I fllow this tutorial to compile keras model to tvm model, but I use cpu device, so I change target = “llvm”, and ctx = tvm.cpu(). But I found that the speed between tvm and keras on resnet50 is shown as follow, I don’t understand why this so huge gap?
File /workspace/.tvm_test_data/data/imagenet1000_clsid_to_human.txt exists, skip.
Relay top-1 id: 278, class name: kit fox, Vulpes macrotis, take time: 14.048312187194824
Keras top-1 id: 278, class name: kit fox, Vulpes macrotis, take time:0.24698472023010254