DQLiu/rrferrglgf-NLP

MobileNetV2

MobileNetV2 是在分辨率 224x224 的 ImageNet-1k 数据集上预训练的图像分类模型。https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet

我们开发了一个 AMD Ryzen AI (https://ryzenai.docs.amd.com/en/latest/) 支持的修改版本 - https://huggingface.co/models?search=mobilenet_v2

详情可参照 https://huggingface.co/amd/mobilenet_v2_1.0_224

安装

按照 https://ryzenai.docs.amd.com/en/latest/inst.html 为 Ryzen AI 准备环境。

运行以下脚本安装pre-requisites包

pip install -r requirements.txt

推理

推理一张图片

import sys
import onnxruntime
import torch
import torchvision.transforms as transforms 
from PIL import Image

image_path = sys.argv[1]
onnx_model = sys.argv[2]

normalize = transforms.Normalize(
  mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
img_transformer = transforms.Compose([
                transforms.Resize(256),
                transforms.CenterCrop(224),
                transforms.ToTensor(),  
                normalize])
img_tensor = img_transformer(Image.open(image_path)).unsqueeze(0)
img_tensor = torch.permute(img_tensor, (0, 2, 3, 1))
so = onnxruntime.SessionOptions()
ort_session = onnxruntime.InferenceSession(
        onnx_model, so, 
        providers=['CPUExecutionProvider'], 
        provider_options=None)
input = img_tensor.numpy()
ort_input = {ort_session.get_inputs()[0].name: input}

output = ort_session.run(None, ort_input)
top5_probabilities, top5_class_indices = torch.topk(torch.nn.functional.softmax(torch.tensor(output[0])), k=5)

vaip_config.json 来自于Ryzen AI 安装包 https://ryzenai.docs.amd.com/en/latest/inst.html, voe-4.0-win_amd64 of ryzen-ai-sw-1.0.zip.

使用 eval_onnx.py 推理 ImageNet validation dataset (50,000 Images).

测试量化模型CPU

python eval_onnx.py  --onnx_model=./mobilenetv2_int8.onnx --data_dir=./{DATA_PATH}

测试量化模型IPU

python eval_onnx.py  --onnx_model=./mobilenetv2_int8.onnx --data_dir=./{DATA_PATH} --ipu --provider_config Path\To\vaip_config.json

Metric	Accuracy on IPU
top1& top5 accuracy	75.62% / 92.52%

DATA_PATH: ImageNet validation目录

引用

@article{MobileNet v2,
  author       = {Mark Sandler and
                  Andrew G. Howard and
                  Menglong Zhu and
                  Andrey Zhmoginov and
                  Liang{-}Chieh Chen},
  title        = {MobileNetV2: Inverted Residuals and Linear Bottlenecks},
  year         = {2018},
  url          = {http://arxiv.org/abs/1801.04381},
}