In our project, we make such an assumption that both AI and Web services will be launched simultaneously, thus we synthesize all of them into a unified docker compose file for the ease of administration. At the moment, we provide three different versions of docker compose configurations, including input_2d, input_3d, input_mixed, and you can run all of them at the same time.
If model is from remote mlflow tracking server, please use mlflow_export_import script to download appropriate experiment at first. You, then, modify appropriate path(s) for backbone's weight in downloaded configuration file. Check this for backbones' weight and our pretrained models.
There are four environment variables, be it DEVICE, WEIGHT_PATH, PORT and CONFIG_PATH, you should pay your attention to. WEIGHT_PATH is path to deployed model's weight with root from /weight. CONFIG_PATH is path to deployed model's configuration json file with root from /weight. PORT is used in case of port conflict circumstance.
This variable is used to determine root directory for plots, scores and videos, which are used to store inferred results.
Two anomalous regions are merged if their frame distance ≤ MERGE_GAP.
The endpoint of AI service.
We applied Savitzky-Golay filter to smooth input anomalous score, which requires WINDOW_LENGTH and WINDOW_LENGTH parameter. See more at here.
We utilized find_peak() from scipy to detect anomalous regions from model's projected scores. See more at here.
The endpoint of BE service.
The exposed port of FE service.