LogFormer: A Pre-train and Tuning Pipeline for Log Anomaly Detection
Authors: Hongcheng Guo, Jian Yang, Jiaheng Liu, Jiaqi Bai, Boyang Wang, Zhoujun Li, Tieqiao Zheng, Bo Zhang, Junran Peng, Qi Tian
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The proposed method is evaluated on three public and one real-world datasets. Experimental results on multiple benchmarks demonstrate the effectiveness of our Log Former with fewer trainable parameters and lower training costs. |
| Researcher Affiliation | Collaboration | Hongcheng Guo1, Jian Yang1, Jiaheng Liu1, Jiaqi Bai1, Boyang Wang1, Zhoujun Li1*, Tieqiao Zheng2, Bo Zhang2, Junran Peng1, Qi Tian3 1State Key Lab of Software Development Environment, Beihang University, Beijing, China 2Cloudwise Research, Beijing, China 3Huawei, Beijing, China |
| Pseudocode | No | The paper describes the Log Former architecture and processes using text and figures, but it does not include any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | two variants of Log Former 3. ... 3https://github.com/HC-Guo/Log Former. |
| Open Datasets | Yes | We conduct experiments on three datasets from Log Hub (He et al. 2020) 1. HDFS (Xu et al. 2010) dataset is generated and collected from the Amazon EC2 platform through running Hadoop-based map-reduce jobs. Thunderbird and BGL datasets (Oliner and Stearley 2007) contain logs collected from a two-supercomputer system at Sandia National Labs (SNL) in Albuquerque. 1https://github.com/logpai/loghub |
| Dataset Splits | No | For each dataset, considering that logs evolve over time, we select the first 80% (according to the timestamp of logs) log sequences for training and the rest 20% for testing, which is consistent with the prior work (Yang et al. 2021a; Du et al. 2017). The paper specifies a training and testing split but does not explicitly mention a separate validation set or split percentage for validation. |
| Hardware Specification | Yes | All runs are trained on 2 NVIDIA A100(40G) with a batch size of 64. |
| Software Dependencies | No | The paper mentions using 'Adam as the optimization algorithm' and 'pre-trained sentence-bert model' for feature extraction, and 'Drain (He et al. 2017)' for log parsing. However, it does not specify version numbers for these software components or any other libraries/frameworks (e.g., Python, PyTorch, TensorFlow versions) that would be needed for reproducibility. |
| Experiment Setup | Yes | In experiments, we use different numbers of transformer encoder layers in {1, 2, 4}. The number of attention heads is 8, and the size of the feedforward network that takes the output of the multi-head self-attention mechanism is 3072. We use Adam as the optimization algorithm whose learning rate is scheduled by One Cycle LR, with β1 = 0.9, β2 = 0.99, and ε = 10 8. All runs are trained on 2 NVIDIA A100(40G) with a batch size of 64. For each dataset, we tune the maximum learning of the One Cycle LR scheduler in {1e 5, 5e 5, 1e 6}. |