ModelConfig has a flag 'enabled'. It defines whether the model counters (n_wt) should be updated as processor receives new batches. When the model is disabled it should be 'read-only'. For disabled models it should be still possible to calculate the accumulative scores (such as perplexity). Right now this is not the case, see processor.cc:
// do not process disabled models.
if (!model.enabled()) return;
The 'enabled' flag should be renamed to 'is_read_only". It defines better what this boolean flag actually stands for.
In "is_read_only=true" mode scores should be calculated
Merger and MasterComponent should respect is_read_only = true, and never "reset" the model.
ModelConfig has a flag 'enabled'. It defines whether the model counters (n_wt) should be updated as processor receives new batches. When the model is disabled it should be 'read-only'. For disabled models it should be still possible to calculate the accumulative scores (such as perplexity). Right now this is not the case, see processor.cc: