[AutoMM] Fix hyperparameters consistency between init() and fit() #3279
Conversation
zhiqiangdon
added
the
model list checked
|
Job PR-3279-28209df is done. |
|
Job PR-3279-70e62f0 is done. |
| @@ -126,6 +126,8 @@ def __init__( | |||
| verbosity: Optional[int] = 3, | |||
| warn_if_exist: Optional[bool] = True, | |||
| enable_progress_bar: Optional[bool] = None, | |||
| pretrained: Optional[bool] = True, | |||
There was a problem hiding this comment.
we shall update docstring to reflect his new param
There was a problem hiding this comment.
MultiModalMatcher is not exposed to users. We can add the docstring later.
There was a problem hiding this comment.
Docstring added.
|
|
||
| if self._config is None: | ||
| # split out the hyperparameters whose values are complex objects | ||
| hyperparameters, advanced_hyperparameters = split_hyperparameters(self._hyperparameters) |
There was a problem hiding this comment.
what's the difference between regular and advanced hps? In future, what's the guiding principal to introduce new ones?
There was a problem hiding this comment.
advanced_hyperparameters are to support more image augmentations from torchvision transforms (https://pytorch.org/vision/stable/transforms.html#geometry). In this case, the values of hyperparameters are callable objects instead of strings or numbers.
| @@ -375,7 +375,7 @@ def object_detection(presets: str = DEFAULT): | |||
| "optimization.top_k_average_method": "best", | |||
| "optimization.warmup_steps": 0.0, | |||
| "optimization.patience": 10, | |||
| "optimization.val_metric": "map", | |||
| # "optimization.val_metric": "map", | |||
There was a problem hiding this comment.
Because this PR changes the design of customizing validation metric. optimization.val_metric has been removed from the default config.
There was a problem hiding this comment.
Removed the commented line. Object detection's default validation metric is map.
|
Job PR-3279-98d6804 is done. |
| @@ -493,14 +487,17 @@ def match_label(self): | |||
|
|
|||
| @property | |||
| def problem_type(self): | |||
There was a problem hiding this comment.
It's a bit confusing here to include self._matcher._pipeline only in self.problem_type but not in self._problem_type
There was a problem hiding this comment.
Updated to set predictor._matcher._pipeline to predictor._problem_type (https://github.com/autogluon/autogluon/pull/3279/files#diff-a73844176f6eae3b4538467b3565e3bdbf4d88997c3c670173bbb44e1a55b7a3R2700) when loading a saved matcher.
|
Job PR-3279-98d6804 is done. |
|
Job PR-3279-3031945 is done. |
Issue #, if available:
#2978
Description of changes:
hyperprametersconsistent between predictor init() and fit().hyperparametersconsistent between matcher init() and fit().config.optimization.val_metricwithvalidation_metricin init(), which is to reduce hacking and simplify the logic.init_scracthwithpretrainedin init() to align with AutoMM motivation and other internal APIs. AutoMM is mainly to finetune pretrained models.pretrainedis more intuitive for users compared toinit_scratch. Some internal APIs already usedpretrained.By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.