In training steps, current only error metric is supported in validation data, other metrics like AUC, recall or precision are also helpful especially for final model selection.
To support such new metrics, two solutions may be helpful:
Using one worker for only evaluation in each several epoch, validation data at this point is assigned to such worker, such worker can be called to evaluator, but it will be limited by memory, how to scale? using another MR jobs can scale but seems heavy.
By computing AUC, Presion, Recall in each worker and merge them in master by using avg/max/min ...
In training steps, current only error metric is supported in validation data, other metrics like AUC, recall or precision are also helpful especially for final model selection.
To support such new metrics, two solutions may be helpful: