This tutorial shows how to execute 8 bit networks through MxNet with the included GoogLeNet v1 model. Each of mode of models has been put into a run script, that passes a few sample images, to show accuracy and measure performance.
For instructions on launching and connecting to instances, see here.
-
Connect to F1
-
Navigate to
/home/centos/xfdnn_17_12_15/mxnet/$ cd /home/centos/xfdnn_17_12_15/mxnet/ $ ls amalgamation cpp-package include MKL_README.md ps-lite tests appveyor.yml cub Jenkinsfile models python tools benchmark DISCLAIMER KEYS mshadow README.md xclbin bin dlpack lib mxnet_docker readthedocs.yml xfdnn_env build dmlc-core LICENSE NEWS.md R-package xlnx_docker cmake docker make nnvm scala-package xlnx_lib CMakeLists.txt docker_build_mxnet.sh Makefile NOTICE setup-utils xlnx_rt_xdnn CODEOWNERS start_docker.sh matlab perl-package snapcraft.yaml config.mk docs mkl plugin snap.python CONTRIBUTORS.md example mklml_lnx_2018.0.20170720.tgz prepare_mkl.sh src -
Execute
./start_docker.shto enter application docker and navigate to/opt/mxnet/examples/xlnx/simple-image-classify/$ ./start_docker.sh # cd /opt/mxnet/example/xlnx/simple-image-classify/ $ ls beagle.jpg classify_fpga.sh sdaccel_profile_summary.csv synset.txt cat.jpg classify.py sdaccel_profile_summary.html wolf.jpg -
Here is a example of running a single image with the model GoogLeNet v1. Run the
./classify_fpga.shscript and pass one of the example images here for it to classify.# ./classify_fpga.sh beagle.jpg XBLAS # FPGAs: 1 [XBLAS] # kernels: 1 [XDNN] using custom DDR banks 0,2,1,1 Device/Slot[0] (/dev/xdma0, 0:0:1d.0) xclProbe found 1 FPGA slots with XDMA driver running CL_PLATFORM_VENDOR Xilinx CL_PLATFORM_NAME Xilinx CL_DEVICE_0: 0x3898a20 CL_DEVICES_FOUND 1, using 0 loading /opt/mxnet/xclbin/kernelSxdnn_hw_f1_16b.xclbin [XBLAS] kernel0: kernelSxdnn_0 XBLAS online! (d=0) WARNING: unaligned host pointer detected, this leads to extra memcpy ------------------- probability=0.805006, class=n02088364 beagle probability=0.109553, class=n02089867 Walker hound, Walker foxhound probability=0.054156, class=n02089973 English foxhound probability=0.003512, class=n02101388 Brittany spaniel probability=0.002501, class=n02088632 bluetick -
To see the agreggated accuracy of GoogLeNet v1, navigate to
/opt/mxnet/example/image-classification/. Execute the example,./score_fpga.sh# cd /opt/mxnet/example/image-classification/ # ls README.md score.pyc symbol_inception-resnet-v1.R test_score.py __init__.py score_fpga.py symbol_inception-resnet-v2.R train_cifar10.R benchmark.py score_fpga.sh symbol_lenet.R train_cifar10.py benchmark_score.py sdaccel_profile_summary.csv symbol_mlp.R train_imagenet.R common sdaccel_profile_summary.html symbol_resnet-28-small.R train_imagenet.py data symbol_alexnet.R symbol_resnet-v2.R train_mnist.R fine-tune.py symbol_googlenet.R symbol_resnet.R train_mnist.py predict-cpp symbol_inception-bn-28-small.R symbol_vgg.R train_model.R score.py symbol_inception-bn.R symbols # ./score_fpga.sh Downloading dataset ... INFO:root:data/val_256_q90.rec exists, skipping download [06:17:30] src/io/iter_image_recordio_2.cc:153: ImageRecordIOParser2: data/val_256_q90.rec, use 3 threads for decoding.. XBLAS # FPGAs: 1 [XBLAS] # kernels: 1 [XDNN] using custom DDR banks 0,2,1,1 Device/Slot[0] (/dev/xdma0, 0:0:1d.0) xclProbe found 1 FPGA slots with XDMA driver running CL_PLATFORM_VENDOR Xilinx CL_PLATFORM_NAME Xilinx CL_DEVICE_0: 0x231a590 CL_DEVICES_FOUND 1, using 0 loading /opt/mxnet/xclbin/kernelSxdnn_hw_f1_16b.xclbin [XBLAS] kernel0: kernelSxdnn_0 XBLAS online! (d=0) WARNING: unaligned host pointer detected, this leads to extra memcpy WARNING: unaligned host pointer detected, this leads to extra memcpy WARNING: unaligned host pointer detected, this leads to extra memcpy WARNING: unaligned host pointer detected, this leads to extra memcpy INFO:root:('accuracy', 0.65239043824701193) INFO:root:('top_k_accuracy_5', 0.8635458167330677)At the end of the run, the top accuracy and top 5 accurary is listed:
INFO:root:('accuracy', 0.65239043824701193) INFO:root:('top_k_accuracy_5', 0.8635458167330677)