Edgenext Small Int8 Imagenet

--- license: mit library_name: onnx base_model: timm/edgenext_small.usi_in1k base_model_relation: quantized pipeline_tag: image-classification datasets: - imagenet-1k tags: - onnx - int8 - quantized - image-classification - imagenet - edge - cpu --- # EdgeNeXt-S: INT8 ONNX, 81.5% ImageNet in 7.2 MB EdgeNeXt-S ([timm/edgenext_small.usi_in1k](https://huggingface.co/timm/edgenext_small.usi_in1k), MIT) at 320px, quantized to INT8 with [Kenosis](https://coreepoch.dev/kenosis/): 128-image calibration, no retraining. 81.53% top-1 from a single 7.2 MB file, measured on CPU with ONNX Runtime and OpenVINO. No accelerator required. ## Accuracy ImageNet-1K validation, 49,872 images (disjoint from the 128 calibration images), top-1. | model | ONNX Runtime | OpenVINO | size | |---|---|---|---| | FP32 baseline | 81.565% | 81.565% | 22.5 MB | | **Kenosis INT8** | **81.535%** | **81.483%** | **7.16 MB** | ## Speed Against the FP32 model on CPU: 32% faster on ONNX Runtime with one thread. ## Run ```python from huggingface_hub import hf_hub_download import numpy as np, onnxruntime as ort from PIL import Image config = hf_hub_download("CoreEpoch/edgenext-small-int8-imagenet", "config.json") path = hf_hub_download("CoreEpoch/edgenext-small-int8-imagenet", "edgenext_s_320_int8_kenosis.onnx") sess = ort.InferenceSession(path, providers=["CPUExecutionProvider"]) img = Image.open("your_image.jpg").convert("RGB") scale = 320 / min(img.size) # shorter side to 320, center crop: measured transform img = img.resize((round(img.width*scale), round(img.height*scale)), Image.BICUBIC) l, t = (img.width-320)//2, (img.height-320)//2 img = img.crop((l, t, l+320, t+320)) x = (np.asarray(img, np.float32)/255.0 - [0.485,0.456,0.406]) / [0.229,0.224,0.225] logits = sess.run(None, {"input": x.transpose(2,0,1)[None].astype(np.float32)})[0] print(int(np.argmax(logits))) ``` Input `1x3x320x320`, RGB, `/255`, ImageNet mean/std. Output `logits [1,1000]`, sorted-synset order. Run this file on a CPU. On OpenVINO's GPU plugin it returns wrong predictions. ## Integrity & license `edgenext_s_320_int8_kenosis.onnx` (7,160,419 B), SHA-256 `D1538E0F392836832E7BBC0C80285396491444B2614E07BC04EE26D38DE7B88C`. MIT (base EdgeNeXt, Maaz et al. 2022, retained). The license covers these model files only; it grants no rights in Kenosis, Core Epoch's quantizer, which remains proprietary. For licensing inqu...

--- license: mit library_name: onnx base_model: timm/edgenext_small.usi_in1k base_model_relation: quantized pipeline_tag: image-classification datasets: - imagenet-1k tags: - onnx - int8 - quantized - image-classification - imagenet - edge - cpu - raspberry-pi --- # EdgeNeXt-S — INT8 ONNX, 81.5% ImageNet in 7.2 MB This is EdgeNeXt-S ([timm/edgenext_small.usi_in1k](https://huggingface.co/timm/edgenext_small.usi_in1k), MIT) evaluated at 320px, quantized to INT8 with [Kenosis](https://coreepoch.dev), Core Epoch's post-training quantizer, using 128 calibration images and no retraining. It scores **81.53% top-1 on ImageNet-1K in a 7.2 MB single file, running 29% faster than its FP32 baseline single-thread**. The same file runs on both ONNX Runtime and OpenVINO, offline, with no GPU required. ## Measured accuracy ImageNet-1K validation, 49,872 images (the full 50,000 minus a 128-image hold-out; calibration used training-split images, never validation data). The evaluation images are the official validation files, verified by byte-level sample audit against the official distribution. All rows were evaluated on identical inputs. | model | top-1 | Δ vs FP32 | file size | |---|---|---|---| | FP32 baseline | 81.565% | — | 22.5 MB | | **this artifact** | **81.535%** | **−0.030** | **7.16 MB** | | ONNX Runtime static quantization at default calibration, same model, same 128 calibration images | 33.74% | −47.8 | 6.4 MB | The FP32 control matches the published checkpoint within 0.01. The third row is ONNX Runtime's own static quantizer on identical inputs at its default MinMax calibration; the exact invocation is in this repo as `quantize_baseline.py`, so the comparison can be re-run. ONNX Runtime also offers Percentile and Entropy calibrators, which are not applied here; choosing and tuning a calibrator per model recovers much of this gap. The artifact above uses no such tuning. Runs 29% faster than FP32 single-thread on ONNX Runtime (i5-13420H). The latency method, the ARM64 portability check, and the OpenVINO cross-check are in the [int8-models repo](https://github.com/CoreEpoch/int8-models). ## Run it ``` pip install onnxruntime numpy pillow huggingface_hub ``` ```python from huggingface_hub import hf_hub_download import numpy as np import onnxruntime as ort from PIL import Image path = hf_hub_download("CoreEpoch/edgenext-small-int8-imagenet", "edgenext_s...

Source context: 51 downloads · 0 likes · Pipeline image-classification · Library onnx · Repo CoreEpoch/edgenext-small-int8-imagenet