don't support other types. If your op does, override this in that op
|
The error I am seeing seem to be coming from mixed type multiplication in https://github.com/dmlc/mxnet/blob/5e000e24082201918023ea243622ec0ee051d15f/src/operator/pooling-inl.h#L95-L98 and https://github.com/dmlc/mxnet/blob/5e000e24082201918023ea243622ec0ee051d15f/src/operator/pooling-inl.h#L144-L145 |
|
@piiswrong @tqchen 0ed4195 has a workaround for the problem I described in dmlc/mshadow#125 |
vchuravy
force-pushed
the
vc/fp16
branch
2 times, most recently
from
e1ecdda
to
b4f903d
Compare
vchuravy
changed the title
|
@piiswrong Could you take another look at this? |
vchuravy
force-pushed
the
vc/fp16
branch
2 times, most recently
from
7119766
to
24b9325
Compare
| CHECK_EQ(in_type->size(), 1); | ||
| int dtype = in_type->at(0); | ||
|
|
||
| if (dtype == 1) { |
There was a problem hiding this comment.
why not double? Also consider use mshadow::kDouble instead of number
|
@piiswrong Thank you for the comments, I rebased and addressed them. |
|
please fix the compile error |
|
done. |
vchuravy
changed the title
|
LeakyReLU has the same operator dispatch problem that Pooling has |
vchuravy
force-pushed
the
vc/fp16
branch
3 times, most recently
from
0f0d5d2
to
1e76151
Compare
|
Travis run through earlier except for a lint problem that I consecutively fixed (https://travis-ci.org/dmlc/mxnet/builds/136388742) and had a seqfault on travis on the second run (https://travis-ci.org/dmlc/mxnet/builds/136410568). I have seen that seqfault before, but it is always transient |
@tqchen Is this the right approach? I seem to be missing something since I am getting