software-mansion/react-native-executorch-zipdepth

zipdepth

This repository hosts ZipDepth Base NPU (single-image depth estimation) exported for the React Native ExecuTorch library as ExecuTorch .pte programs, ready to run on device.

Upstream model: fabiotosi92/ZipDepth, checkpoint zipdepth_base_npu.pth, revision 91f3fd21e131641f51e8d35736d1958350180e3a.

Variants

Path	Backend	Precision
base/coreml/zipdepth_base_384_coreml_fp32.pte	coreml	fp32
base/vulkan/zipdepth_base_384_vulkan_fp16.pte	vulkan	fp16
base/xnnpack/zipdepth_base_384_xnnpack_fp32.pte	xnnpack	fp32

Core ML precision

The Core ML export uses FP32 because FP16 reduced numerical agreement with the float reference: across seven photos/illustrations, its worst range-normalized RMSE was 6.31% and minimum correlation was 0.970. FP32 reduced the worst range-normalized RMSE to below 0.00012%.

Unlike the FP16 Core ML exports, this FP32 ML program runs on CPU/GPU and cannot use the Apple Neural Engine, even with compute units set to ALL; see Apple's typed-execution guide.

Repository structure

config.json                                       30 B
base/coreml/config.json                           847 B
base/coreml/zipdepth_base_384_coreml_fp32.pte     24.9 MB
base/vulkan/config.json                           847 B
base/vulkan/zipdepth_base_384_vulkan_fp16.pte     12.8 MB
base/xnnpack/config.json                          849 B
base/xnnpack/zipdepth_base_384_xnnpack_fp32.pte   24.6 MB

Compatibility

These files are published for the ExecuTorch v1.4.1 runtime. ExecuTorch gives no forward compatibility guarantee, so an older runtime may fail to load them.

For React Native ExecuTorch model loading, see the documentation. To load these files in your own runtime, read the compatibility note first.