MobileNet-v1-1.0-224 was the first image classification workload [contributed to MLPerf Inference by dividiti] (https://github.com/mlperf/inference/pull/5). It consists of 3 equivalent implementations: TF (Python), TF (C++), TFLite. The latest TFLite instructions (not yet merged) are here.
All three implementations share the same package which can be installed as follows:
$ ck install package --tags=tensorflowmodel,mobilenet,mlperf
While TFLite is the preferred ArmNN frontend, perhaps this program can be made more universal to accept either tflite or tf (case-insensitive) via the CK_ARMNN_FRONTEND environment variable e.g.:
$ ck benchmark program:image-classification-armnn-mobilenet --env.CK_ARMNN_FRONTEND=tflite
Moreover, perhaps this program can be made even more universal to accept any image classification workload e.g.:
psyhtest
commented
5 years ago
MobileNet-v1-1.0-224 was the first image classification workload [contributed to MLPerf Inference by dividiti] (https://github.com/mlperf/inference/pull/5). It consists of 3 equivalent implementations: TF (Python), TF (C++), TFLite. The latest TFLite instructions (not yet merged) are here.
All three implementations share the same package which can be installed as follows:
While TFLite is the preferred ArmNN frontend, perhaps this program can be made more universal to accept either
tfliteortf(case-insensitive) via theCK_ARMNN_FRONTENDenvironment variable e.g.:Moreover, perhaps this program can be made even more universal to accept any image classification workload e.g.:
similarly to ck-tensorflow:program:image-classification-tf-cpp and its friends.