chinakook commented on issue #10599: [MKLDNN Bug] MKLDNN eats lots of memory
and then crash down.
URL:
https://github.com/apache/incubator-mxnet/issues/10599#issuecomment-383005826
@lionel92 The model is a widely used Caffe-Resnet101 with some 1x1
convolutions added and with big input
chinakook commented on issue #10599: [MKLDNN Bug] MKLDNN eats lots of memory
and then crash down.
URL:
https://github.com/apache/incubator-mxnet/issues/10599#issuecomment-382980794
I think the MKLDNN may use its own allocation engine, other than MXNet
allocation engine. The former may
chinakook commented on issue #10599: [MKLDNN Bug] MKLDNN eats lots of memory
and then crash down.
URL:
https://github.com/apache/incubator-mxnet/issues/10599#issuecomment-382980794
I think the MKLDNN may use its own allocation engine, other than MXNet
allocation engine. The former may
chinakook commented on issue #10599: [MKLDNN Bug] MKLDNN eats lots of memory
and then crash down.
URL:
https://github.com/apache/incubator-mxnet/issues/10599#issuecomment-382382640
Thanks! As I've tested before, when inferencing a scene segmentation network
such as Deeplab or U-Net,
