Renesas/EfficientNetV2-B0-ONNX

EfficientNetV2-B0 (ONNX) – Renesas X5H

Introduction

This repository hosts EfficientNetV2-B0 in ONNX FP32 format, targeting the Renesas R-Car X5H platform for image classification inference on the NPX6 NPU.

• Model Architecture: EfficientNetV2-B0 — a parameter-efficient convolutional neural network optimized for high accuracy and efficient inference.
• Source Framework: MMPreTrain (OpenMMLab)
• Export Format: ONNX FP32
• Graph Optimization: ONNX Optimizer
• Task: Image Classification (ImageNet ILSVRC2012, 1000 classes)
• Parameters: 7.1M (7.11M counted from fp32/efficientnetv2-b0_3rdparty_in1k_sim_opt.onnx initializers)

### Deployment Flow

```text
EfficientNetV2-B0 (MMPreTrain)
            │
            ▼
       ONNX Export
            │
            ▼
      ONNX Optimizer
            │
            ▼
efficientnetv2-b0_3rdparty_in1k_sim_opt.onnx (FP32)
            │
            ├─▶ ONNX Runtime (Custom NPU EP) ─▶ INT8 auto-cast ─▶ NPX6 NPU
            │
            └─▶ MWMX Runtime                 ─▶ INT8 auto-cast ─▶ NPX6 NPU

Provided Artifacts

| Artifact | Status | Notes |
|:---------:|:------:|--------|
| **FP32 ONNX** | ✅ Provided | Exported from MMPreTrain and optimized using ONNX Optimizer |

Runtime Details

### ONNX Model

- **Training Framework:** MMPreTrain (OpenMMLab)
- **Export Format:** ONNX FP32
- **Graph Optimizations:** ONNX Optimizer passes applied
- **Input Resolution:** 224 × 224
- **Output:** 1000 ImageNet classes

Download

huggingface-cli download Renesas/EfficientNetV2-B0-ONNX fp32/efficientnetv2_b0.onnx

Performance

Benchmark YAML files live under each artifact's benchmarks/ folder. For B0 (int8/benchmarks/), there are now two independent measurement sources on the same X5H hardware:

Source	Runtime	Slice	Latency (ms)	FPS	Type
x5h_mwmx_npu.yaml	MWMX	1× NPU · 1 Core · 850 MHz	2.0818	480	Measured (hil)
x5h_ppa_npu.yaml	PPA Estimator	1× NPU · 1 Core · 850 MHz	1.517	650	Estimation
x5h_mwmx_npu_apm50_1core.yaml	MWMX (APM50 CI)	1× NPU · 1 Core · 850 MHz	1.739605	—	Measured (hil)
x5h_mwmx_npu_apm50_12core.yaml	MWMX (APM50 CI)	1× NPU · 12 Cores · 850 MHz	2.417156	—	Measured (hil)

Note: the apm50 files are a second, independent benchmark run captured via the internal "APM50" CI pipeline (metawaremx_runtime), reported separately from the original x5h_mwmx_npu.yaml / x5h_ppa_npu.yaml measurements above rather than replacing them.

Additional Sizes (B1–B3)

This repo is named for B0 but now also hosts B1, B2, and B3 as additional sibling artifacts (b1/, b2/, b3/, each with their own .metadata.yaml and benchmarks/), measured via the same APM50 / metawaremx_runtime MWMX CI pipeline on Renesas R-Car X5H.

Model	Checkpoint	Parameters	Slice: 1 Core (ms)	Slice: 12 Cores (ms)
EfficientNetV2-B1	efficientnetv2_b1_3rdparty_in1k	8.1M	2.897629	3.27878
EfficientNetV2-B2	efficientnetv2_b2_3rdparty_in1k	10.1M	4.379195	4.300261
EfficientNetV2-B3	efficientnetv2_b3_3rdparty_in1k	14.4M	7.498649	6.245134

This repo is named for B0 but now also hosts B1–B3 as additional artifacts; consider renaming to EfficientNetV2-ONNX in a future cleanup.