Anomalib#

Contents

Supported version in EVAM#

  • v0.7.0

Training#

Anomalib includes ready to use anomaly detection models. A model can be trained based on default config provided for the model or can be customized for a particular dataset and category.

More details on Training and exporting a trained model can be found here.

Supported models in EVAM#

Currently a subset of available anomaly detection models in anomalib are supported in EVAM:

Inference using Anomalib UDF#

When trained/exported model is available, [EVAM_WORKDIR]/user_scripts/udfs/python/anomalib_udf/ can be used for running inference. As an example, a custom STFPM model trained on Amazon’s Visual Anomaly (Visa) dataset is included to detect anomalies in PCB.

Inferencing could be configured to be based on

  • openvino - default openvino inferencer provided by Anomalib or

  • openvino_nomask - custom openvino no mask inferencer derived from openvino inferencer which allows for image resolution to not have an impact on inference time.

For more details on the configuration of the UDF, refer to this [EVAM_WORKDIR]//user_scripts/udfs/python/anomalib_udf/README.md.