python

服务器端

global_model = MyModel() for round in range(num_rounds): selected_clients = random.sample(clients, m) meta_grads = [] for client in selected_clients: # 客户端上传的是元梯度 meta_grad = client.local_meta_update(copy.copy(global_model)) meta_grads.append(meta_grad) # 聚合元梯度并更新全局模型 agg_grad = sum(meta_grads) / len(meta_grads) with torch.no_grad(): for param, grad in zip(global_model.parameters(), agg_grad): param -= outer_lr * grad

客户端方法

def local_meta_update(self, model): support_loader, query_loader = self.split_data() # 内循环：在支持集上更新模型 model_copy = copy.deepcopy(model) opt_inner = SGD(model_copy.parameters(), lr=inner_lr) for x, y in support_loader: loss = loss_fn(model_copy(x), y) opt_inner.zero_grad() loss.backward() opt_inner.step() # 外循环：在查询集上计算损失，并反向传播至初始模型（需要保持计算图） opt_outer = SGD(model.parameters(), lr=outer_lr) # 这里的model是初始模型 for x, y in query_loader: pred = model_copy(x) # 使用适应后的模型 loss = loss_fn(pred, y) # 获取元梯度（需要二阶导数，可通过create_graph=True实现） grads = torch.autograd.grad(loss, model.parameters(), create_graph=True) # 返回元梯度 return [g.detach() for g in grads]