elasticdl
elasticdl copied to clipboard
None of the checkpoints saved from training process is evaluated.
In EDL train command, there are two parameters checkpoint_steps and evaluation_steps. Per checkpoint_steps, the training checkpoint will be saved. Per evaluation_steps, the checkpoint file for evaluation will be saved and the evaluation task for this checkpoint will generated and assigned to the workers. But the checkpoint files from training process are not evaluated at all.
Please refer to #1252 , I think we might have two kinds of evaluation.