| --- |
| license: bsd-3-clause |
| tags: |
| - executorch |
| - xnnpack |
| - pte |
| - on-device |
| - image-classification |
| --- |
| # efficientnet_b1 — ExecuTorch |
| |
| - **Source**: torchvision efficientnet_b1 IMAGENET1K_V2 |
| - **License**: BSD-3-Clause |
| - **Input**: [[1, 3, 240, 240]] — RGB, ImageNet norm, 240x240 |
| - **Output**: ImageNet logits [1,1000] |
| |
| ## Variants |
| |
| All variants take and return fp32 tensors — swap the `.pte` file, keep your app code. |
| |
| | build | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* | |
| |-----------|------|-----------|------------------------------------|------------------| |
| | fp32 | `efficientnet_b1_xnnpack_fp32.pte` | 31.2 | 1.000000 | 9.3 | |
| | fp16 | `efficientnet_b1_xnnpack_fp16.pte` | 28.8 | 0.999816 | 54.9 | |
| | Core ML (fp16, iOS) | `efficientnet_b1_coreml_all.pte` | 16.3 | 0.992817 — see below | 0.6 | |
|
|
|
|
| The Core ML build is the same graph lowered to Apple's Neural Engine instead of |
| XNNPACK, which is CPU-only. Measured on an iPhone 17 Pro across seven models, it |
| runs **3.5x to 13.9x faster (median 12x)** at roughly half the file size — for |
| example Depth-Anything-V2-Small at 500.8 ms against 42.7 ms, and MODNet at 81.7 ms |
| against 5.9 ms. It computes in fp16 and is iOS-only; the XNNPACK files stay the |
| portable option and are what runs on Android. |
|
|
| \*Mac arm64, single process, median of 10 — a reference point for relative cost |
| only, not a device number (torch eager fp32 on the same machine: 352.7 ms). |
| |
| ### Checked in the task's own units |
| |
| Correlation is a first filter. These are the numbers that decide: |
| |
| - **Core ML (fp16, iOS)** — measured in the units that matter for this model — fraction of images keeping the fp32 top-1 label: 9 of 10 images keep the fp32 top-1 label. |
| |
| ### Builds that did not earn a slot |
| |
| - **int8 is not shipped**: measured in the units that matter for this model — fraction of images keeping the fp32 top-1 label: 0 of 10 images keep the fp32 top-1 label. |
| |
| ## Verification (executorch 1.4.0, torch 2.13.0) |
| |
| Parity is measured against the fp32 eager model on real image input; `corr` is |
| the correlation over all elements of each output tensor. |
| |
| | output | shape | max_abs_diff | corr | |
| |--------|-------|--------------|------| |
| | 0 | [1, 1000] | 1.490e-05 | 1.000000 | |
| |
| XNNPACK delegate coverage (fp32): 100.0% (410/410 ops) |
| |
| ## Conversion |
| |
| torch.export -> to_edge_transform_and_lower(partitioner) -> .pte |
| (conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models)) |
| |
