[Retiarii] Rename APIs and refine documentation #3404
Conversation
docs/en_US/NAS/retiarii/Tutorial.rst
Outdated
|
|
||
| Create a Trainer and Exploration Strategy | ||
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ | ||
| Create a Evaluator and Exploration Strategy |
There was a problem hiding this comment.
i will revise the doc in another pr
| @@ -0,0 +1,2 @@ | |||
| from .base import PyTorchImageClassificationTrainer, PyTorchMultiModelTrainer | |||
There was a problem hiding this comment.
why not change Trainer to Evaluator?
There was a problem hiding this comment.
The base file is deprecated. These are not public APIs.
|
|
||
|
|
||
| class Lightning(TrainingConfig): | ||
| class Lightning(ModelEvaluator): |
There was a problem hiding this comment.
what is the difference between ModelEvaluator, Evaluator and Trainer?
There was a problem hiding this comment.
Renamed modelevaluator -> evaluator. Trainer is a concept in oneshot only.
There was a problem hiding this comment.
Maybe we can deprecate the concept of trainer and rename DartsTrainer -> DARTS, EnasTrainer -> ENAS? Leave that for discussion.
| if model_id not in evaluator_slot: | ||
| phy_model.evaluator.kwargs['model_kwargs'].append(new_node.graph.model.evaluator.kwargs.copy()) | ||
| evaluator_slot[model_id] = len(phy_model.evaluator.kwargs['model_kwargs']) - 1 | ||
| slot = evaluator_slot[model_id] | ||
| phy_model.evaluator.kwargs['model_kwargs'][slot]['model_id'] = model_id | ||
| phy_model.evaluator.kwargs['model_kwargs'][slot]['use_input'] = False | ||
| phy_model.evaluator.kwargs['model_kwargs'][slot]['use_output'] = False | ||
| else: | ||
| slot = training_config_slot[model_id] | ||
| slot = evaluator_slot[model_id] | ||
| # If a model's inputs/outputs are not used in the multi-model | ||
| # the codegen and trainer should not generate and use them | ||
| # "use_input" and "use_output" are used to mark whether | ||
| # an input/output of a model is used in a multi-model | ||
| if new_node.operation.type == '_inputs': | ||
| input_slot_mapping[new_node] = slot | ||
| phy_model.training_config.kwargs['model_kwargs'][slot]['use_input'] = True | ||
| phy_model.evaluator.kwargs['model_kwargs'][slot]['use_input'] = True | ||
| if new_node.operation.type == '_outputs': | ||
| output_slot_mapping[new_node] = slot | ||
| phy_model.training_config.kwargs['model_kwargs'][slot]['use_output'] = True | ||
| phy_model.evaluator.kwargs['model_kwargs'][slot]['use_output'] = True |
There was a problem hiding this comment.
@hzhua could you please help review this part?
There was a problem hiding this comment.
No problem if evaluator has the same structure with training_config.
No description provided.