数据划分

Key changes made:

  • Split trajectories into train/test before any sampling

  • Generate anomalies (detours and switches) using the original non-sampled test trajectories

  • Apply sampling rate to the generated anomalies after they're created

  • Keep the same sampling approach for train, test, and OOD data

This modification ensures that:

  • Anomaly generation uses complete trajectory information

  • The sampling rate is consistently applied to all final datasets

  • The relative structure of the anomalies is preserved before sampling

得到的结果，相差是否很大？以西安为例

sample_rates = [1, 0.99, 0.95, 0.92, 0.90]

Results for xian:

==============================

Client client_0:

Training time: 1.37 minutes

Inference time: 2.59 minutes

Total time: 3.96 minutes

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.7410, PR_AUC: 0.6993

Normal & Switch - ROC_AUC: 0.7998, PR_AUC: 0.7754

OOD & Detour - ROC_AUC: 0.5806, PR_AUC: 0.6753

OOD & Switch - ROC_AUC: 0.6483, PR_AUC: 0.7361

Evaluation time: 0.03 minutes

Client client_1:

Training time: 0.94 minutes

Inference time: 2.34 minutes

Total time: 3.27 minutes

Training completed at: 2024-12-25 23:46:56

Inference completed at: 2024-12-25 23:49:16

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.6327, PR_AUC: 0.5078

Normal & Switch - ROC_AUC: 0.6616, PR_AUC: 0.5527

OOD & Detour - ROC_AUC: 0.5115, PR_AUC: 0.5666

OOD & Switch - ROC_AUC: 0.5417, PR_AUC: 0.5988

Evaluation time: 0.02 minutes

Client client_2:

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.5152, PR_AUC: 0.3801

Normal & Switch - ROC_AUC: 0.5509, PR_AUC: 0.3985

OOD & Detour - ROC_AUC: 0.4260, PR_AUC: 0.4795

OOD & Switch - ROC_AUC: 0.4631, PR_AUC: 0.4982

Evaluation time: 0.02 minutes

Client client_3:

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.5012, PR_AUC: 0.3721

Normal & Switch - ROC_AUC: 0.5502, PR_AUC: 0.4088

OOD & Detour - ROC_AUC: 0.5308, PR_AUC: 0.5529

OOD & Switch - ROC_AUC: 0.5797, PR_AUC: 0.5914

Evaluation time: 0.01 minutes

Client client_4:

Training time: 0.42 minutes

Inference time: 1.91 minutes

Total time: 2.33 minutes

Training completed at: 2024-12-25 23:55:16

Inference completed at: 2024-12-25 23:57:11

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.4824, PR_AUC: 0.3438

Normal & Switch - ROC_AUC: 0.5304, PR_AUC: 0.3740

OOD & Detour - ROC_AUC: 0.5116, PR_AUC: 0.5309

OOD & Switch - ROC_AUC: 0.5590, PR_AUC: 0.5639

Evaluation time: 0.01 minutes

可以看见采样率取值一般是在0.9-1之间比较合适

2. 然后，再用Fedavg确认效果

System Setup:

  • The system initializes multiple clients (default is 5)

  • Each client gets their own trainer with specific data paths

  • All clients share the same model architecture but start with different weights

  • Clients have their own local datasets for training and testing

Training Process:

  • The training happens in rounds (federated rounds)

  • In each round:

    • First, each client trains independently on their local data

    • Then, all client models are collected centrally

    • The central server averages all model parameters (this is the FedAvg step)

    • Finally, all clients receive the averaged model back

Parameter Averaging:

  • After local training, each client's model parameters are collected

  • The system creates an average by:

    • Taking each layer's parameters from all clients

    • Computing the mathematical average for each parameter

    • Creating a new model with these averaged parameters

  • This averaged model is then distributed back to all clients

Exclude road embedding in average:

Most parameters are averaged across clients (VAE, confidence layers, projection heads)

BUT road embeddings are specifically kept local to each client through:

  • Storing original road embeddings for each client

  • Not including them in the averaging process

  • Returning them back to their original clients unchanged

以西安的FedAvg为例，各个客户端使用FedAvg反而下降了，说明直接聚合是不可行的

Client client_0:

Training time: 8.60 minutes

Inference time: 2.65 minutes

Total time: 11.25 minutes

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.6848, PR_AUC: 0.6206

Normal & Switch - ROC_AUC: 0.7600, PR_AUC: 0.7263

OOD & Detour - ROC_AUC: 0.5455, PR_AUC: 0.6310

OOD & Switch - ROC_AUC: 0.6268, PR_AUC: 0.7111

Evaluation time: 0.04 minutes

Client client_1:

Evaluation Metrics:

Normal & Detour - ROC_AUC: 0.5955, PR_AUC: 0.4640

Normal & Switch - ROC_AUC: 0.6609, PR_AUC: 0.5461

OOD & Detour - ROC_AUC: 0.4926, PR_AUC: 0.5418

OOD & Switch - ROC_AUC: 0.5586, PR_AUC: 0.6050

Client client_2:

Normal & Detour - ROC_AUC: 0.5039, PR_AUC: 0.3612

Normal & Switch - ROC_AUC: 0.5601, PR_AUC: 0.3896

OOD & Detour - ROC_AUC: 0.4298, PR_AUC: 0.4740

OOD & Switch - ROC_AUC: 0.4862, PR_AUC: 0.5030

Client client_3:

Normal & Detour - ROC_AUC: 0.4737, PR_AUC: 0.3356

Normal & Switch - ROC_AUC: 0.5273, PR_AUC: 0.3672

OOD & Detour - ROC_AUC: 0.5216, PR_AUC: 0.5350

OOD & Switch - ROC_AUC: 0.5744, PR_AUC: 0.5708

Evaluation time: 0.04 minutes

Client client_4:

Normal & Detour - ROC_AUC: 0.4614, PR_AUC: 0.3180

Normal & Switch - ROC_AUC: 0.5204, PR_AUC: 0.3524

OOD & Detour - ROC_AUC: 0.5107, PR_AUC: 0.5238

OOD & Switch - ROC_AUC: 0.5678, PR_AUC: 0.5638

posted @ 2024-12-29 13:44  GraphL  阅读(6)  评论(0)    收藏  举报