cv_resnet50_live-category

1. 简介

本文档记录 iic/cv_resnet50_live-category（ResNet-50 直播分类模型）在昇腾 NPU（Ascend910）环境的快速部署与验证结果。

ResNet-50 直播场景图像分类模型，基于 PyTorch torchvision 框架。

2. 验证环境

组件	版本
`torch`	`2.5.1`
`torch_npu`	`2.5.1`
`torchvision`	`>=0.20.0`
`CANN`	`8.5.RC1`

NPU：Ascend910（单卡）
输入尺寸：224x224
推理框架：PyTorch + torchvision

3. 快速部署

3.1 环境准备

pip install torch torchvision pillow

3.2 推理代码

import torch
import torchvision.models as models
from PIL import Image
from torchvision import transforms

device = torch.device("npu:0" if torch.npu.is_available() else "cpu")

ckpt = torch.load("pytorch_model.pt", map_location="cpu")
num_classes = ckpt['fc.weight'].shape[0] if 'fc.weight' in ckpt else 1000
model = models.resnet50(weights=None, num_classes=num_classes)
model.load_state_dict(ckpt, strict=False)
model = model.to(device).eval()

transform = transforms.Compose([
    transforms.Resize(256), transforms.CenterCrop(224),
    transforms.ToTensor(), transforms.Normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]),
])
img = Image.new("RGB", (224, 224), (128, 128, 128))
input_tensor = transform(img).unsqueeze(0).to(device)

with torch.no_grad():
    output = model(input_tensor)
pred = output.argmax(-1).item()
print(f"Predicted class: {pred}")

4. Smoke 验证

python3 inference.py

5. 性能参考

指标	数值
平均推理时间	`4.81ms`
测试次数	`50`

6. 精度评测

指标	数值
Top-1 一致性	`4/4`
Top-5 一致性	`4/4`
最大 logits 相对误差	`0.09%`
平均 KL 散度	`0.000003`
结论	`PASS`

cv_resnet50_live-category

1. 简介

本文档记录 iic/cv_resnet50_live-category（ResNet-50 直播分类模型）在昇腾 NPU（Ascend910）环境的快速部署与验证结果。

ResNet-50 直播场景图像分类模型，基于 PyTorch torchvision 框架。

2. 验证环境

组件	版本
`torch`	`2.5.1`
`torch_npu`	`2.5.1`
`torchvision`	`>=0.20.0`
`CANN`	`8.5.RC1`

NPU：Ascend910（单卡）
输入尺寸：224x224
推理框架：PyTorch + torchvision

3. 快速部署

3.1 环境准备

pip install torch torchvision pillow

3.2 推理代码

import torch
import torchvision.models as models
from PIL import Image
from torchvision import transforms

device = torch.device("npu:0" if torch.npu.is_available() else "cpu")

ckpt = torch.load("pytorch_model.pt", map_location="cpu")
num_classes = ckpt['fc.weight'].shape[0] if 'fc.weight' in ckpt else 1000
model = models.resnet50(weights=None, num_classes=num_classes)
model.load_state_dict(ckpt, strict=False)
model = model.to(device).eval()

transform = transforms.Compose([
    transforms.Resize(256), transforms.CenterCrop(224),
    transforms.ToTensor(), transforms.Normalize(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]),
])
img = Image.new("RGB", (224, 224), (128, 128, 128))
input_tensor = transform(img).unsqueeze(0).to(device)

with torch.no_grad():
    output = model(input_tensor)
pred = output.argmax(-1).item()
print(f"Predicted class: {pred}")

4. Smoke 验证

python3 inference.py

5. 性能参考

指标	数值
平均推理时间	`4.81ms`
测试次数	`50`

6. 精度评测

指标	数值
Top-1 一致性	`4/4`
Top-5 一致性	`4/4`
最大 logits 相对误差	`0.09%`
平均 KL 散度	`0.000003`
结论	`PASS`