3rd_Anti-UAV_CVPR23 Ascend NPU 部署指南

3rd_Anti-UAV_CVPR23-ascend/
├── inference.py          # 推理测试脚本
├── log.txt               # 测试日志 (完整)
├── README.md             # 本文档
├── test_sample.pt        # 测试样本
├── test_sample_info.json # 测试样本信息
├── inference_result.json # 推理结果
└── precision_result.json # 精度测试结果

docker exec -it test-modelagent bash

source /usr/local/Ascend/ascend-toolkit/set_env.sh

pip install torch torch_npu -i https://pypi.huaweicloud.com/repository/pypi/simple/

cd /data/ysws/agentsp/5-19-1/3rd_Anti-UAV_CVPR23-ascend/

# 普通推理模式
python3 inference.py

# 首次运行会有模型加载和编译开销

cd /data/ysws/agentsp/5-19-1/3rd_Anti-UAV_CVPR23-ascend/

# 精度测试模式
python3 inference.py precision_test

============================================================
3rd_Anti-UAV_CVPR23 NPU Test Suite
Model: SiamFC-based UAV Tracking
Output: /data/ysws/agentsp/5-19-1/3rd_Anti-UAV_CVPR23-ascend
============================================================

Mode: PRECISION TEST (CPU vs NPU)

============================================================
3rd_Anti-UAV_CVPR23 NPU Inference Test
============================================================
Device: npu:0
Model: /data/ysws/agentsp/5-19-1/3rd_Anti-UAV_CVPR23/iic/3rd_Anti-UAV_CVPR23/pytorch_model.pt
Loading model state dict...
Loaded 30 state dict entries
Building SiamFC model...
Model built successfully
Input shape: torch.Size([1, 3, 255, 255])
Running inference...
Output shape: torch.Size([1, 256, 22, 22])
Inference time: 5.7031s

============================================================
Creating Test Samples
============================================================
Saved test sample: /data/ysws/agentsp/5-19-1/3rd_Anti-UAV_CVPR23-ascend/test_sample.pt
Shape: torch.Size([1, 3, 255, 255])
Seed: 42 (for reproducibility)

============================================================
3rd_Anti-UAV_CVPR23 Precision Test (CPU vs NPU)
============================================================
Device: npu:0
Loading model state dict...
Building CPU model...
Building NPU model...
Input shape: torch.Size([1, 3, 255, 255])
Input seed: 42 (for reproducibility)
Running inference on CPU...
CPU inference time: 0.3300s
CPU output shape: torch.Size([1, 256, 22, 22])
Running inference on NPU...
NPU inference time: 0.0017s
NPU output shape: torch.Size([1, 256, 22, 22])

============================================================
Precision Results
============================================================
CPU inference time: 0.3300s
NPU inference time: 0.0017s
Speedup: 191.39x
Max absolute error: 1.222447e-03
Max relative error: 5.782230e-04 (0.0578%)
Threshold: 1.0%
Status: PASS

============================================================
Test Complete!
============================================================

import torch
import torch.nn as nn
import torch.nn.functional as F

MODEL_PATH = "/data/ysws/agentsp/5-19-1/3rd_Anti-UAV_CVPR23/iic/3rd_Anti-UAV_CVPR23/pytorch_model.pt"

class SiamFCModel(nn.Module):
    def __init__(self, state_dict):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 96, 11, stride=2)
        self.bn1 = nn.BatchNorm2d(96)
        self.conv2 = nn.Conv2d(96, 256, 5, groups=2)
        self.bn2 = nn.BatchNorm2d(256)
        self.conv3 = nn.Conv2d(256, 384, 3)
        self.bn3 = nn.BatchNorm2d(384)
        self.conv4 = nn.Conv2d(384, 384, 3, groups=2)
        self.bn4 = nn.BatchNorm2d(384)
        self.conv5 = nn.Conv2d(384, 256, 3, groups=2)
        # Load weights from state_dict...

    def forward(self, x):
        x = torch.relu(self.bn1(self.conv1(x)))
        x = F.max_pool2d(x, 3, stride=2)
        x = torch.relu(self.bn2(self.conv2(x)))
        x = F.max_pool2d(x, 3, stride=2)
        x = torch.relu(self.bn3(self.conv3(x)))
        x = torch.relu(self.bn4(self.conv4(x)))
        x = self.conv5(x)
        return x

# Load and run
state_dict = torch.load(MODEL_PATH, map_location="cpu")
model = SiamFCModel(state_dict).to("npu:0")
model.eval()

test_input = torch.randn(1, 3, 255, 255).to("npu:0")
with torch.no_grad():
    output = model(test_input)

print(f"Output shape: {output.shape}")  # torch.Size([1, 256, 22, 22])

# 提取 exemplar 特征
exemplar = torch.randn(1, 3, 127, 127).to("npu:0")  # 127x127 exemplar
exemplar_feat = model(exemplar)  # [1, 256, 6, 6]

# 提取 search 区域特征
search = torch.randn(1, 3, 255, 255).to("npu:0")  # 255x255 search
search_feat = model(search)  # [1, 256, 22, 22]

# 交叉相关计算响应图
response = F.conv2d(search_feat, exemplar_feat)
print(f"Response shape: {response.shape}")  # [1, 1, H, W]