donalee / wetas

Weakly Supervised Temporal Anomaly Segmentation

WETAS: Weakly Supervised Time-Series Anomaly Segmentation

Figure 1. Two different strategies for localizing temporal anomalies.

• This is the author code of "Weakly Supervised Temporal Anomaly Segmentation with Dynamic Time Warping" (ICCV 2021).
• We employ (and customize) the publicly availabe implementation of soft-dtw, please refer to this repository.

Overview

Figure 2. The overall framework of WETAS, optimized by both the classification loss and the alignment loss.

Our proposed framework, termed as WETAS, optimizes the parameters of the dilated CNN by leveraging only weak supervision (i.e., instance-level anomaly labels, rather than point-level anomaly labels). To fully utilize the given instance-level anomaly labels, two different types of losses are considered (Figure 2, right).

1. The classification loss for correctly classifying an input instance as its instance-level anomaly label
2. The alignment loss for matching the input instance with the sequential anomaly label, which is synthesized by the model by distilling the instance-level label.

For temporal anomaly segmentation on a test input instance, it utilizes dynamic time warping (DTW) which outputs the optimal alignment between a target instance and the sequential anomaly label (Figure 2, left).

Run the codes

STEP 1. Install the python libraries / packages

• numpy
• numba
• scikit-learn
• pytorch

STEP 2. Download the real-world datasets for temporal anomaly segmentation

• You can download the raw datasets from the following links.

  • Electromyography Dataset (EMG) : http://archive.ics.uci.edu/ml/datasets/EMG+data+for+gestures
  • Gasoil Plant Heating Loop Dataset (GHL) : https://kas.pr/ics-research/dataset_ghl_1
  • Server Machine Dataset (SMD) : https://github.com/smallcowbaby/OmniAnomaly
  • Subway Entrance/Exit Dataset (Subway) : available for public use upon request to the authors of (Adam et al., 2008)

• You need to split the whole temporal data into a training set, a validation set, and a test set.

• Then, place the files in the corresponding directories.

  • ./data/{DATASET}/train, ./data/{DATASET}/valid, and ./data/{DATASET}/test

• We provide the preprocessed EMG dataset as an example.

STEP 3. Train and evaluate the WETAS framework

• You can simply run the code with the default setting, by using the following command.

python train_classifier.py

• For the EMG dataset, the training process will be printed like as below.

Epoch [25/200], step [15/15], Train Loss : 0.561529 (BCE : 0.482434, DTW : 0.079095), Valid loss : 0.530925 (BCE : 0.448849, DTW : 0.082075)
	Valid (WEAK) AUC : 0.850989, AUPRC : 0.537640, Best F1 : 0.597938,  Precision : 0.547619, Recall : 0.605263, threshold : 0.214676
	Test  (WEAK) AUC : 0.894690, AUPRC : 0.655222, Best F1 : 0.634146, Precision : 0.418182, Recall : 0.901961
	Test (DENSE) F1 : 0.040875, Precision : 0.061622, Recall : 0.030579, IoU : 0.020864

Epoch [50/200], step [15/15], Train Loss : 0.387882 (BCE : 0.308741, DTW : 0.079141), Valid loss : 0.441821 (BCE : 0.359525, DTW : 0.082296)
	Valid (WEAK) AUC : 0.896036, AUPRC : 0.738797, Best F1 : 0.688822,  Precision : 0.453125, Recall : 0.763158, threshold : 0.247649
	Test  (WEAK) AUC : 0.899510, AUPRC : 0.743431, Best F1 : 0.742857, Precision : 0.523810, Recall : 0.862745
	Test (DENSE) F1 : 0.560338, Precision : 0.405383, Recall : 0.907055, IoU : 0.389215

Epoch [75/200], step [15/15], Train Loss : 0.351386 (BCE : 0.272270, DTW : 0.079117), Valid loss : 0.461061 (BCE : 0.378665, DTW : 0.082396)
	Valid (WEAK) AUC : 0.906921, AUPRC : 0.783545, Best F1 : 0.716846,  Precision : 0.409091, Recall : 0.710526, threshold : 0.174542
	Test  (WEAK) AUC : 0.905147, AUPRC : 0.777514, Best F1 : 0.763636, Precision : 0.623188, Recall : 0.843137
	Test (DENSE) F1 : 0.580832, Precision : 0.453260, Recall : 0.808342, IoU : 0.409276

Epoch [100/200], step [15/15], Train Loss : 0.319351 (BCE : 0.240257, DTW : 0.079095), Valid loss : 0.394851 (BCE : 0.312770, DTW : 0.082081)
	Valid (WEAK) AUC : 0.931307, AUPRC : 0.826301, Best F1 : 0.763889,  Precision : 0.375000, Recall : 0.868421, threshold : 0.244523
	Test  (WEAK) AUC : 0.917729, AUPRC : 0.807198, Best F1 : 0.773585, Precision : 0.652174, Recall : 0.882353
	Test (DENSE) F1 : 0.603051, Precision : 0.450450, Recall : 0.912017, IoU : 0.431691

• You can specify the details of the framework and its optimization by input arguments.

Citation

@inproceedings{lee2021weakly,
  title={Weakly Supervised Temporal Anomaly Segmentation with Dynamic Time Warping},
  author={Lee, Dongha and Yu, Sehun and Ju, Hyunjun and Yu, Hwanjo},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={7355--7364},
  year={2021}
}

wetas's Issues

There is a problem with the 'Activation map' formula in the code

@donalee
The location of the problem is line 149 of the dtw_loss() funtion of the 'model.py' file.
def dtw_loss(self, out, wlabel): h = self.fc(out).squeeze(dim=2) dscore = torch.sigmoid(self.fc(out).squeeze(dim=2)) with torch.no_grad(): # Activation map actmap = h actmin = torch.min(actmap, dim=1)[0] actmap = actmap - actmin.unsqueeze(dim=1) actmax = torch.max(actmap, dim=1)[0] actmap = actmap / actmax.unsqueeze(dim=1)
Min-Max normalization should be (actmap - actmin)/(actmax - actmin), but in the code is (actmap - actmin)/actmax.

Hello Mr. Big, how to preprocess this EMG dataset and do you have any relevant code references

Hello Mr. Big, how to preprocess this EMG dataset and do you have any relevant code references