You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@mxnet.apache.org by GitBox <gi...@apache.org> on 2020/10/16 18:49:21 UTC
[GitHub] [incubator-mxnet] Zha0q1 opened a new issue #19362: Float16 Argmin/max gives wrong results
Zha0q1 opened a new issue #19362:
URL: https://github.com/apache/incubator-mxnet/issues/19362
``` python
import mxnet as mx
from mxnet import np, nd, npx, profiler
import numpy as _np
from mxnet.gluon import HybridBlock
AA = np.random.normal(0, 10, size=(2,5,99999), dtype='float16', ctx=mx.gpu())
BB = np.argmin(AA, 2)
CC = _np.argmin(AA.asnumpy(), 2)
print(BB)
print(CC)
```
outputs:
```
[ 39904 48736 58816
2708 36896 11056
45600 9223372036854775807 8020
9223372036854775807 9223372036854775807 9223372036854775807
9223372036854775807 34176 30160
19456 9223372036854775807 12008
9223372036854775807 52640] @gpu(0)
[39904 48728 58812 2709 36908 11059 45615 69962 8019 68762 83795 96961
89943 34176 30167 19451 65951 12011 75882 52637]
```
This is reproducible on master and 1.7, both on cpu and gpu. This has not been noticed before because in `test_np_argmin_argmax` we were only using very small tensors. I added in a large tensor case yesterday and multiple ci failed https://github.com/apache/incubator-mxnet/pull/19359.
This might be a memory alignment issue. We probably also need to see if any other operators are affected too.
----------------------------------------------------------------
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:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@mxnet.apache.org
For additional commands, e-mail: issues-help@mxnet.apache.org
[GitHub] [incubator-mxnet] sxjscience commented on issue #19362: Float16 Argmin/max gives wrong results
Posted by GitBox <gi...@apache.org>.
sxjscience commented on issue #19362:
URL: https://github.com/apache/incubator-mxnet/issues/19362#issuecomment-710419795
I can also reproduce.
----------------------------------------------------------------
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:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@mxnet.apache.org
For additional commands, e-mail: issues-help@mxnet.apache.org
[GitHub] [incubator-mxnet] sandeep-krishnamurthy closed issue #19362: Float16 Argmin/max gives wrong results
Posted by GitBox <gi...@apache.org>.
sandeep-krishnamurthy closed issue #19362:
URL: https://github.com/apache/incubator-mxnet/issues/19362
----------------------------------------------------------------
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:
users@infra.apache.org
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@mxnet.apache.org
For additional commands, e-mail: issues-help@mxnet.apache.org