vision/src/autogluon/vision/_gluoncv/image_classification.py [302:355]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def fit(self, train_data, val_data=None, train_size=0.9, random_state=None, time_limit=None):
        """Fit auto estimator given the input data.
        Parameters
        ----------
        train_data : pd.DataFrame or iterator
            Training data.
        val_data : pd.DataFrame or iterator, optional
            Validation data, optional. If `train_data` is DataFrame, `val_data` will be split from
            `train_data` given `train_size`.
        train_size : float
            The portion of train data split from original `train_data` if `val_data` is not provided.
        random_state : int
            Random state for splitting, for `np.random.seed`.
        time_limit : int, default is None
            The wall clock time limit(second) for fit process, if `None`, time limit is not enforced.
            If `fit` takes longer than `time_limit`, the process will terminate early and return the
            model prematurally.
            Due to callbacks and additional validation functions, the `time_limit` may not be very precise
            (few minutes allowance), but you can use it to safe-guard a very long training session.
            If `time_limits` key set in __init__ with config, the `time_limit` value will overwrite configuration
            if not `None`.
        Returns
        -------
        Estimator
            The estimator obtained by training on the specified dataset.
        """
        config = self._config.copy()
        if time_limit is None:
            if config.get('time_limits', None):
                time_limit = config['time_limits']
            else:
                time_limit = math.inf
        elif not isinstance(time_limit, int):
            raise TypeError(f'Invalid type `time_limit={time_limit}`, int or None expected')
        self.scheduler_options['time_out'] = time_limit
        wall_clock_tick = time.time() + time_limit
        # split train/val before HPO to make fair comparisons
        if not isinstance(train_data, pd.DataFrame):
            assert val_data is not None, \
                "Please provide `val_data` as we do not know how to split `train_data` of type: \
                {}".format(type(train_data))

        if val_data is None:
            assert 0 <= train_size <= 1.0
            if random_state:
                np.random.seed(random_state)
            split_mask = np.random.rand(len(train_data)) < train_size
            train = train_data[split_mask]
            val = train_data[~split_mask]
            self._logger.info('Randomly split train_data into train[%d]/validation[%d] splits.',
                              len(train), len(val))
            train_data, val_data = train, val

        estimator = config.get('estimator', None)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



vision/src/autogluon/vision/_gluoncv/object_detection.py [268:322]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def fit(self, train_data, val_data=None, train_size=0.9, random_state=None, time_limit=None):
        """Fit auto estimator given the input data.
        Parameters
        ----------
        train_data : pd.DataFrame or iterator
            Training data.
        val_data : pd.DataFrame or iterator, optional
            Validation data, optional. If `train_data` is DataFrame, `val_data` will be split from
            `train_data` given `train_size`.
        train_size : float
            The portion of train data split from original `train_data` if `val_data` is not provided.
        random_state : int
            Random state for splitting, for `np.random.seed`.
        time_limit : int, default is None
            The wall clock time limit(second) for fit process, if `None`, time limit is not enforced.
            If `fit` takes longer than `time_limit`, the process will terminate early and return the
            model prematurally.
            Due to callbacks and additional validation functions, the `time_limit` may not be very precise
            (few minutes allowance), but you can use it to safe-guard a very long training session.
            If `time_limits` key set in __init__ with config, the `time_limit` value will overwrite configuration
            if not `None`.
        Returns
        -------
        Estimator
            The estimator obtained by training on the specified dataset.
        """
        config = self._config.copy()
        if time_limit is None:
            if config.get('time_limits', None):
                time_limit = config['time_limits']
            else:
                time_limit = math.inf
        elif not isinstance(time_limit, int):
            raise TypeError(f'Invalid type `time_limit={time_limit}`, int or None expected')
        self.scheduler_options['time_out'] = time_limit
        wall_clock_tick = time.time() + time_limit
        # split train/val before HPO to make fair comparisons
        if not isinstance(train_data, pd.DataFrame):
            assert val_data is not None, \
                "Please provide `val_data` as we do not know how to split `train_data` of type: \
                {}".format(type(train_data))

        if val_data is None:
            assert 0 <= train_size <= 1.0
            if random_state:
                np.random.seed(random_state)
            split_mask = np.random.rand(len(train_data)) < train_size
            train = train_data[split_mask]
            val = train_data[~split_mask]
            self._logger.info('Randomly split train_data into train[%d]/validation[%d] splits.',
                              len(train), len(val))
            train_data, val_data = train, val

        # automatically suggest some hyperparameters based on the dataset statistics(experimental)
        estimator = config.get('estimator', None)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -