ashokei commented on issue #12502: [NGRAPH] MXNet - nGraph initial integration URL: https://github.com/apache/incubator-mxnet/pull/12502#issuecomment-454131909 @zheng-da MKLDNN issue has been resolved, this PR uses the default MXNet MKLDNN. So, when users build with `make USE_NGRAPH=1` , all the following options work ``` MXNET_SUBGRAPH_BACKEND="" python example/image-classification/benchmark_score.py --network resnet-50 MXNET_SUBGRAPH_BACKEND="MKLDNN" python example/image-classification/benchmark_score.py --network resnet-50 MXNET_SUBGRAPH_BACKEND="ngraph" python example/image-classification/benchmark_score.py --network resnet-50 ``` With this PR (built with `make USE_NGRAPH=1`) , We ran resnet-50 on AWS c5.18xlarge instance with benchmark_score.py, below are the results (`MXNET_SUBGRAPH_BACKEND=""` uses default MXNet MKLDNN ops) ``` ubuntu@aws-c5-18xlarge:~/mxnet-upstream$ MXNET_SUBGRAPH_BACKEND="" OMP_NUM_THREADS=18 env KMP_BLOCKTIME=1 env KMP_AFFINITY=granularity=fine,compact,1,0 python example/image-classification/benchmark_score.py --network resnet-50 INFO:root:run batchsize [1, 32, 64, 128, 256] by default, set --batch-size to run a specific one INFO:root:network: resnet-50 INFO:root:device: cpu(0) /home/ubuntu/mxnet-upstream/python/mxnet/module/base_module.py:67: UserWarning: Data provided by label_shapes don't match names specified by label_names ([] vs. ['softmax_label']) warnings.warn(msg) INFO:root:batch size 1, dtype float32, images/sec: 90.258317 INFO:root:batch size 32, dtype float32, images/sec: 152.033810 INFO:root:batch size 64, dtype float32, images/sec: 151.357087 INFO:root:batch size 128, dtype float32, images/sec: 150.144761 INFO:root:batch size 256, dtype float32, images/sec: 148.233113 ubuntu@aws-c5-18xlarge:~/mxnet-upstream$ MXNET_SUBGRAPH_BACKEND="MKLDNN" OMP_NUM_THREADS=18 env KMP_BLOCKTIME=1 env KMP_AFFINITY=granularity=fine,compact,1,0 python example/image-classification/benchmark_score.py --network resnet-50 INFO:root:run batchsize [1, 32, 64, 128, 256] by default, set --batch-size to run a specific one INFO:root:network: resnet-50 INFO:root:device: cpu(0) /home/ubuntu/mxnet-upstream/python/mxnet/module/base_module.py:67: UserWarning: Data provided by label_shapes don't match names specified by label_names ([] vs. ['softmax_label']) warnings.warn(msg) [19:27:50] src/operator/subgraph/mkldnn/mkldnn_conv_property.cc:138: Start to execute MKLDNN Convolution optimization pass. INFO:root:batch size 1, dtype float32, images/sec: 110.765558 [19:27:50] src/operator/subgraph/mkldnn/mkldnn_conv_property.cc:138: Start to execute MKLDNN Convolution optimization pass. INFO:root:batch size 32, dtype float32, images/sec: 191.871125 [19:27:53] src/operator/subgraph/mkldnn/mkldnn_conv_property.cc:138: Start to execute MKLDNN Convolution optimization pass. INFO:root:batch size 64, dtype float32, images/sec: 191.062513 [19:27:58] src/operator/subgraph/mkldnn/mkldnn_conv_property.cc:138: Start to execute MKLDNN Convolution optimization pass. INFO:root:batch size 128, dtype float32, images/sec: 193.064107 [19:28:09] src/operator/subgraph/mkldnn/mkldnn_conv_property.cc:138: Start to execute MKLDNN Convolution optimization pass. INFO:root:batch size 256, dtype float32, images/sec: 191.515350 ubuntu@aws-c5-18xlarge:~/mxnet-upstream$ MXNET_SUBGRAPH_BACKEND="ngraph" OMP_NUM_THREADS=18 env KMP_BLOCKTIME=1 env KMP_AFFINITY=granularity=fine,compact,1,0 python example/image-classification/benchmark_score.py --network resnet-50 INFO:root:run batchsize [1, 32, 64, 128, 256] by default, set --batch-size to run a specific one INFO:root:network: resnet-50 INFO:root:device: cpu(0) /home/ubuntu/mxnet-upstream/python/mxnet/module/base_module.py:67: UserWarning: Data provided by label_shapes don't match names specified by label_names ([] vs. ['softmax_label']) warnings.warn(msg) INFO:root:batch size 1, dtype float32, images/sec: 125.942216 INFO:root:batch size 32, dtype float32, images/sec: 224.910995 INFO:root:batch size 64, dtype float32, images/sec: 227.315916 INFO:root:batch size 128, dtype float32, images/sec: 225.936182 INFO:root:batch size 256, dtype float32, images/sec: 226.969117 ```
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
