import numpy as np import torch from tqdm import tqdm from sklearn.metrics import roc_auc_score, accuracy_score import copy import torch.nn.functional as F import random def train_deepmodel(device, model, loader, optimizer, criterion, epoch, model_name): model.train() running_loss = 0.0 corrects = 0.0 alpha = 1 beta = 0.1 for inputs, labels in tqdm(loader, desc=f'Training {model_name} Epoch {epoch + 1}', unit='batch'): inputs, labels = inputs.float().to(device), labels.to(device) # 确保数据在 device 上 optimizer.zero_grad() outputs, re_img = model(inputs) loss = criterion(outputs.squeeze(), labels.float()) loss_F1 = F.l1_loss(re_img, inputs) loss = alpha * loss + beta * loss_F1 loss.backward() optimizer.step() running_loss += loss.item() avg_loss = running_loss / len(loader) print(f'{model_name} Training Loss: {avg_loss:.4f}') return avg_loss def validate_deepmodel(device, model, loader, criterion, epoch, model_name): model.eval() running_loss = 0.0 correct, total = 0, 0 all_labels, all_preds = [], [] val_corrects = 0.0 alpha = 1 beta = 0.1 with torch.no_grad(): for inputs, labels in tqdm(loader, desc=f'Validating {model_name} Epoch {epoch + 1}', unit='batch'): inputs, labels = inputs.float().to(device), labels.to(device) # 确保数据在 device 上 outputs, re_img = model(inputs) # 将 logits 转换为预测 predicted = torch.sigmoid(outputs).data all_preds.extend(predicted.cpu().numpy()) all_labels.extend(labels.cpu().numpy()) # loss = criterion(outputs.squeeze(), labels.float()) loss = criterion(outputs.squeeze(), labels.float()) loss_F1 = F.l1_loss(re_img, inputs) loss = alpha * loss + beta * loss_F1 running_loss += loss.item() auc = roc_auc_score(all_labels, all_preds) predicted_labels = (np.array(all_preds) >= 0.5).astype(int) # 确保转换为 NumPy 数组 acc = accuracy_score(all_labels, predicted_labels) avg_loss = running_loss / len(loader) print(f'{model_name} Validation Loss: {avg_loss:.4f}, Accuracy: {acc:.4f}, AUC: {auc:.4f}') return avg_loss, acc, auc def test_deepmodel(device, model, loader): model.eval() all_labels, all_preds = [], [] with torch.no_grad(): for inputs, labels in tqdm(loader, desc=f'Testing', unit='batch'): inputs, labels = inputs.float().to(device), labels.to(device) # 确保数据在 device 上 outputs, re_img = model(inputs) predicted = torch.sigmoid(outputs).data # 将 logits 转换为预测 # 收集预测值和真实标签 all_preds.extend(predicted.cpu().numpy()) all_labels.extend(labels.cpu().numpy()) # 将预测值转换为二值标签 predicted_labels = (np.array(all_preds) >= 0.5).astype(int) # 计算准确率和AUC acc = accuracy_score(all_labels, predicted_labels) auc = roc_auc_score(all_labels, all_preds) print(f'Test Accuracy: {acc:.4f}, Test AUC: {auc:.4f}') return acc, auc # def train_model(device, model, loader, optimizer, criterion, epoch, model_name): # model.train() # running_loss = 0.0 # for i, (inputs, labels) in enumerate(tqdm(loader, desc=f'Training {model_name} Epoch {epoch + 1}', unit='batch')): # inputs, labels = inputs.float().to(device), labels.float().to(device) # 确保数据格式正确 # optimizer.zero_grad() # # outputs = model(inputs) # loss = criterion(outputs.squeeze(), labels) # # # 随机打印部分输出和标签,检查格式 # if i % 10 == 0: # 每100个批次打印一次 # print(f"Batch {i} - Sample Output: {outputs[0].item():.4f}, Sample Label: {labels[0].item()}") # # # 检查损失值是否异常 # if loss.item() < 0: # print(f"Warning: Negative loss detected at batch {i}. Loss: {loss.item()}") # # loss.backward() # optimizer.step() # # running_loss += loss.item() # # avg_loss = running_loss / len(loader) # print(f'{model_name} Training Loss: {avg_loss:.4f}') # return avg_loss def train_model(device, model, loader, optimizer, criterion, epoch, model_name): model.train() running_loss = 0.0 for inputs, labels in tqdm(loader, desc=f'Training {model_name} Epoch {epoch + 1}', unit='batch'): inputs, labels = inputs.float().to(device), labels.to(device) # 确保数据在 device 上 optimizer.zero_grad() outputs = model(inputs) loss = criterion(outputs.squeeze(), labels.float()) loss.backward() optimizer.step() running_loss += loss.item() avg_loss = running_loss / len(loader) print(f'{model_name} Training Loss: {avg_loss:.4f}') return avg_loss def validate_model(device, model, loader, criterion, epoch, model_name): model.eval() running_loss = 0.0 correct, total = 0, 0 all_labels, all_preds = [], [] with torch.no_grad(): for inputs, labels in tqdm(loader, desc=f'Validating {model_name} Epoch {epoch + 1}', unit='batch'): inputs, labels = inputs.float().to(device), labels.to(device) # 确保数据在 device 上 outputs = model(inputs) # 将 logits 转换为预测 predicted = torch.sigmoid(outputs).data all_preds.extend(predicted.cpu().numpy()) all_labels.extend(labels.cpu().numpy()) # loss = criterion(outputs.squeeze(), labels.float()) loss = criterion(outputs.squeeze(), labels.float()) running_loss += loss.item() auc = roc_auc_score(all_labels, all_preds) predicted_labels = (np.array(all_preds) >= 0.5).astype(int) # 确保转换为 NumPy 数组 acc = accuracy_score(all_labels, predicted_labels) avg_loss = running_loss / len(loader) print(f'{model_name} Validation Loss: {avg_loss:.4f}, Accuracy: {acc:.4f}, AUC: {auc:.4f}') return avg_loss, acc, auc # 权重聚合函数 def aggregate_weights(weights_list, alpha=1 / 3, beta=1 / 3, gamma=1 / 3): new_state_dict = copy.deepcopy(weights_list[0]) # 从模型a复制权重结构 for key in new_state_dict.keys(): new_state_dict[key] = (alpha * weights_list[0][key] + beta * weights_list[1][key] + gamma * weights_list[2][key]) return new_state_dict def v3_update_model_weights( epoch, model_to_update, other_models, global_model, losses, val_loader, device, val_auc_threshold, # 当前需要更新模型的验证 AUC 阈值 validate_model, criterion, update_frequency ): """ 根据给定的条件更新模型的权重。 参数: epoch (int): 当前训练轮次。 model_to_update: 需要更新的模型。 other_models (list): 其他模型列表,用于计算全局模型权重。 global_model: 全局模型。 losses (list): 各模型的损失值列表。 val_loader: 验证数据的 DataLoader。 device: 设备 ('cuda' 或 'cpu')。 val_auc_threshold (float): 当前需要更新模型的验证 AUC。 aggregate_weights (function): 权重聚合函数。 validate_model (function): 验证模型的函数。 update_frequency (int): 权重更新的频率。 返回: val_acc (float): 全局模型的验证精度。 val_auc (float): 全局模型的验证 AUC。 updated_val_auc_threshold (float): 更新后的验证 AUC。 """ if (epoch + 1) % update_frequency == 0: # 获取所有模型的权重 all_weights = [model.state_dict() for model in other_models] avg_weights = aggregate_weights(all_weights) # 聚合权重 # 更新全局模型权重 global_model.load_state_dict(avg_weights) # 计算加权平均损失 weighted_loss = sum(loss * 0.33 for loss in losses) print(f"Weighted Average Loss: {weighted_loss:.4f}") # 验证全局模型 val_loss, val_acc, val_auc = validate_model(device, global_model, val_loader, criterion, epoch, 'global_model') print(f'global_model Validation Accuracy: {val_acc:.4f}, global_model Validation AUC: {val_auc:.4f}') # 如果全局模型的 AUC 更高,则更新目标模型 if val_auc > val_auc_threshold: print(f'Updating model at epoch {epoch + 1}') model_to_update.load_state_dict(global_model.state_dict()) val_auc_threshold = val_auc # 更新 AUC 阈值 return val_acc, val_auc, val_auc_threshold return None, None, val_auc_threshold def update_model_weights( epoch, model_to_update, other_models, global_model, losses, val_loader, device, val_auc_threshold, # 当前需要更新模型的验证 AUC 阈值 validate_model, criterion, update_frequency ): """ 根据给定的条件更新模型的权重。 参数: epoch (int): 当前训练轮次。 model_to_update: 需要更新的模型。 other_models (list): 其他模型列表,用于计算全局模型权重。 global_model: 全局模型。 losses (list): 各模型的损失值列表。 val_loader: 验证数据的 DataLoader。 device: 设备 ('cuda' 或 'cpu')。 val_auc_threshold (float): 当前需要更新模型的验证 AUC。 aggregate_weights (function): 权重聚合函数。 validate_model (function): 验证模型的函数。 update_frequency (int): 权重更新的频率。 返回: val_acc (float): 全局模型的验证精度。 val_auc (float): 全局模型的验证 AUC。 updated_val_auc_threshold (float): 更新后的验证 AUC。 """ if (epoch + 1) % update_frequency == 0: # 获取所有模型的权重 all_weights = [model.state_dict() for model in other_models] avg_weights = aggregate_weights(all_weights) # 聚合权重 # 更新全局模型权重 global_model.load_state_dict(avg_weights) # 计算加权平均损失 weighted_loss = sum(loss * 0.33 for loss in losses) print(f"Weighted Average Loss: {weighted_loss:.4f}") # 验证全局模型 val_loss, val_acc, val_auc = validate_deepmodel(device, global_model, val_loader, criterion, epoch, 'global_model') print(f'global_model Validation Accuracy: {val_acc:.4f}, global_model Validation AUC: {val_auc:.4f}') # 如果全局模型的 AUC 更高,则更新目标模型 if val_auc > val_auc_threshold: print(f'Updating model at epoch {epoch + 1}') model_to_update.load_state_dict(global_model.state_dict()) val_auc_threshold = val_auc # 更新 AUC 阈值 return val_acc, val_auc, val_auc_threshold return None, None, val_auc_threshold def f_update_model_weights( epoch, model_to_update, other_models, global_model, losses, val_loader, device, val_auc_threshold, # 当前需要更新模型的验证 AUC 阈值 aggregate_weights, # 权重聚合函数 validate_model, criterion, update_frequency ): """ 根据给定的条件更新模型的权重。 参数: epoch (int): 当前训练轮次。 model_to_update: 需要更新的模型。 other_models (list): 其他模型列表,用于计算全局模型权重。 global_model: 全局模型。 losses (list): 各模型的损失值列表。 val_loader: 验证数据的 DataLoader。 device: 设备 ('cuda' 或 'cpu')。 val_auc_threshold (float): 当前需要更新模型的验证 AUC 阈值。 aggregate_weights (function): 权重聚合函数。 validate_model (function): 验证模型的函数。 criterion: 损失函数。 update_frequency (int): 权重更新的频率。 返回: val_acc (float): 全局模型的验证精度。 val_auc (float): 全局模型的验证 AUC。 updated_val_auc_threshold (float): 更新后的验证 AUC 阈值。 """ # 每隔指定的 epoch 更新一次模型权重 if (epoch + 1) % update_frequency == 0: print(f"\n[Epoch {epoch + 1}] Updating global model weights...") # 获取其他模型的权重 all_weights = [model.state_dict() for model in other_models] # 使用聚合函数计算全局权重 avg_weights = aggregate_weights(all_weights) print("Global model weights aggregated.") # 更新全局模型权重 global_model.load_state_dict(avg_weights) # 计算加权平均损失 weighted_loss = sum(loss * (1 / len(losses)) for loss in losses) # 平均加权 print(f"Weighted Average Loss: {weighted_loss:.4f}") # 验证全局模型性能 val_loss, val_acc, val_auc = validate_deepmodel(device, global_model, val_loader, criterion, epoch, 'global_model') print(f"[Global Model] Validation Loss: {val_loss:.4f}, Accuracy: {val_acc:.4f}, AUC: {val_auc:.4f}") # 如果全局模型 AUC 高于阈值,则更新目标模型权重 if val_auc > val_auc_threshold: print(f"Global model AUC improved ({val_auc:.4f} > {val_auc_threshold:.4f}). Updating target model.") model_to_update.load_state_dict(global_model.state_dict()) val_auc_threshold = val_auc # 更新 AUC 阈值 else: print( f"Global model AUC did not improve ({val_auc:.4f} <= {val_auc_threshold:.4f}). No update to target model.") return val_acc, val_auc, val_auc_threshold # 如果未到达更新频率,返回当前的 AUC 阈值 return None, None, val_auc_threshold