Sync model repo (text/metadata)

.gitattributes

@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 deit-tiny-int8-ethos-u-executorch.pte filter=lfs diff=lfs merge=lfs -text
+sample_input.jpg filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,267 @@
+---
+library_name: executorch
+display_name: DeiT-Tiny INT8 — Ethos-U85 (ExecuTorch)
+license: apache-2.0
+base_model: facebook/deit-tiny-patch16-224
+base_model_relation: quantized
+tags:
+  - image-classification
+  - deit
+  - vision-transformer
+  - int8
+  - quantized
+  - smoothquant
+  - ethos-u85
+  - npu
+  - arm
+  - executorch
+  - edge-ai
+  - imagenet-1k
+pipeline_tag: image-classification
+datasets:
+  - imagenet-1k
+metrics:
+  - accuracy
+model-index:
+  - name: deit-tiny-int8-ethos-u-executorch
+    results:
+      - task:
+          type: image-classification
+          name: Image Classification
+        dataset:
+          type: imagenet-1k
+          name: ImageNet-1K val
+          split: validation
+          args:
+            evaluation_samples: 50000
+        metrics:
+          - type: accuracy
+            value: 67.95
+            name: "Top-1 Accuracy (PyTorch INT8 simulation, not measured on-device)"
+          - type: accuracy
+            value: 88.47
+            name: "Top-5 Accuracy (PyTorch INT8 simulation, not measured on-device)"
+---
+
+# DeiT-Tiny INT8 SmoothQuant (ExecuTorch + Ethos-U85)
+
+This is an INT8-quantized version of [facebook/deit-tiny-patch16-224](https://huggingface.co/facebook/deit-tiny-patch16-224) (DeiT-Tiny, a data-efficient Vision Transformer from Meta / Facebook AI Research), optimized for the Arm Ethos-U85 NPU. The model was quantized using static PT2E quantization with SmoothQuant activation smoothing (INT8 activations/weights, with the MLP-down projection widened to INT16 activations for accuracy recovery), then ahead-of-time compiled with Arm's Vela compiler and exported to the ExecuTorch `.pte` format for execution on the Ethos-U85 NPU via the Corstone-320 Fixed Virtual Platform (FVP) or equivalent Ethos-U85 silicon (e.g., Alif Ensemble DK-E8).
+
+**Important:** the accuracy numbers reported below come from a **PyTorch eager-mode INT8-vs-FP32 comparison pass**, not from running the exported `.pte` on the Corstone-320 FVP. The FVP-based on-device accuracy check (`fvp_subsample` mode) was disabled for this run. Treat these numbers as a proxy for expected NPU behavior, not a measured on-hardware result. See [Known Limitations](#known-limitations) for details.
+
+## Key Highlights
+
+Compared to the FP32 baseline:
+
+- **3.93x smaller PyTorch state dict** — 21.89 MB to 5.57 MB; the exported Ethos-U85 `.pte` artifact itself is 4.71 MB
+- **99.91% of ops NPU-delegated at compile time** — Vela's static analysis delegates 3,494 of 3,497 ops to the Ethos-U85, with only 3 ops falling back to the host CPU (compile-time metric, not a latency measurement)
+- **-4.18 pp Top-1 / -2.66 pp Top-5 accuracy** vs FP32, measured via PyTorch INT8 simulation on the full 50,000-image ImageNet-1K validation set — not yet confirmed on the exported `.pte`/FVP
+
+## Model Details
+
+### Model Description
+
+DeiT-Tiny is a data-efficient Vision Transformer for ImageNet-1K classification, distilled and trained by Facebook AI Research. This is the INT8-quantized, Ethos-U85-targeted export of that model — no finetuning of weights was performed beyond quantization.
+
+- **Developed by:** Facebook AI Research (Meta)
+- **Model type:** Image Classification (Vision Transformer / DeiT)
+- **License:** Apache 2.0
+- **Base model:** [facebook/deit-tiny-patch16-224](https://huggingface.co/facebook/deit-tiny-patch16-224) — quantized, not finetuned
+
+### Model Sources
+
+- **Repository:** https://huggingface.co/facebook/deit-tiny-patch16-224
+
+## How to Get Started with the Model
+
+### Install dependencies
+
+When running from this repository, install `uv` and synchronize the project environment:
+
+```bash
+make install-uv
+uv sync
+```
+
+The example itself uses NumPy and Pillow for preprocessing/postprocessing. It does **not** use the desktop ExecuTorch Python runtime; the compiled model is executed by `arm_executor_runner` inside Corstone-320 FVP.
+
+### Download the model
+
+```python
+from huggingface_hub import hf_hub_download
+
+model_path = hf_hub_download(
+    repo_id="your-org/deit-tiny-int8-smoothquant-ethos-u85-executorch",
+    filename="deit-tiny_ethos_ethosu_optimized.pte",
+)
+```
+
+Place the downloaded `deit-tiny_ethos_ethosu_optimized.pte` next to `example.py`, or pass a different path with `--model`.
+
+### Install Corstone-320 and the runner
+
+The example requires both:
+
+```text
+third_party/corstone-320-fvp/bin/FVP_Corstone_SSE-320
+third_party/executorch-arm-runner/cmake-out/arm_executor_runner
+```
+
+On Linux, install the supported toolchain from the repository root:
+
+```bash
+make install-corstone
+```
+
+#### macOS setup
+
+Corstone-320 has no native macOS binary in this setup. It runs through the [`FVPs-on-Mac`](https://github.com/Arm-Examples/FVPs-on-Mac) Docker wrapper. Docker Desktop must be installed and running:
+
+```bash
+open -a Docker
+docker info
+```
+
+If the wrapper is not installed yet:
+
+```bash
+git clone https://github.com/Arm-Examples/FVPs-on-Mac.git ~/tools/FVPs-on-Mac
+cd ~/tools/FVPs-on-Mac
+./build.sh
+echo 'export PATH="$PATH:$HOME/tools/FVPs-on-Mac/bin"' >> ~/.zshrc
+source ~/.zshrc
+FVP_Corstone_SSE-320 --version
+```
+
+The wrapper requires an activated Arm User Based License, normally cached under `~/.armlm`. It mounts that cache into the FVP container automatically. Once Docker, the wrapper and license are ready, return to this repository and build/install the runner:
+
+```bash
+export CXXFLAGS="-Wno-invalid-specialization"
+export CFLAGS="-Wno-invalid-specialization"
+make install-corstone
+```
+
+If `docker info` reports `Cannot connect to the Docker daemon`, wait for Docker Desktop to finish starting before continuing. The first FVP invocation may build its local Docker image and take longer.
+
+### Run inference
+
+From the repository root:
+
+```bash
+uv run python models_reports/deit_tiny_ethosu85_executorch/huggingface/example.py
+```
+
+The script performs genuine inference on Corstone-320 FVP: it preprocesses `sample_input.jpg`, stages the `.pte` and FP32 input through semihosting, invokes `arm_executor_runner`, reads the `[1, 1000]` logits, prints the Top-5 ImageNet predictions and overwrites:
+
+```text
+models_reports/deit_tiny_ethosu85_executorch/huggingface/predictions.json
+```
+
+Useful overrides are exposed as command-line arguments:
+
+```bash
+uv run python models_reports/deit_tiny_ethosu85_executorch/huggingface/example.py \
+  --model /path/to/model.pte \
+  --image /path/to/image.jpg \
+  --labels /path/to/imagenet_classes.json \
+  --fvp-bin /path/to/FVP_Corstone_SSE-320 \
+  --runner-elf /path/to/arm_executor_runner \
+  --timelimit 1800
+```
+
+The `.pte` cannot run through the standard desktop ExecuTorch CPU/XNNPACK runtime because it contains an Ethos-U85 delegate. Actual Ethos-U85 silicon is also supported, but this particular example script is an FVP semihosting runner.
+
+## Evaluation
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+Evaluated on 50,000 images from ImageNet-1K (full validation split).
+
+#### Metrics
+
+- **Top-1 / Top-5 Accuracy** — standard ImageNet classification metrics, computed here as a **PyTorch eager-mode FP32-vs-INT8 comparison**, not a measurement of the exported `.pte` running on the Corstone-320 FVP
+
+### Results
+
+#### Accuracy
+
+| Metric | FP32 (Original) | INT8 SmoothQuant (PyTorch simulation) | Delta |
+|:---|:---:|:---:|:---:|
+| Top-1 Accuracy | 72.13% | 67.95% | -4.18 pp |
+| Top-5 Accuracy | 91.13% | 88.47% | -2.66 pp |
+
+These numbers come from a PyTorch-only FP32-vs-INT8 comparison pass over all 50,000 ImageNet-1K validation images. They are a proxy for what the Ethos-U85 NPU should produce, not a value measured from the exported `.pte` on the FVP in this run.
+
+#### Efficiency
+
+| Metric | Value | What it measures |
+|:---|:---:|:---|
+| PyTorch State Dict Size — FP32 | 21.89 MB | Serialized `state_dict()` size |
+| PyTorch State Dict Size — INT8 | 5.57 MB (3.93x smaller) | Serialized `state_dict()` size |
+| Ethos-U85 `.pte` Size — INT8 | 4.71 MB | Real exported artifact file size |
+| NPU Delegation Coverage (Vela) | 99.91% (3,494 delegated / 3 host ops) | Static, export-time compiler metric — not runtime |
+| Vela MAC Count (per batch) | 1,271,114,184 | Static, export-time compiler metric — not runtime |
+
+**No inference latency was measured in this experiment.** No `RuntimeEvaluator` / Corstone-320 FVP timing run was performed for this report; the delegation coverage and MAC count above are static compiler outputs from Vela, not wall-clock measurements. See [Known Limitations](#known-limitations).
+
+## Technical Specifications
+
+### Objective
+
+1,000-class image classification on ImageNet using a tiny Vision Transformer (DeiT) architecture.
+
+### Quantization
+
+- **Method:** Static post-training quantization (PT2E) with SmoothQuant activation smoothing applied before observer insertion, targeting the Ethos-U85 Vela compile path
+- **Base precision:** INT8 symmetric weights, per-tensor (`is_per_channel: false`), INT8 activations (a8w8) globally
+- **Per-layer override:** `output.dense` (the MLP-down projection in each transformer block) widened to INT16 activations / INT8 weights (a16w8) for extra accuracy recovery
+- **Activation observer:** Histogram
+- **Calibration:** 2,000 randomly selected ImageNet-1K images
+- **Target:** `ethos-u85-256`, system config `Ethos_U85_SRAM_MRAM`, memory mode `Shared_Sram`
+
+### Export Pipeline
+
+1. Load pretrained FP32 `facebook/deit-tiny-patch16-224`
+2. Apply SmoothQuant activation smoothing
+3. Capture model graph via `torch.export`
+4. Insert quantization observers (global a8w8, with `output.dense` overridden to a16w8)
+5. Calibrate with 2,000 representative ImageNet-1K images
+6. Convert observers to Q/DQ pairs
+7. Ahead-of-time compile the quantized graph with Vela for the `ethos-u85-256` target
+8. Export to ExecuTorch `.pte` for semihosting execution via the Ethos-U85 NPU delegate
+
+### Preprocessing
+
+| Property | Value |
+|---|---|
+| Input shape | `[1, 3, 224, 224]` (BCHW) |
+| Data type | float32 |
+| Value range | [0.0, 1.0] |
+| Color space | RGB |
+
+**Steps**:
+1. Resize shortest side to 256 (bicubic interpolation)
+2. Center crop to 224x224
+3. Convert to tensor ([0-255] uint8 -> [0, 1] float32)
+
+**Normalization**: ImageNet mean/std (`mean=[0.485, 0.456, 0.406]`, `std=[0.229, 0.224, 0.225]`).
+
+### Postprocessing
+
+| Property | Value |
+|---|---|
+| Output shape | `[1, 1000]` |
+| Output format | Raw class logits (unnormalized), ImageNet-1K class ordering |
+
+**Steps**:
+1. Apply softmax over the 1,000 logits
+2. Take the top-5 classes by probability
+
+## Known Limitations
+
+- **Accuracy is a PyTorch simulation, not an on-device measurement.** The Top-1/Top-5 numbers above come from a PyTorch eager-mode FP32-vs-INT8 comparison over the full ImageNet-1K validation set. The FVP-based accuracy mode (`accuracy_modes.fvp_subsample`) that would compare PyTorch vs the actual exported `.pte` running on the Corstone-320 FVP was disabled for this run — there is no measured on-FVP/on-silicon accuracy number to report yet.
+- **No runtime latency was measured.** This experiment did not run a `RuntimeEvaluator` / Corstone-320 FVP timing pass. The 99.91% NPU delegation coverage and the ~1.27B MAC count are static, export-time compiler metrics produced by Vela — they indicate how much of the graph the NPU can execute and how much compute it involves, not how fast it runs.
+- **The `.pte` requires Ethos-U85 hardware or the FVP to execute.** It cannot be run for real inference on a standard desktop/CPU ExecuTorch Python runtime, since the `EthosUBackend` delegate is not registered there — see the `RuntimeError` behavior in [How to Get Started](#how-to-get-started-with-the-model).
+- **Fixed input size** of 224x224, batch size 1 — no dynamic shapes.

benchmarks/README.md

@@ -0,0 +1,21 @@
+# Benchmarks
+
+Machine-readable measurement records for this model (`vision-classification` schema).
+
+## Files
+
+- `deit-tiny-int8-ethos-u-executorch-alif-dk-e8-int8.yaml` — optimized INT8 measurements on the Alif DK-E8
+  (Cortex-M55 host + Ethos-U85 NPU) via the ExecuTorch `ethos-u` backend.
+- `deit-tiny-int8-ethos-u-executorch-alif-dk-e8-fp32.yaml` — FP32 reference accuracy (PyTorch; the baseline was
+  not compiled to the NPU, so no on-device performance).
+
+## Ethos-U metrics deferred
+
+These are **regular vision-classification** records for now. The delivery also carries
+NPU-specific metrics that the 0.6.0 schema's `additionalProperties: false` rejects —
+target `npu_model`/`npu_num_macs` and performance `sram_peak_bytes`, `flash_footprint_mib`,
+`mac_utilization_pct`, `total_cycles`, `active_cycles`, `effective_cycles_ratio_pct`. They
+are **preserved in `metadata.yaml`** and will be added back to these records once a schema
+holds the embedded-NPU profile (see [`upstream.md`](../../../docs/marvik/upstream.md) and
+[ai-portal-model-schema#4](https://github.com/Arm-Collaboration/ai-portal-model-schema/pull/4)).
+`peak_memory_mb` here is the SRAM working-memory peak (`sram_peak_bytes / 1048576`).

benchmarks/deit-tiny-int8-ethos-u-executorch-alif-dk-e8-fp32.yaml

@@ -0,0 +1,39 @@
+# FP32 baseline measurement record — DeiT-Tiny INT8 reference accuracy.
+# The FP32 baseline was NOT compiled/run on the Ethos-U85 (delivered as a PyTorch
+# reference, performance: null), so this carries reference accuracy only — no on-device
+# latency/memory. runtime/benchmark describe the PyTorch eval on the ImageNet workload.
+version: 1.0.0
+task: vision-classification
+created_at: '2026-07-14T21:00:10Z'
+context:
+  model:
+    id: Arm/deit-tiny-int8-ethos-u-executorch
+    filename: not-applicable
+    base_model_id: facebook/deit-tiny-patch16-224
+    profile: Baseline
+    weight_dtype: fp32
+    model_size_mb: 21.887
+    format: pytorch_state_dict
+  target:
+    name: Alif DK-E8
+    hardware_class: Embedded NPU
+    cpu_architecture: armv8.1-m
+    cpu_model: Cortex-M55
+    cpu_core_count: 1
+    cpu_clock_ghz: 0.4
+    os: bare-metal
+  runtime:
+    name: pytorch
+    execution_backend: cpu
+  dataset:
+    name: ImageNet-1k
+    sample_count: 50000
+    reference_url: https://www.image-net.org/
+    slice: val
+  benchmark:
+    batch_size: 1
+    input_resolution: 224x224
+    num_runs: 1
+accuracy:
+  top_1_accuracy_pct: 72.132
+  top_5_accuracy_pct: 91.134

benchmarks/deit-tiny-int8-ethos-u-executorch-alif-dk-e8-int8.yaml

@@ -0,0 +1,66 @@
+# Optimized measurement record — DeiT-Tiny INT8 on Alif DK-E8 (Ethos-U85 NPU, ExecuTorch).
+# Onboarded as a regular vision-classification model; NPU-specific metrics are omitted
+# here until the schema supports them (docs/marvik/upstream.md + ai-portal-model-schema#4):
+# target npu_model/npu_num_macs and performance sram_peak_bytes, flash_footprint_mib,
+# mac_utilization_pct, total_cycles, active_cycles, effective_cycles_ratio_pct — all
+# retained in metadata.yaml. peak_memory_mb is the SRAM working-memory peak
+# (sram_peak_bytes / 1048576).
+version: 1.0.0
+task: vision-classification
+created_at: '2026-07-14T21:00:10Z'
+context:
+  model:
+    id: Arm/deit-tiny-int8-ethos-u-executorch
+    filename: deit-tiny-int8-ethos-u-executorch.pte
+    base_model_id: facebook/deit-tiny-patch16-224
+    profile: Arm-Optimized
+    weight_dtype: int8
+    quantization:
+      method: PTQ-static
+      weight_bits: 8
+      activation_bits: 8
+      symmetric: true
+      mode: static
+      weight_granularity: per-tensor
+    calibration:
+      dataset_name: ImageNet-1k
+      sample_count: 2000
+      selection: random
+    model_size_mb: 4.707
+    format: pte
+  target:
+    name: Alif DK-E8
+    hardware_class: Embedded NPU
+    cpu_architecture: armv8.1-m
+    cpu_model: Cortex-M55
+    cpu_core_count: 1
+    cpu_clock_ghz: 0.4
+    os: bare-metal
+  runtime:
+    name: executorch
+    execution_backend: ethos-u
+    config:
+      optimisations:
+      - CMSIS-NN
+      - Ethos-U-Driver
+    version: 1.1.0
+  dataset:
+    name: ImageNet-1k
+    sample_count: 50000
+    reference_url: https://www.image-net.org/
+    slice: val
+  benchmark:
+    batch_size: 1
+    input_resolution: 224x224
+    num_runs: 100
+    warmup_runs: 10
+performance:
+  end_to_end_latency_ms:
+    p50: 81.5947
+    p90: 81.5948
+  peak_memory_mb: 2.558
+  frames_per_second: 12.2557
+  delegation_pct: 99.91
+accuracy:
+  top_1_accuracy_pct: 67.95
+  top_5_accuracy_pct: 88.47

config.yaml

@@ -0,0 +1,14 @@
+input:
+  shape: [1, 3, 224, 224]
+  dtype: float32
+  range: [0.0, 1.0]
+  preprocessing:
+    - to_tensor
+    - normalize: { mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225] }
+
+output:
+  shape: [1, 1000]
+  format: "Raw class logits (unnormalized), ImageNet-1k class ordering"
+  postprocessing:
+    softmax: true
+    top_k: 5

example.py

@@ -0,0 +1,267 @@
+"""Run DeiT-Tiny Ethos-U85 ExecuTorch inference on Corstone-320 FVP.
+
+This example preprocesses one image, stages the compiled ``.pte`` and raw
+FP32 input tensor into a semihosting workdir, invokes ``FVP_Corstone_SSE-320``
+with the bare-metal ``arm_executor_runner`` ELF, then reads the FP32 logits
+written back by the runner and prints the top-5 ImageNet predictions.
+
+Prerequisites:
+  * ``FVP_Corstone_SSE-320`` is on PATH.
+  * ``arm_executor_runner`` was built with ``--pte=semihosting``.
+
+From this repository, the default runner path produced by ``make
+install-corstone`` is:
+
+  third_party/executorch-arm-runner/cmake-out/arm_executor_runner
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import shutil
+import subprocess
+import uuid
+from pathlib import Path
+
+import numpy as np
+from PIL import Image
+
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+REPO_ROOT = SCRIPT_DIR.parents[2]
+
+MODEL_PATH = SCRIPT_DIR / "deit-tiny-int8-ethos-u-executorch.pte"
+IMAGE_PATH = SCRIPT_DIR / "sample_input.jpg"
+LABELS_PATH = SCRIPT_DIR / "imagenet_classes.json"
+DEFAULT_WORKDIR = SCRIPT_DIR / "fvp_workdir"
+DEFAULT_RUNNER_ELF = (
+    REPO_ROOT / "third_party/executorch-arm-runner/cmake-out/arm_executor_runner"
+)
+DEFAULT_FVP_BIN = REPO_ROOT / "third_party/corstone-320-fvp/bin/FVP_Corstone_SSE-320"
+
+INPUT_SIZE = (224, 224)  # (H, W)
+RESIZE_SIZE = 256
+IMAGENET_MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
+IMAGENET_STD = np.array([0.229, 0.224, 0.225], dtype=np.float32)
+TOP_K = 5
+NUM_MACS = 256
+SEMIHOSTING_CMD_LINE_MAX = 256
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Run a DeiT-Tiny Ethos-U85 .pte on Corstone-320 FVP."
+    )
+    parser.add_argument("--model", type=Path, default=MODEL_PATH)
+    parser.add_argument("--image", type=Path, default=IMAGE_PATH)
+    parser.add_argument("--labels", type=Path, default=LABELS_PATH)
+    parser.add_argument("--workdir", type=Path, default=DEFAULT_WORKDIR)
+    parser.add_argument(
+        "--fvp-bin",
+        default=os.environ.get("FVP_Corstone_SSE_320", str(DEFAULT_FVP_BIN)),
+        help="Path to FVP_Corstone_SSE-320, or a binary name available on PATH.",
+    )
+    parser.add_argument(
+        "--runner-elf",
+        type=Path,
+        default=Path(os.environ.get("ARM_EXECUTOR_RUNNER", DEFAULT_RUNNER_ELF)),
+        help="Path to arm_executor_runner built with --pte=semihosting.",
+    )
+    parser.add_argument("--timelimit", type=int, default=1800)
+    return parser.parse_args()
+
+
+def preprocess(image_path: Path) -> np.ndarray:
+    """Resize -> center crop -> normalize -> NCHW FP32."""
+    image = Image.open(image_path).convert("RGB")
+
+    w, h = image.size
+    if w < h:
+        new_w, new_h = RESIZE_SIZE, round(h * RESIZE_SIZE / w)
+    else:
+        new_h, new_w = RESIZE_SIZE, round(w * RESIZE_SIZE / h)
+    image = image.resize((new_w, new_h), Image.BICUBIC)
+
+    left = (new_w - INPUT_SIZE[1]) // 2
+    top = (new_h - INPUT_SIZE[0]) // 2
+    image = image.crop((left, top, left + INPUT_SIZE[1], top + INPUT_SIZE[0]))
+
+    array = np.asarray(image, dtype=np.float32) / 255.0
+    array = array.transpose(2, 0, 1)
+    array = (array - IMAGENET_MEAN.reshape(3, 1, 1)) / IMAGENET_STD.reshape(3, 1, 1)
+    return np.expand_dims(array.astype(np.float32), axis=0)
+
+
+def load_labels(labels_path: Path) -> list[str] | None:
+    if not labels_path.exists():
+        return None
+    with labels_path.open() as f:
+        labels = json.load(f)
+    if isinstance(labels, dict):
+        return [labels[str(i)] for i in range(len(labels))]
+    return labels
+
+
+def build_fvp_cmd(
+    *,
+    fvp_bin: str,
+    runner_elf: Path,
+    workdir: Path,
+    cmd_line: str,
+    timelimit: int,
+) -> list[str]:
+    return [
+        fvp_bin,
+        "-C",
+        f"mps4_board.subsystem.ethosu.num_macs={NUM_MACS}",
+        "-C",
+        "mps4_board.visualisation.disable-visualisation=1",
+        "-C",
+        "vis_hdlcd.disable_visualisation=1",
+        "-C",
+        "mps4_board.telnetterminal0.start_telnet=0",
+        "-C",
+        "mps4_board.uart0.out_file=-",
+        "-C",
+        "mps4_board.uart0.shutdown_on_eot=1",
+        "-C",
+        "mps4_board.subsystem.cpu0.semihosting-enable=1",
+        "-C",
+        "mps4_board.subsystem.ethosu.extra_args='--fast'",
+        "-C",
+        "mps4_board.subsystem.cpu0.semihosting-stack_base=0",
+        "-C",
+        "mps4_board.subsystem.cpu0.semihosting-heap_limit=0",
+        "-C",
+        f"mps4_board.subsystem.cpu0.semihosting-cwd={workdir}",
+        "-C",
+        f"mps4_board.subsystem.cpu0.semihosting-cmd_line='{cmd_line}'",
+        "-a",
+        str(runner_elf),
+        "--timelimit",
+        str(timelimit),
+    ]
+
+
+def run_fvp(
+    *,
+    model_path: Path,
+    input_tensor: np.ndarray,
+    workdir: Path,
+    fvp_bin: str,
+    runner_elf: Path,
+    timelimit: int,
+) -> np.ndarray:
+    workdir.mkdir(parents=True, exist_ok=True)
+
+    run_id = uuid.uuid4().hex[:8]
+    staged_model = workdir / model_path.name
+    input_path = workdir / f"input_{run_id}.bin"
+    output_base = f"out_{run_id}"
+    output_path = workdir / f"{output_base}-0.bin"
+
+    if not staged_model.exists() or staged_model.stat().st_size != model_path.stat().st_size:
+        shutil.copyfile(model_path, staged_model)
+    input_path.write_bytes(np.ascontiguousarray(input_tensor, dtype=np.float32).tobytes())
+
+    cmd_line = f"arm_executor_runner -m {staged_model.name} -i {input_path.name} -o {output_base}"
+    if len(cmd_line) > SEMIHOSTING_CMD_LINE_MAX:
+        raise ValueError(
+            f"semihosting cmd_line too long ({len(cmd_line)} > {SEMIHOSTING_CMD_LINE_MAX})"
+        )
+
+    fvp_cmd = build_fvp_cmd(
+        fvp_bin=fvp_bin,
+        runner_elf=runner_elf,
+        workdir=workdir.resolve(),
+        cmd_line=cmd_line,
+        timelimit=timelimit,
+    )
+    result = subprocess.run(fvp_cmd, capture_output=True, text=True, check=False)
+    if result.returncode != 0:
+        raise RuntimeError(
+            f"FVP exited with code {result.returncode}\n"
+            f"stdout tail:\n{result.stdout[-2048:]}\n"
+            f"stderr tail:\n{result.stderr[-2048:]}"
+        )
+    if not output_path.exists():
+        raise RuntimeError(
+            f"Runner finished but {output_path} was not produced.\n"
+            f"stdout tail:\n{result.stdout[-2048:]}\n"
+            f"stderr tail:\n{result.stderr[-2048:]}"
+        )
+
+    logits = np.fromfile(output_path, dtype=np.float32)
+    if logits.size != 1000:
+        raise ValueError(f"Expected 1000 output logits, got {logits.size}")
+    return logits.reshape(1, 1000)
+
+
+def postprocess(raw_output: np.ndarray, labels: list[str] | None) -> list[dict]:
+    logits = raw_output.astype(np.float64)
+    logits = logits - logits.max(axis=-1, keepdims=True)
+    probabilities = np.exp(logits)
+    probabilities = probabilities / probabilities.sum(axis=-1, keepdims=True)
+    indices = np.argsort(probabilities[0])[::-1][:TOP_K]
+
+    results = []
+    for index in indices:
+        label = labels[int(index)] if labels is not None else None
+        results.append(
+            {
+                "class_index": int(index),
+                "label": label,
+                "probability": float(probabilities[0, index]),
+            }
+        )
+    return results
+
+
+def save_results(results: list[dict]) -> None:
+    output_path = SCRIPT_DIR / "predictions.json"
+    with output_path.open("w") as f:
+        json.dump(results, f, indent=2)
+    print(f"Saved predictions to {output_path}")
+
+
+def main() -> None:
+    args = parse_args()
+    model_path = args.model.resolve()
+    image_path = args.image.resolve()
+    labels_path = args.labels.resolve()
+    runner_elf = args.runner_elf.resolve()
+
+    if not model_path.exists():
+        raise FileNotFoundError(f"Model not found: {model_path}")
+    if not image_path.exists():
+        raise FileNotFoundError(f"Image not found: {image_path}")
+    if not runner_elf.exists():
+        raise FileNotFoundError(
+            f"arm_executor_runner not found: {runner_elf}\n"
+            "Build it with make install-corstone, or pass --runner-elf."
+        )
+
+    labels = load_labels(labels_path)
+    input_tensor = preprocess(image_path)
+    raw_output = run_fvp(
+        model_path=model_path,
+        input_tensor=input_tensor,
+        workdir=args.workdir,
+        fvp_bin=args.fvp_bin,
+        runner_elf=runner_elf,
+        timelimit=args.timelimit,
+    )
+    results = postprocess(raw_output, labels)
+
+    print("Top-5 predictions:")
+    for rank, item in enumerate(results, start=1):
+        name = item["label"] if item["label"] is not None else f"class_{item['class_index']}"
+        print(f"  {rank}. {name} (index {item['class_index']}): {item['probability']:.4f}")
+
+    save_results(results)
+
+
+if __name__ == "__main__":
+    main()

imagenet_classes.json

@@ -0,0 +1,1002 @@
+[
+  "tench",
+  "goldfish",
+  "great white shark",
+  "tiger shark",
+  "hammerhead",
+  "electric ray",
+  "stingray",
+  "cock",
+  "hen",
+  "ostrich",
+  "brambling",
+  "goldfinch",
+  "house finch",
+  "junco",
+  "indigo bunting",
+  "robin",
+  "bulbul",
+  "jay",
+  "magpie",
+  "chickadee",
+  "water ouzel",
+  "kite",
+  "bald eagle",
+  "vulture",
+  "great grey owl",
+  "European fire salamander",
+  "common newt",
+  "eft",
+  "spotted salamander",
+  "axolotl",
+  "bullfrog",
+  "tree frog",
+  "tailed frog",
+  "loggerhead",
+  "leatherback turtle",
+  "mud turtle",
+  "terrapin",
+  "box turtle",
+  "banded gecko",
+  "common iguana",
+  "American chameleon",
+  "whiptail",
+  "agama",
+  "frilled lizard",
+  "alligator lizard",
+  "Gila monster",
+  "green lizard",
+  "African chameleon",
+  "Komodo dragon",
+  "African crocodile",
+  "American alligator",
+  "triceratops",
+  "thunder snake",
+  "ringneck snake",
+  "hognose snake",
+  "green snake",
+  "king snake",
+  "garter snake",
+  "water snake",
+  "vine snake",
+  "night snake",
+  "boa constrictor",
+  "rock python",
+  "Indian cobra",
+  "green mamba",
+  "sea snake",
+  "horned viper",
+  "diamondback",
+  "sidewinder",
+  "trilobite",
+  "harvestman",
+  "scorpion",
+  "black and gold garden spider",
+  "barn spider",
+  "garden spider",
+  "black widow",
+  "tarantula",
+  "wolf spider",
+  "tick",
+  "centipede",
+  "black grouse",
+  "ptarmigan",
+  "ruffed grouse",
+  "prairie chicken",
+  "peacock",
+  "quail",
+  "partridge",
+  "African grey",
+  "macaw",
+  "sulphur-crested cockatoo",
+  "lorikeet",
+  "coucal",
+  "bee eater",
+  "hornbill",
+  "hummingbird",
+  "jacamar",
+  "toucan",
+  "drake",
+  "red-breasted merganser",
+  "goose",
+  "black swan",
+  "tusker",
+  "echidna",
+  "platypus",
+  "wallaby",
+  "koala",
+  "wombat",
+  "jellyfish",
+  "sea anemone",
+  "brain coral",
+  "flatworm",
+  "nematode",
+  "conch",
+  "snail",
+  "slug",
+  "sea slug",
+  "chiton",
+  "chambered nautilus",
+  "Dungeness crab",
+  "rock crab",
+  "fiddler crab",
+  "king crab",
+  "American lobster",
+  "spiny lobster",
+  "crayfish",
+  "hermit crab",
+  "isopod",
+  "white stork",
+  "black stork",
+  "spoonbill",
+  "flamingo",
+  "little blue heron",
+  "American egret",
+  "bittern",
+  "crane bird",
+  "limpkin",
+  "European gallinule",
+  "American coot",
+  "bustard",
+  "ruddy turnstone",
+  "red-backed sandpiper",
+  "redshank",
+  "dowitcher",
+  "oystercatcher",
+  "pelican",
+  "king penguin",
+  "albatross",
+  "grey whale",
+  "killer whale",
+  "dugong",
+  "sea lion",
+  "Chihuahua",
+  "Japanese spaniel",
+  "Maltese dog",
+  "Pekinese",
+  "Shih-Tzu",
+  "Blenheim spaniel",
+  "papillon",
+  "toy terrier",
+  "Rhodesian ridgeback",
+  "Afghan hound",
+  "basset",
+  "beagle",
+  "bloodhound",
+  "bluetick",
+  "black-and-tan coonhound",
+  "Walker hound",
+  "English foxhound",
+  "redbone",
+  "borzoi",
+  "Irish wolfhound",
+  "Italian greyhound",
+  "whippet",
+  "Ibizan hound",
+  "Norwegian elkhound",
+  "otterhound",
+  "Saluki",
+  "Scottish deerhound",
+  "Weimaraner",
+  "Staffordshire bullterrier",
+  "American Staffordshire terrier",
+  "Bedlington terrier",
+  "Border terrier",
+  "Kerry blue terrier",
+  "Irish terrier",
+  "Norfolk terrier",
+  "Norwich terrier",
+  "Yorkshire terrier",
+  "wire-haired fox terrier",
+  "Lakeland terrier",
+  "Sealyham terrier",
+  "Airedale",
+  "cairn",
+  "Australian terrier",
+  "Dandie Dinmont",
+  "Boston bull",
+  "miniature schnauzer",
+  "giant schnauzer",
+  "standard schnauzer",
+  "Scotch terrier",
+  "Tibetan terrier",
+  "silky terrier",
+  "soft-coated wheaten terrier",
+  "West Highland white terrier",
+  "Lhasa",
+  "flat-coated retriever",
+  "curly-coated retriever",
+  "golden retriever",
+  "Labrador retriever",
+  "Chesapeake Bay retriever",
+  "German short-haired pointer",
+  "vizsla",
+  "English setter",
+  "Irish setter",
+  "Gordon setter",
+  "Brittany spaniel",
+  "clumber",
+  "English springer",
+  "Welsh springer spaniel",
+  "cocker spaniel",
+  "Sussex spaniel",
+  "Irish water spaniel",
+  "kuvasz",
+  "schipperke",
+  "groenendael",
+  "malinois",
+  "briard",
+  "kelpie",
+  "komondor",
+  "Old English sheepdog",
+  "Shetland sheepdog",
+  "collie",
+  "Border collie",
+  "Bouvier des Flandres",
+  "Rottweiler",
+  "German shepherd",
+  "Doberman",
+  "miniature pinscher",
+  "Greater Swiss Mountain dog",
+  "Bernese mountain dog",
+  "Appenzeller",
+  "EntleBucher",
+  "boxer",
+  "bull mastiff",
+  "Tibetan mastiff",
+  "French bulldog",
+  "Great Dane",
+  "Saint Bernard",
+  "Eskimo dog",
+  "malamute",
+  "Siberian husky",
+  "dalmatian",
+  "affenpinscher",
+  "basenji",
+  "pug",
+  "Leonberg",
+  "Newfoundland",
+  "Great Pyrenees",
+  "Samoyed",
+  "Pomeranian",
+  "chow",
+  "keeshond",
+  "Brabancon griffon",
+  "Pembroke",
+  "Cardigan",
+  "toy poodle",
+  "miniature poodle",
+  "standard poodle",
+  "Mexican hairless",
+  "timber wolf",
+  "white wolf",
+  "red wolf",
+  "coyote",
+  "dingo",
+  "dhole",
+  "African hunting dog",
+  "hyena",
+  "red fox",
+  "kit fox",
+  "Arctic fox",
+  "grey fox",
+  "tabby",
+  "tiger cat",
+  "Persian cat",
+  "Siamese cat",
+  "Egyptian cat",
+  "cougar",
+  "lynx",
+  "leopard",
+  "snow leopard",
+  "jaguar",
+  "lion",
+  "tiger",
+  "cheetah",
+  "brown bear",
+  "American black bear",
+  "ice bear",
+  "sloth bear",
+  "mongoose",
+  "meerkat",
+  "tiger beetle",
+  "ladybug",
+  "ground beetle",
+  "long-horned beetle",
+  "leaf beetle",
+  "dung beetle",
+  "rhinoceros beetle",
+  "weevil",
+  "fly",
+  "bee",
+  "ant",
+  "grasshopper",
+  "cricket",
+  "walking stick",
+  "cockroach",
+  "mantis",
+  "cicada",
+  "leafhopper",
+  "lacewing",
+  "dragonfly",
+  "damselfly",
+  "admiral",
+  "ringlet",
+  "monarch",
+  "cabbage butterfly",
+  "sulphur butterfly",
+  "lycaenid",
+  "starfish",
+  "sea urchin",
+  "sea cucumber",
+  "wood rabbit",
+  "hare",
+  "Angora",
+  "hamster",
+  "porcupine",
+  "fox squirrel",
+  "marmot",
+  "beaver",
+  "guinea pig",
+  "sorrel",
+  "zebra",
+  "hog",
+  "wild boar",
+  "warthog",
+  "hippopotamus",
+  "ox",
+  "water buffalo",
+  "bison",
+  "ram",
+  "bighorn",
+  "ibex",
+  "hartebeest",
+  "impala",
+  "gazelle",
+  "Arabian camel",
+  "llama",
+  "weasel",
+  "mink",
+  "polecat",
+  "black-footed ferret",
+  "otter",
+  "skunk",
+  "badger",
+  "armadillo",
+  "three-toed sloth",
+  "orangutan",
+  "gorilla",
+  "chimpanzee",
+  "gibbon",
+  "siamang",
+  "guenon",
+  "patas",
+  "baboon",
+  "macaque",
+  "langur",
+  "colobus",
+  "proboscis monkey",
+  "marmoset",
+  "capuchin",
+  "howler monkey",
+  "titi",
+  "spider monkey",
+  "squirrel monkey",
+  "Madagascar cat",
+  "indri",
+  "Indian elephant",
+  "African elephant",
+  "lesser panda",
+  "giant panda",
+  "barracouta",
+  "eel",
+  "coho",
+  "rock beauty",
+  "anemone fish",
+  "sturgeon",
+  "gar",
+  "lionfish",
+  "puffer",
+  "abacus",
+  "abaya",
+  "academic gown",
+  "accordion",
+  "acoustic guitar",
+  "aircraft carrier",
+  "airliner",
+  "airship",
+  "altar",
+  "ambulance",
+  "amphibian",
+  "analog clock",
+  "apiary",
+  "apron",
+  "ashcan",
+  "assault rifle",
+  "backpack",
+  "bakery",
+  "balance beam",
+  "balloon",
+  "ballpoint",
+  "Band Aid",
+  "banjo",
+  "bannister",
+  "barbell",
+  "barber chair",
+  "barbershop",
+  "barn",
+  "barometer",
+  "barrel",
+  "barrow",
+  "baseball",
+  "basketball",
+  "bassinet",
+  "bassoon",
+  "bathing cap",
+  "bath towel",
+  "bathtub",
+  "beach wagon",
+  "beacon",
+  "beaker",
+  "bearskin",
+  "beer bottle",
+  "beer glass",
+  "bell cote",
+  "bib",
+  "bicycle-built-for-two",
+  "bikini",
+  "binder",
+  "binoculars",
+  "birdhouse",
+  "boathouse",
+  "bobsled",
+  "bolo tie",
+  "bonnet",
+  "bookcase",
+  "bookshop",
+  "bottlecap",
+  "bow",
+  "bow tie",
+  "brass",
+  "brassiere",
+  "breakwater",
+  "breastplate",
+  "broom",
+  "bucket",
+  "buckle",
+  "bulletproof vest",
+  "bullet train",
+  "butcher shop",
+  "cab",
+  "caldron",
+  "candle",
+  "cannon",
+  "canoe",
+  "can opener",
+  "cardigan",
+  "car mirror",
+  "carousel",
+  "carpenter's kit",
+  "carton",
+  "car wheel",
+  "cash machine",
+  "cassette",
+  "cassette player",
+  "castle",
+  "catamaran",
+  "CD player",
+  "cello",
+  "cellular telephone",
+  "chain",
+  "chainlink fence",
+  "chain mail",
+  "chain saw",
+  "chest",
+  "chiffonier",
+  "chime",
+  "china cabinet",
+  "Christmas stocking",
+  "church",
+  "cinema",
+  "cleaver",
+  "cliff dwelling",
+  "cloak",
+  "clog",
+  "cocktail shaker",
+  "coffee mug",
+  "coffeepot",
+  "coil",
+  "combination lock",
+  "computer keyboard",
+  "confectionery",
+  "container ship",
+  "convertible",
+  "corkscrew",
+  "cornet",
+  "cowboy boot",
+  "cowboy hat",
+  "cradle",
+  "crane",
+  "crash helmet",
+  "crate",
+  "crib",
+  "Crock Pot",
+  "croquet ball",
+  "crutch",
+  "cuirass",
+  "dam",
+  "desk",
+  "desktop computer",
+  "dial telephone",
+  "diaper",
+  "digital clock",
+  "digital watch",
+  "dining table",
+  "dishrag",
+  "dishwasher",
+  "disk brake",
+  "dock",
+  "dogsled",
+  "dome",
+  "doormat",
+  "drilling platform",
+  "drum",
+  "drumstick",
+  "dumbbell",
+  "Dutch oven",
+  "electric fan",
+  "electric guitar",
+  "electric locomotive",
+  "entertainment center",
+  "envelope",
+  "espresso maker",
+  "face powder",
+  "feather boa",
+  "file",
+  "fireboat",
+  "fire engine",
+  "fire screen",
+  "flagpole",
+  "flute",
+  "folding chair",
+  "football helmet",
+  "forklift",
+  "fountain",
+  "fountain pen",
+  "four-poster",
+  "freight car",
+  "French horn",
+  "frying pan",
+  "fur coat",
+  "garbage truck",
+  "gasmask",
+  "gas pump",
+  "goblet",
+  "go-kart",
+  "golf ball",
+  "golfcart",
+  "gondola",
+  "gong",
+  "gown",
+  "grand piano",
+  "greenhouse",
+  "grille",
+  "grocery store",
+  "guillotine",
+  "hair slide",
+  "hair spray",
+  "half track",
+  "hammer",
+  "hamper",
+  "hand blower",
+  "hand-held computer",
+  "handkerchief",
+  "hard disc",
+  "harmonica",
+  "harp",
+  "harvester",
+  "hatchet",
+  "holster",
+  "home theater",
+  "honeycomb",
+  "hook",
+  "hoopskirt",
+  "horizontal bar",
+  "horse cart",
+  "hourglass",
+  "iPod",
+  "iron",
+  "jack-o'-lantern",
+  "jean",
+  "jeep",
+  "jersey",
+  "jigsaw puzzle",
+  "jinrikisha",
+  "joystick",
+  "kimono",
+  "knee pad",
+  "knot",
+  "lab coat",
+  "ladle",
+  "lampshade",
+  "laptop",
+  "lawn mower",
+  "lens cap",
+  "letter opener",
+  "library",
+  "lifeboat",
+  "lighter",
+  "limousine",
+  "liner",
+  "lipstick",
+  "Loafer",
+  "lotion",
+  "loudspeaker",
+  "loupe",
+  "lumbermill",
+  "magnetic compass",
+  "mailbag",
+  "mailbox",
+  "maillot",
+  "maillot tank suit",
+  "manhole cover",
+  "maraca",
+  "marimba",
+  "mask",
+  "matchstick",
+  "maypole",
+  "maze",
+  "measuring cup",
+  "medicine chest",
+  "megalith",
+  "microphone",
+  "microwave",
+  "military uniform",
+  "milk can",
+  "minibus",
+  "miniskirt",
+  "minivan",
+  "missile",
+  "mitten",
+  "mixing bowl",
+  "mobile home",
+  "Model T",
+  "modem",
+  "monastery",
+  "monitor",
+  "moped",
+  "mortar",
+  "mortarboard",
+  "mosque",
+  "mosquito net",
+  "motor scooter",
+  "mountain bike",
+  "mountain tent",
+  "mouse",
+  "mousetrap",
+  "moving van",
+  "muzzle",
+  "nail",
+  "neck brace",
+  "necklace",
+  "nipple",
+  "notebook",
+  "obelisk",
+  "oboe",
+  "ocarina",
+  "odometer",
+  "oil filter",
+  "organ",
+  "oscilloscope",
+  "overskirt",
+  "oxcart",
+  "oxygen mask",
+  "packet",
+  "paddle",
+  "paddlewheel",
+  "padlock",
+  "paintbrush",
+  "pajama",
+  "palace",
+  "panpipe",
+  "paper towel",
+  "parachute",
+  "parallel bars",
+  "park bench",
+  "parking meter",
+  "passenger car",
+  "patio",
+  "pay-phone",
+  "pedestal",
+  "pencil box",
+  "pencil sharpener",
+  "perfume",
+  "Petri dish",
+  "photocopier",
+  "pick",
+  "pickelhaube",
+  "picket fence",
+  "pickup",
+  "pier",
+  "piggy bank",
+  "pill bottle",
+  "pillow",
+  "ping-pong ball",
+  "pinwheel",
+  "pirate",
+  "pitcher",
+  "plane",
+  "planetarium",
+  "plastic bag",
+  "plate rack",
+  "plow",
+  "plunger",
+  "Polaroid camera",
+  "pole",
+  "police van",
+  "poncho",
+  "pool table",
+  "pop bottle",
+  "pot",
+  "potter's wheel",
+  "power drill",
+  "prayer rug",
+  "printer",
+  "prison",
+  "projectile",
+  "projector",
+  "puck",
+  "punching bag",
+  "purse",
+  "quill",
+  "quilt",
+  "racer",
+  "racket",
+  "radiator",
+  "radio",
+  "radio telescope",
+  "rain barrel",
+  "recreational vehicle",
+  "reel",
+  "reflex camera",
+  "refrigerator",
+  "remote control",
+  "restaurant",
+  "revolver",
+  "rifle",
+  "rocking chair",
+  "rotisserie",
+  "rubber eraser",
+  "rugby ball",
+  "rule",
+  "running shoe",
+  "safe",
+  "safety pin",
+  "saltshaker",
+  "sandal",
+  "sarong",
+  "sax",
+  "scabbard",
+  "scale",
+  "school bus",
+  "schooner",
+  "scoreboard",
+  "screen",
+  "screw",
+  "screwdriver",
+  "seat belt",
+  "sewing machine",
+  "shield",
+  "shoe shop",
+  "shoji",
+  "shopping basket",
+  "shopping cart",
+  "shovel",
+  "shower cap",
+  "shower curtain",
+  "ski",
+  "ski mask",
+  "sleeping bag",
+  "slide rule",
+  "sliding door",
+  "slot",
+  "snorkel",
+  "snowmobile",
+  "snowplow",
+  "soap dispenser",
+  "soccer ball",
+  "sock",
+  "solar dish",
+  "sombrero",
+  "soup bowl",
+  "space bar",
+  "space heater",
+  "space shuttle",
+  "spatula",
+  "speedboat",
+  "spider web",
+  "spindle",
+  "sports car",
+  "spotlight",
+  "stage",
+  "steam locomotive",
+  "steel arch bridge",
+  "steel drum",
+  "stethoscope",
+  "stole",
+  "stone wall",
+  "stopwatch",
+  "stove",
+  "strainer",
+  "streetcar",
+  "stretcher",
+  "studio couch",
+  "stupa",
+  "submarine",
+  "suit",
+  "sundial",
+  "sunglass",
+  "sunglasses",
+  "sunscreen",
+  "suspension bridge",
+  "swab",
+  "sweatshirt",
+  "swimming trunks",
+  "swing",
+  "switch",
+  "syringe",
+  "table lamp",
+  "tank",
+  "tape player",
+  "teapot",
+  "teddy",
+  "television",
+  "tennis ball",
+  "thatch",
+  "theater curtain",
+  "thimble",
+  "thresher",
+  "throne",
+  "tile roof",
+  "toaster",
+  "tobacco shop",
+  "toilet seat",
+  "torch",
+  "totem pole",
+  "tow truck",
+  "toyshop",
+  "tractor",
+  "trailer truck",
+  "tray",
+  "trench coat",
+  "tricycle",
+  "trimaran",
+  "tripod",
+  "triumphal arch",
+  "trolleybus",
+  "trombone",
+  "tub",
+  "turnstile",
+  "typewriter keyboard",
+  "umbrella",
+  "unicycle",
+  "upright",
+  "vacuum",
+  "vase",
+  "vault",
+  "velvet",
+  "vending machine",
+  "vestment",
+  "viaduct",
+  "violin",
+  "volleyball",
+  "waffle iron",
+  "wall clock",
+  "wallet",
+  "wardrobe",
+  "warplane",
+  "washbasin",
+  "washer",
+  "water bottle",
+  "water jug",
+  "water tower",
+  "whiskey jug",
+  "whistle",
+  "wig",
+  "window screen",
+  "window shade",
+  "Windsor tie",
+  "wine bottle",
+  "wing",
+  "wok",
+  "wooden spoon",
+  "wool",
+  "worm fence",
+  "wreck",
+  "yawl",
+  "yurt",
+  "web site",
+  "comic book",
+  "crossword puzzle",
+  "street sign",
+  "traffic light",
+  "book jacket",
+  "menu",
+  "plate",
+  "guacamole",
+  "consomme",
+  "hot pot",
+  "trifle",
+  "ice cream",
+  "ice lolly",
+  "French loaf",
+  "bagel",
+  "pretzel",
+  "cheeseburger",
+  "hotdog",
+  "mashed potato",
+  "head cabbage",
+  "broccoli",
+  "cauliflower",
+  "zucchini",
+  "spaghetti squash",
+  "acorn squash",
+  "butternut squash",
+  "cucumber",
+  "artichoke",
+  "bell pepper",
+  "cardoon",
+  "mushroom",
+  "Granny Smith",
+  "strawberry",
+  "orange",
+  "lemon",
+  "fig",
+  "pineapple",
+  "banana",
+  "jackfruit",
+  "custard apple",
+  "pomegranate",
+  "hay",
+  "carbonara",
+  "chocolate sauce",
+  "dough",
+  "meat loaf",
+  "pizza",
+  "potpie",
+  "burrito",
+  "red wine",
+  "espresso",
+  "cup",
+  "eggnog",
+  "alp",
+  "bubble",
+  "cliff",
+  "coral reef",
+  "geyser",
+  "lakeside",
+  "promontory",
+  "sandbar",
+  "seashore",
+  "valley",
+  "volcano",
+  "ballplayer",
+  "groom",
+  "scuba diver",
+  "rapeseed",
+  "daisy",
+  "yellow lady's slipper",
+  "corn",
+  "acorn",
+  "hip",
+  "buckeye",
+  "coral fungus",
+  "agaric",
+  "gyromitra",
+  "stinkhorn",
+  "earthstar",
+  "hen-of-the-woods",
+  "bolete",
+  "ear",
+  "toilet tissue"
+]

metadata.yaml

@@ -0,0 +1,66 @@
+version: 1.0.0
+task: vision-classification
+created_at: '2026-07-14T21:00:10Z'
+context:
+  model:
+    id: Arm/deit-tiny-int8-ethos-u-executorch
+    base_model_id: facebook/deit-tiny-patch16-224
+    profile: Arm-Optimized
+    weight_dtype: int8
+    quantization:
+      method: PTQ-static
+      weight_bits: 8
+      activation_bits: 8
+      symmetric: true
+      mode: static
+      weight_granularity: per-tensor
+    calibration:
+      dataset_name: ImageNet-1k
+      sample_count: 2000
+      selection: random
+    model_size_mb: 4.707
+    format: pte
+    filename: deit-tiny-int8-ethos-u-executorch.pte
+  target:
+    name: Alif DK-E8
+    hardware_class: Embedded NPU
+    cpu_architecture: armv8.1-m
+    cpu_model: Cortex-M55
+    cpu_core_count: 1
+    cpu_clock_ghz: 0.4
+    npu_model: Ethos-U85
+    npu_num_macs: 256
+    os: bare-metal
+  runtime:
+    name: executorch
+    execution_backend: ethos-u
+    config:
+      optimisations:
+      - CMSIS-NN
+      - Ethos-U-Driver
+    version: 1.1.0
+  dataset:
+    name: ImageNet-1k
+    sample_count: 50000
+    reference_url: https://www.image-net.org/
+    slice: val
+  benchmark:
+    batch_size: 1
+    input_resolution: 224x224
+    num_runs: 100
+    warmup_runs: 10
+performance:
+  end_to_end_latency_ms:
+    p50: 81.5947
+    p90: 81.5948
+  sram_peak_bytes: 2682173
+  flash_footprint_mib: 4.7069
+  frames_per_second: 12.2557
+  delegation_pct: 99.91
+  mac_utilization_pct: 15.2133
+  total_cycles: 32637876
+  active_cycles: 32626806
+  effective_cycles_ratio_pct: 99.9661
+accuracy:
+  top_1_accuracy_pct: 67.95
+  top_5_accuracy_pct: 88.47

predictions.json

@@ -0,0 +1,27 @@
+[
+  {
+    "class_index": 311,
+    "label": "grasshopper",
+    "probability": 0.3301173782135408
+  },
+  {
+    "class_index": 312,
+    "label": "cricket",
+    "probability": 0.21808608542863353
+  },
+  {
+    "class_index": 599,
+    "label": "honeycomb",
+    "probability": 0.011976920119533177
+  },
+  {
+    "class_index": 698,
+    "label": "palace",
+    "probability": 0.003887443093009725
+  },
+  {
+    "class_index": 913,
+    "label": "wreck",
+    "probability": 0.0034532177617875135
+  }
+]