边缘AI部署在资源受限环境运行模型前言我们有一个用户场景需要在没有网络的工厂环境中使用 AI。传统的云端 AI 方案完全不行必须在边缘设备上运行模型。经过几个月的探索我们成功将模型部署到了树莓派和工业电脑上。今天分享边缘 AI 部署的经验。一、边缘AI的特点1.1 边缘 vs 云端维度边缘部署云端部署延迟极低取决于网络隐私高数据不离开设备中数据上传云端成本一次性硬件成本按需付费网络依赖无必须有网络计算能力有限强大模型大小受限无限制1.2 边缘场景EDGE_SCENARIOS { iot: {device: 树莓派, ram: 1-4GB, suitable: 轻量模型}, industrial: {device: 工业PC, ram: 8-16GB, suitable: 中量模型}, mobile: {device: 手机, ram: 4-8GB, suitable: 量化模型}, embedded: {device: MCU, ram: 512KB-2MB, suitable: Tiny模型} }二、模型优化2.1 模型剪枝import torch.nn.utils.prune as prune class ModelPruner: def __init__(self, model): self.model model def prune_weights(self, amount: float 0.3): 权重剪枝 for name, module in self.model.named_modules(): if isinstance(module, torch.nn.Linear): prune.l1_unstructured(module, nameweight, amountamount) def remove_pruning(self): 移除剪枝重新参数化 for name, module in self.model.named_modules(): if isinstance(module, torch.nn.Linear): prune.remove(module, weight)2.2 模型量化class ModelQuantizer: def __init__(self): self.quantization_config { compute_dtype: torch.float16, weight_dtype: torch.qint8 } def quantize_dynamic(self, model): 动态量化 return torch.quantization.quantize_dynamic( model, {torch.nn.Linear, torch.nn.LSTM}, dtypetorch.qint8 ) def quantize_static(self, model, calibration_data): 静态量化 model.qconfig torch.quantization.get_default_qconfig(fbgemm) torch.quantization.prepare(model, inplaceTrue) # 校准 with torch.no_grad(): for data in calibration_data: model(data) torch.quantization.convert(model, inplaceTrue) return model三、推理框架3.1 ONNX Runtimeimport onnxruntime as ort class ONNXInference: def __init__(self, model_path: str): self.session ort.InferenceSession( model_path, providers[CPUExecutionProvider] ) def predict(self, input_data): 推理 input_name self.session.get_inputs()[0].name output_name self.session.get_outputs()[0].name result self.session.run( [output_name], {input_name: input_data} ) return result[0]3.2 TensorRTclass TensorRTInference: def __init__(self, engine_path: str): import tensorrt as trt logger trt.Logger(trt.Logger.WARNING) runtime trt.Runtime(logger) with open(engine_path, rb) as f: self.engine runtime.deserialize_cuda_engine(f.read()) self.context self.engine.create_execution_context() def predict(self, input_data, output_data): 推理 import pycuda.driver as cuda cuda.init() context cuda.Context() stream cuda.Stream() # 内存分配和拷贝 d_input cuda.mem_alloc(input_data.nbytes) d_output cuda.mem_alloc(output_data.nbytes) cuda.memcpy_htod_async(d_input, input_data, stream) # 执行 self.context.execute_async_v2( bindings[int(d_input), int(d_output)], stream_handlestream.handle ) cuda.memcpy_dtoh_async(output_data, d_output, stream) stream.synchronize() return output_data四、设备适配4.1 树莓派部署# requirements.txt for Raspberry Pi # torch2.0.0 # torchvision0.15.0 # onnxruntime1.15.0 class RaspberryPiDeployer: def optimize_for_pi(self, model): 为树莓派优化 # 使用 PyTorch Mobile model.eval() # 量化 model_quantized torch.quantization.quantize_dynamic( model, {torch.nn.Linear}, dtypetorch.qint8 ) return model_quantized def export_scripted(self, model, input_shape): 导出为 TorchScript traced torch.jit.trace(model, torch.randn(input_shape)) return traced4.2 工业设备部署class IndustrialDeployer: def deploy(self, model, device_type: str): 部署到工业设备 if device_type jetson_nano: return self._deploy_jetson(model) elif device_type jetson_xavier: return self._deploy_jetson(model, use_tensorrtTrue) elif device_type industrial_pc: return self._deploy_pc(model) def _deploy_jetson(self, model, use_tensorrtTrue): 部署到 Jetson if use_tensorrt: # 转换为 TensorRT return self._convert_to_tensorrt(model) else: # 使用 PyTorch Native return model.cuda()五、性能优化5.1 批处理优化class BatchOptimizer: def __init__(self, max_batch_size: int 8): self.max_batch_size max_batch_size self.pending_requests [] def add_request(self, data): 添加请求 self.pending_requests.append(data) if len(self.pending_requests) self.max_batch_size: return self._process_batch() return None def force_process(self): 强制处理 if self.pending_requests: return self._process_batch() return None def _process_batch(self): 批量处理 batch self.pending_requests[:self.max_batch_size] self.pending_requests self.pending_requests[self.max_batch_size:] return batch5.2 缓存优化class EdgeCache: def __init__(self, max_size_mb: int 100): self.max_size max_size_mb * 1024 * 1024 self.cache {} self.access_times {} def get(self, key): 获取缓存 if key in self.cache: self.access_times[key] datetime.now() return self.cache[key] return None def set(self, key, value): 设置缓存 size self._get_size(value) while self._get_total_size() size self.max_size: self._evict_lru() self.cache[key] value self.access_times[key] datetime.now()六、监控与维护6.1 边缘监控class EdgeMonitor: def __init__(self): self.metrics { cpu_usage: [], memory_usage: [], inference_count: 0, errors: [] } def record(self, metric_type: str, value): 记录指标 if metric_type in [cpu_usage, memory_usage]: self.metrics[metric_type].append({ value: value, timestamp: datetime.now() }) else: self.metrics[metric_type] value def get_health_report(self): 健康报告 return { cpu_avg: sum(m[value] for m in self.metrics[cpu_usage]) / len(self.metrics[cpu_usage]) if self.metrics[cpu_usage] else 0, memory_avg: sum(m[value] for m in self.metrics[memory_usage]) / len(self.metrics[memory_usage]) if self.metrics[memory_usage] else 0, total_inferences: self.metrics[inference_count], error_count: len(self.metrics[errors]) }6.2 OTA 更新class EdgeOTA: def __init__(self): self.update_server https://updates.example.com def check_update(self, current_version: str) - dict: 检查更新 import requests response requests.get( f{self.update_server}/check, params{version: current_version} ) return response.json() def download_update(self, model_id: str, progress_callbackNone): 下载更新 import requests response requests.get( f{self.update_server}/download/{model_id}, streamTrue ) total_size int(response.headers.get(content-length, 0)) downloaded 0 with open(/tmp/model_update.onnx, wb) as f: for chunk in response.iter_content(chunk_size8192): f.write(chunk) downloaded len(chunk) if progress_callback: progress_callback(downloaded / total_size) return /tmp/model_update.onnx七、最佳实践7.1 部署策略✅渐进更新先小范围测试再全量✅版本管理保持多个版本可回滚✅监控告警实时监控设备状态✅自动恢复异常时自动重启7.2 性能优化✅模型优化剪枝、量化、蒸馏✅批处理提高 GPU 利用率✅缓存减少重复计算✅异步非阻塞推理八、总结边缘 AI 让 AI 能力延伸到每一个角落。关键在于模型优化适配硬件限制推理框架选择合适的运行时性能优化榨干硬件性能运维监控确保稳定运行记住边缘不是将就而是必然。
