MobileNetV3-Large is the hardware-aware successor to MobileNetV2, adding squeeze-excite blocks and h-swish (~5.5M parameters). 224x224 RGB in, 1000 ImageNet class scores out.
MobileNetV3-Large image classifier (torchvision export), compiled for the
VKNN Vulkan inference engine.
Every op runs on the GPU (fp16 weights + fp16 compute, fp32 host boundary); the CPU backend serves
as the bit-accuracy oracle in VKNN's device gates, and this model's GPU outputs are gated against
an fp32 onnxruntime golden (cosine > 0.999, argmax agreement) on-device.
1# any VKNN example binary, e.g. the IO runner:
2./vknn_run_io mobilenetv3_fp16.vxm out_dir input.bin
