Initial upload of DeepLabV3Plus-R50-ONNX

.metadata.yaml

@@ -0,0 +1,46 @@
+# family is `deeplab` with version `v3+`, not `deeplabv3plus`: app.js familyOf()
+# rendered the old slug as "Deeplabv3plus".
+# PARTIAL MODEL. The source checkpoint name ends in
+# `custom_seg_split_4_split_2` — this artifact is ONE segment of a 4-way split
+# network, not the full end-to-end model. Latency below reflects only that segment;
+# do not quote it as whole-model latency.
+# PARAMETERS DELIBERATELY ABSENT for exactly that reason: the commonly-cited ~41M
+# is the FULL DeepLabV3+/ResNet50 model, so attaching it to one of four segments
+# would overstate this artifact several-fold. Count this segment from the graph.
+
+model:
+  name: deeplabv3plus-r50
+  display_name: DeepLabV3Plus-R50
+  # upstream: intentionally absent — these weights have no HuggingFace repo.
+  #            See the `source` block below. Never write "TBD" here: the
+  #            generator copies it into base_model and renders it as the
+  #            model's architecture label in the catalog.
+  source:
+    kind: openmmlab
+    id: deeplabv3plus_r50_d8_4xb2_80k_cityscapes_512x1024
+    url: https://github.com/open-mmlab/mmsegmentation/blob/main/configs/deeplabv3plus/metafile.yaml
+
+architecture:
+  family: deeplab   # lineage only — no version, no size
+  version: "v3+"
+  backbone: resnet50-d8
+  dataset: cityscapes
+  input_resolution: 512x1024
+  num_classes: 19
+  modality:
+    - vision
+  # parameters: intentionally absent — see the note above. Fill it with the
+  #             exact count instead of an estimate:
+  #   import onnx; from onnx import numpy_helper
+  #   m = onnx.load('model.onnx')
+  #   sum(numpy_helper.to_array(t).size for t in m.graph.initializer)
+  parameters_source: unknown
+
+format:
+  type: onnx
+  # opset: read it off the graph rather than guessing —
+  #   onnx.load('model.onnx').opset_import[0].version
+  # NOTE: `version` is a GGUF-only field and must not be used for ONNX.
+
+tasks:
+  - semantic-segmentation

README.md

@@ -0,0 +1,105 @@
+---
+license: apache-2.0
+base_model: []
+pipeline_tag: image-segmentation
+tags:
+- semantic-segmentation
+- computer-vision
+- renesas
+- x5h
+- onnx
+- deeplabv3plus
+- resnet50
+- cityscapes
+---
+
+# DeepLabV3Plus-R50 (ONNX) – Renesas X5H
+
+> ⏳ **Model file not yet uploaded.** Benchmark results on this page were published ahead of the model weights — see **Provided Artifacts** below. Download/deployment steps will not work until the file is added to this repository.
+
+> **⚠️ Partial-model caveat.** The source checkpoint name ends in
+> `custom_seg_split_4_split_2` — this artifact is **one segment of a 4-way split network**, not
+> the full end-to-end DeepLabV3+ model. The latency below reflects only that segment; do not
+> quote it as whole-model latency until the other splits are accounted for.
+
+## Introduction
+
+This repository hosts **DeepLabV3+ (ResNet50-D8 backbone)** targeting the **Renesas R-Car X5H**
+platform for semantic segmentation inference on the NPX6 NPU.
+
+- **Model Architecture:** DeepLabV3+ with ResNet50-D8 backbone
+- **Source Model:** OpenMMLab config [`deeplabv3plus_r50_d8_4xb2_80k_cityscapes_512x1024`](https://github.com/open-mmlab/mmsegmentation/blob/main/configs/deeplabv3plus/metafile.yaml)
+  *(no HuggingFace mirror of these weights; see `model.source` in `.metadata.yaml`)*
+- **Task:** Semantic Segmentation
+- **Dataset:** Cityscapes (inferred from checkpoint name)
+- **Input Resolution:** 512 × 1024 (explicit in checkpoint name)
+- **Parameters:** not published — count them from the ONNX graph (`sum(numpy_helper.to_array(t).size for t in model.graph.initializer)`)
+
+## Deployment Flow
+
+The FP32 ONNX model is auto-cast to **INT8** by the Renesas MWMX toolchain at compile time — no
+separate quantization step is required.
+
+```
+deeplabv3plus_..._split_2.onnx (FP32)
+        │
+        └─▶  MWMX Runtime  ──▶  INT8 auto-cast  ──▶  NPX6 NPU
+```
+
+## Provided Artifacts
+
+| Artifact | Status | Notes |
+|----------|--------|-------|
+| **FP32 (ONNX)** | ⏳ Not yet uploaded | Benchmark numbers below exist; the model file has not been published to this repo yet |
+
+## Performance
+
+Measured on **Renesas R-Car X5H** via the MWMX runtime (APM80 ship-performance CI pipeline).
+
+> **Benchmark configuration:** Single NPU · Batch size: 1 · Input: 3 × 512 × 1024
+>
+> The 1-AI-core slice failed to compile in the source CI pipeline, so only the 12-core result
+> is available. Latency is for the `split_2` segment only (see caveat above).
+
+| Runtime | Precision | Device | Latency (ms) | Type |
+|---------|-----------|--------|---------------|------|
+| MWMX Runtime | INT8 (auto) | X5H · 1× NPU · 12 Cores · 850 MHz | 38.104 | Measured |
+
+### Accuracy (mIoU)
+
+TBD — not yet measured/published for this repo.
+
+---
+
+## Runtime Details
+
+### MWMX Runtime
+
+- **Engine:** Renesas MWMX (Middleware MX) native inference runtime
+- **Input format:** FP32 ONNX (compiled by the MWMX toolchain)
+- **NPU execution precision:** INT8 (auto-cast by MWMX toolchain)
+- **Execution target:** NPX6-48K NPU on R-Car X5H
+
+---
+
+## Prerequisites
+
+To run inference on Renesas R-Car X5H, you need:
+
+1. **Renesas R-Car X5H board** with NPX6 NPU
+2. **Renesas MWMX Runtime**
+3. **Hugging Face CLI** to download the model (once the model file is published)
+
+## Download
+
+TBD — model file not yet published to this repository.
+
+---
+
+## Benchmark Methodology
+
+- **HIL runs:** Hardware-in-the-loop — measured on physical R-Car X5H silicon via the MWMX
+  runtime (`metawaremx_runtime` CI pipeline, "APM80" ship-performance target)
+- **Precision:** FP32 ONNX input; INT8 execution (auto-cast by MWMX)
+- **Slices:** only the 12-AI-core result is available (1-core compile failed)
+- **Scope:** this artifact is a single segment (`split_2` of 4) of the full segmentation pipeline

int8/.metadata.yaml

@@ -0,0 +1,14 @@
+variant:
+  id: onnx_int8
+  format: onnx
+  precision: int8            # INT8 quantized — auto-cast by NPU execution provider at runtime
+  method: onnx_export        # base model exported to ONNX; quantization applied by runtime
+
+quantization:
+  datatype: int8
+  scope:
+    - weights
+    - activations
+  granularity: per-tensor    # typical for runtime-cast INT8
+  calibration: ptq           # post-training quantization applied by the NPU EP / MWMX toolchain
+  toolchain: onnx

int8/benchmarks/x5h_mwmx_npu_apm80_12core.yaml

@@ -0,0 +1,44 @@
+hardware:
+  vendor: renesas
+  chip: rcar-x5h
+  cpu: arm-cortex-a720
+  npu: npx6-48k
+  npu_count: 2
+  npu_cores: 12
+  npu_default_freq_mhz: 1066
+  accelerator:
+    - npu
+
+runtime:
+  engine: mwmx
+  toolchain_version: "MWMX SDK v4.35.0"
+  format: onnx
+  execution_provider: npu
+  execution_precision: int8
+
+configuration:
+  npu_instances: 1
+  npu_cores_per_instance: 12
+  npu_freq_mhz: 850
+
+benchmark:
+  type: hil
+  parameters:
+    batch_size: 1
+    input_resolution: [1, 3, 512, 1024]   # explicit in the source checkpoint name (512x1024, Cityscapes)
+
+performance:
+  fps: null
+  latency: 38.103941        # APM80 pipeline run (metawaremx_runtime CI); this artifact is ONE segment
+  # of a 4-way split model ("custom_seg_split_4_split_2") — not full end-to-end latency.
+  # 1-core run for this artifact failed to compile — no figure available for that slice.
+
+metrics:
+  accuracy: null
+  top5_accuracy: null
+
+memory:
+  peak_mb: null
+
+power:
+  avg_w: null