 b/aggmap/AggMapNet.py
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Aug 16 17:10:53 2020
+@author: wanxiang.shen@u.nus.edu
+"""
+import warnings, os
+warnings.filterwarnings("ignore")
+import numpy as np
+import pandas as pd
+import tensorflow as tf
+tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.ERROR)
+from sklearn.utils.validation import check_X_y, check_array, check_is_fitted
+from sklearn.base import BaseEstimator, ClassifierMixin, RegressorMixin
+from sklearn.utils.multiclass import unique_labels
+from sklearn.metrics import get_scorer, SCORERS
+from aggmap import aggmodel
+from aggmap.aggmodel.explain_dev import GlobalIMP, LocalIMP
+from aggmap.aggmodel.explainer import shapley_explainer, simply_explainer
+from joblib import dump, load
+from  copy import copy
+from tensorflow.keras.models import load_model as load_tf_model
+import gc
+import tensorflow.keras.backend as K
+import tensorflow as tf
+def clean(clf):
+    del clf._model
+    del clf._performance
+    del clf
+    gc.collect()
+    K.clear_session()
+    tf.compat.v1.reset_default_graph() # TF graph isn't same as Keras graph
+def save_model(model, model_path):
+    if not os.path.exists(model_path):
+        os.makedirs(model_path)
+    print('saving model to %s' % model_path)
+    model_new = copy(model)
+    model_new._model.save(os.path.join(model_path, 'inner_model.h5'))
+    model_new._model = None
+    model_new._performance = None
+    res = dump(model_new,  os.path.join(model_path, 'outer_model.est'))
+    return res
+def load_model(model_path, gpuid=None):
+    '''
+    gpuid: load model to specific gpu: {None, 0, 1, 2, 3,..}
+    '''
+    model = load(os.path.join(model_path, 'outer_model.est'))
+    if gpuid==None:
+        gpuid = model.gpuid
+    else:
+        gpuid = str(gpuid)
+    os.environ["CUDA_VISIBLE_DEVICES"]= gpuid
+    model.gpuid = gpuid
+    model._model = load_tf_model(os.path.join(model_path, 'inner_model.h5'))
+    return model
+class RegressionEstimator(BaseEstimator, RegressorMixin):
+    """ An AggMap CNN Regression estimator (each sample belongs to only one class)
+    Parameters
+    ----------
+    epochs : int, default = 200
+        A parameter used for training epochs.
+    conv1_kernel_size: int, default = 13
+        A parameter used for the kernel size of first covolutional layers
+    dense_layers: list, default = [128]
+        A parameter used for the dense layers.
+    batch_size: int, default: 128
+        A parameter used for the batch size.
+    lr: float, default: 1e-4
+        A parameter used for the learning rate.
+    loss:string or function, default: 'mse'
+        A parameter used for the loss function
+    batch_norm: bool, default: False
+        batch normalization after convolution layers.
+    n_inception: int, default:2
+        Number of the inception layers.
+    dense_avf: str, default is 'relu'
+        activation fuction in the dense layers.
+    dropout: float, default: 0
+        A parameter used for the dropout of the dense layers.
+    monitor: str, default: 'val_loss'
+        {'val_loss', 'val_r2'}, a monitor for model selection
+    metric: str, default: 'r2'
+        {'r2', 'rmse'},  a matric parameter
+    patience: int, default: 10000
+        A parameter used for early stopping
+    gpuid: int, default: 0
+        A parameter used for specific gpu card
+    verbose: int, default: 0
+        if positive, then the log infomation of AggMapNet will be print
+        if negative, then the log infomation of orignal model will be print
+    random_state: int, default: 32
+        random seed.
+    Examples
+    --------
+    >>> from aggmap import AggModel
+    >>> clf = AggModel.RegressionEstimator()
+    """
+    def __init__(self,
+                 epochs = 200,
+                 conv1_kernel_size = 13,
+                 dense_layers = [128],
+                 dense_avf = 'relu',
+                 batch_size = 128,
+                 lr = 1e-4,
+                 loss = 'mse',
+                 batch_norm = False,
+                 n_inception = 2,
+                 dropout = 0.0,
+                 monitor = 'val_loss',
+                 metric = 'r2',
+                 patience = 10000,
+                 verbose = 0,
+                 random_state = 32,
+                 gpuid = 0,
+                ):
+        self.epochs = epochs
+        self.dense_layers = dense_layers
+        self.conv1_kernel_size = conv1_kernel_size
+        self.dense_avf = dense_avf
+        self.batch_size = batch_size
+        self.lr = lr
+        self.loss = loss
+        self.batch_norm = batch_norm
+        self.n_inception = n_inception
+        self.dropout = dropout
+        self.monitor = monitor
+        self.metric = metric
+        self.patience = patience
+        self.gpuid = str(gpuid)
+        os.environ["CUDA_VISIBLE_DEVICES"]= self.gpuid
+        self.verbose = verbose
+        self.random_state = random_state
+        self.is_fit = False
+        self.name = "AggMap Regression Estimator"
+        #print(self.get_params())
+        self.history = {}
+        self.history_model = {}
+        if self.verbose > 0:
+            self.verbose1 = self.verbose
+            self.verbose2 = 0
+        elif self.verbose ==0:
+            self.verbose1 = 0
+            self.verbose2 = 0
+        elif self.verbose < 0:
+            self.verbose1 = 0
+            self.verbose2 = abs(self.verbose)
+        print(self)
+    def get_params(self, deep=True):
+        model_paras =  {"epochs": self.epochs,
+                        "lr":self.lr,
+                        "loss":self.loss,
+                        "conv1_kernel_size": self.conv1_kernel_size,
+                        "dense_layers": self.dense_layers,
+                        "dense_avf":self.dense_avf,
+                        "batch_size":self.batch_size,
+                        "dropout":self.dropout,
+                        "batch_norm":self.batch_norm,
+                        "n_inception":self.n_inception,
+                        "monitor": self.monitor,
+                        "metric":self.metric,
+                        "patience":self.patience,
+                        "random_state":self.random_state,
+                        "verbose":self.verbose,
+                        "name":self.name,
+                        "gpuid": self.gpuid,
+                       }
+        return model_paras
+    def set_params(self, **parameters):
+        for parameter, value in parameters.items():
+            setattr(self, parameter, value)
+        return self
+    def fit(self, X, y,  X_valid = None, y_valid = None):
+        # Check that X and y have correct shape
+        if  X.ndim != 4:
+            raise ValueError("Found array X with dim %d. %s expected == 4." % (X.ndim, self.name))
+        if  y.ndim != 2:
+            raise ValueError("Found array y with dim %d. %s expected == 2." % (y.ndim, self.name))
+        self.X_ = X
+        self.y_ = y
+        if (X_valid is None) | (y_valid is None):
+            X_valid = X
+            y_valid = y
+        np.random.seed(self.random_state)
+        tf.compat.v1.set_random_seed(self.random_state)
+        model = aggmodel.net._AggMapNet(X.shape[1:],
+                                       n_outputs = y.shape[-1],
+                                       conv1_kernel_size = self.conv1_kernel_size,
+                                       batch_norm = self.batch_norm,
+                                       n_inception = self.n_inception,
+                                       dense_layers = self.dense_layers,
+                                       dense_avf = self.dense_avf,
+                                       dropout = self.dropout,
+                                       last_avf = 'linear')
+        opt = tf.keras.optimizers.Adam(lr=self.lr, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) #
+        model.compile(optimizer = opt, loss = self.loss)
+        performance = aggmodel.cbks.Reg_EarlyStoppingAndPerformance((X, y),
+                                                                    (X_valid, y_valid),
+                                                                    patience = self.patience,
+                                                                    criteria = self.monitor,
+                                                                    verbose = self.verbose1,)
+        history = model.fit(X, y,
+                            batch_size=self.batch_size,
+                            epochs= self.epochs, verbose= self.verbose2, shuffle = True,
+                            validation_data = (X_valid, y_valid),
+                            callbacks=[performance])
+        self._model = model
+        self._performance = performance
+        self.history = self._performance.history
+        self.history_model = history.history
+        self.is_fit = True
+        # Return the classifier
+        return self
+    def predict(self, X):
+        """
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features_w, n_features_h, n_features_c)
+            Vector to be scored, where `n_samples` is the number of samples and
+        Returns
+        -------
+        T : array-like of shape (n_samples, n_classes)
+            Returns the predicted values
+        """
+        y_pred = self._model.predict(X, verbose = self.verbose)
+        return y_pred
+    def score(self, X, y, scoring = 'r2', sample_weight=None):
+        """Returns the score using the `scoring` option on the given
+        test data and labels.
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            Test samples.
+        y : array-like of shape (n_samples,)
+            True labels for X.
+        scoring: str, please refer to: https://scikit-learn.org/stable/modules/model_evaluation.html#scoring-parameter
+        sample_weight : array-like of shape (n_samples,), default=None
+            Sample weights.
+        Returns
+        -------
+        score : float
+            Score of self.predict(X) wrt. y.
+        """
+        assert scoring in SCORERS.keys(), 'scoring is not in %s' % SCORERS.keys()
+        scoring = get_scorer(scoring)
+        return scoring(self, X, y, sample_weight=sample_weight)
+    def plot_model(self, to_file='model.png',
+                   show_shapes=True,
+                   show_layer_names=True,
+                   rankdir='TB',
+                   expand_nested=False,
+                   dpi=96):
+        if self.is_fit:
+            tf.keras.utils.plot_model(self._model,
+                       to_file=to_file,
+                       show_shapes=show_shapes,
+                       show_layer_names=show_layer_names,
+                       rankdir=rankdir,
+                       expand_nested=expand_nested,
+                       dpi=dpi)
+        else:
+            print('Please fit first!')
+    def save_model(self, model_path):
+        return save_model(self, model_path)
+    def load_model(self, model_path, gpuid=None):
+        return load_model(model_path, gpuid=gpuid)
+    def explain_model(self, mp, X, y,
+                      explain_format = 'global',
+                      apply_logrithm = False,
+                      apply_smoothing = False,
+                      kernel_size = 3, sigma = 1.2):
+        '''
+        Feature importance calculation
+        Parameters
+        --------------
+        mp: aggmap object
+        X: trianing or test set X arrays
+        y: trianing or test set y arrays
+        explain_format: {'local', 'global'}, default: 'global'
+            local or global feature importance, if local, then X must be one sample
+        apply_logrithm: {True, False}, default: False
+            whether apply a logarithm transformation on the importance values
+        apply_smoothing: {True, False}, default: False
+            whether apply a smoothing transformation on the importance values
+        kernel_size: odd number, the kernel size to perform the smoothing
+        sigma: float, sigma for gaussian smoothing
+        Returns
+        ------------
+        DataFrame of feature importance
+        '''
+        if explain_format == 'global':
+            explain_func = GlobalIMP
+        else:
+            explain_func = LocalIMP
+        dfe = explain_func(self, mp, X, y,
+                        task_type = 'regression',
+                        sigmoidy = False,
+                        apply_logrithm = apply_logrithm,
+                        apply_smoothing = apply_smoothing,
+                        kernel_size = kernel_size, sigma = sigma)
+        return dfe
+    @property
+    def clean(self):
+        clean(self)
+class MultiClassEstimator(BaseEstimator, ClassifierMixin):
+    """ An AggMap CNN MultiClass estimator (each sample belongs to only one class)
+    Parameters
+    ----------
+    epochs : int, default = 200
+        A parameter used for training epochs.
+    conv1_kernel_size: int, default = 13
+        A parameter used for the kernel size of first covolutional layers
+    dense_layers: list, default = [128]
+        A parameter used for the dense layers.
+    batch_size: int, default: 128
+        A parameter used for the batch size.
+    lr: float, default: 1e-4
+        A parameter used for the learning rate.
+    loss: string or function, default: 'categorical_crossentropy'
+        A parameter used for the loss function
+    batch_norm: bool, default: False
+        batch normalization after convolution layers.
+    n_inception: int, default:2
+        Number of the inception layers.
+    dense_avf: str, default is 'relu'
+        activation fuction in the dense layers.
+    dropout: float, default: 0
+        A parameter used for the dropout of the dense layers.
+    monitor: str, default: 'val_loss'
+        {'val_loss', 'val_metric'}, a monitor for model selection.
+    metric: str, default: 'ACC'
+        {'ROC', 'ACC', 'PRC'},  a matric parameter.
+    patience: int, default: 10000
+        A parameter used for early stopping.
+    gpuid: int, default: 0
+        A parameter used for specific gpu card.
+    verbose: int, default: 0
+        if positive, then the log infomation of AggMapNet will be print,
+        if negative, then the log infomation of orignal model will be print.
+    random_state: int, default: 32
+        Random seed.
+    Examples
+    --------
+    >>> from aggmap import AggModel
+    >>> clf = AggModel.MultiClassEstimator()
+    """
+    def __init__(self,
+                 epochs = 200,
+                 conv1_kernel_size = 13,
+                 dense_layers = [128],
+                 dense_avf = 'relu',
+                 batch_size = 128,
+                 lr = 1e-4,
+                 loss = 'categorical_crossentropy',
+                 batch_norm = False,
+                 n_inception = 2,
+                 dropout = 0.0,
+                 monitor = 'val_loss',
+                 metric = 'ACC',
+                 patience = 10000,
+                 verbose = 0,
+                 last_avf = 'softmax',
+                 random_state = 32,
+                 gpuid=0,
+                ):
+        self.epochs = epochs
+        self.dense_layers = dense_layers
+        self.conv1_kernel_size = conv1_kernel_size
+        self.dense_avf = dense_avf
+        self.batch_size = batch_size
+        self.lr = lr
+        self.loss = loss
+        self.last_avf = last_avf
+        self.batch_norm = batch_norm
+        self.n_inception = n_inception
+        self.dropout = dropout
+        self.monitor = monitor
+        self.metric = metric
+        self.patience = patience
+        self.gpuid = str(gpuid)
+        os.environ["CUDA_VISIBLE_DEVICES"]= self.gpuid
+        self.verbose = verbose
+        self.random_state = random_state
+        self.name = "AggMap MultiClass Estimator"
+        self.is_fit = False
+        #print(self.get_params())
+        self.history = {}
+        self.history_model = {}
+        if self.verbose > 0:
+            self.verbose1 = self.verbose
+            self.verbose2 = 0
+        elif self.verbose ==0:
+            self.verbose1 = 0
+            self.verbose2 = 0
+        elif self.verbose < 0:
+            self.verbose1 = 0
+            self.verbose2 = abs(self.verbose)
+        print(self)
+    def get_params(self, deep=True):
+        model_paras =  {"epochs": self.epochs,
+                        "lr":self.lr,
+                        "loss":self.loss,
+                        "conv1_kernel_size": self.conv1_kernel_size,
+                        "dense_layers": self.dense_layers,
+                        "dense_avf":self.dense_avf,
+                        "last_avf":self.last_avf,
+                        "batch_size":self.batch_size,
+                        "dropout":self.dropout,
+                        "batch_norm":self.batch_norm,
+                        "n_inception":self.n_inception,
+                        "monitor": self.monitor,
+                        "metric":self.metric,
+                        "patience":self.patience,
+                        "random_state":self.random_state,
+                        "verbose":self.verbose,
+                        "name":self.name,
+                        "gpuid": self.gpuid,
+                       }
+        return model_paras
+    def set_params(self, **parameters):
+        for parameter, value in parameters.items():
+            setattr(self, parameter, value)
+        return self
+    def fit(self, X, y,
+            X_valid = None,
+            y_valid = None,
+            class_weight = None,
+           ):
+        # Check that X and y have correct shape
+        if  X.ndim != 4:
+            raise ValueError("Found array X with dim %d. %s expected == 4." % (X.ndim, self.name))
+        if  y.ndim != 2:
+            raise ValueError("Found array y with dim %d. %s expected == 2." % (y.ndim, self.name))
+        # Store the classes seen during fit
+        self.classes_ = unique_labels(y)
+        self.X_ = X
+        self.y_ = y
+        if (X_valid is None) | (y_valid is None):
+            X_valid = X
+            y_valid = y
+        np.random.seed(self.random_state)
+        tf.compat.v1.set_random_seed(self.random_state)
+        model = aggmodel.net._AggMapNet(X.shape[1:],
+                                        n_outputs = y.shape[-1],
+                                        conv1_kernel_size = self.conv1_kernel_size,
+                                        batch_norm = self.batch_norm,
+                                        n_inception = self.n_inception,
+                                        dense_layers = self.dense_layers,
+                                        dense_avf = self.dense_avf,
+                                        dropout = self.dropout,
+                                        last_avf = self.last_avf)
+        opt = tf.keras.optimizers.Adam(lr=self.lr, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) #
+        model.compile(optimizer = opt, loss = self.loss, metrics = ['accuracy'])
+        performance = aggmodel.cbks.CLA_EarlyStoppingAndPerformance((X, y),
+                                                                    (X_valid, y_valid),
+                                                                    patience = self.patience,
+                                                                    criteria = self.monitor,
+                                                                    metric = self.metric,
+                                                                    last_avf= self.last_avf,
+                                                                    verbose = self.verbose1,)
+        history = model.fit(X, y,
+                  batch_size=self.batch_size,
+                  epochs= self.epochs, verbose= self.verbose2, shuffle = True,
+                  validation_data = (X_valid, y_valid), class_weight = class_weight,
+                  callbacks=[performance])
+        self._model = model
+        self._performance = performance
+        self.history = self._performance.history
+        self.history_model = history.history
+        self.is_fit = True
+        # Return the classifier
+        return self
+    def predict_proba(self, X):
+        """
+        Probability estimates.
+        The returned estimates for all classes are ordered by the
+        label of classes.
+        For a multi_class problem, if multi_class is set to be "multinomial"
+        the softmax function is used to find the predicted probability of
+        each class.
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            Vector to be scored, where `n_samples` is the number of samples and
+            `n_features` is the number of features.
+        Returns
+        -------
+        T : array-like of shape (n_samples, n_classes)
+            Returns the probability of the sample for each class in the model,
+            where classes are ordered as they are in ``self.classes_``.
+        """
+        # Check is fit had been called
+        check_is_fitted(self)
+        # Input validation
+        if  X.ndim != 4:
+            raise ValueError("Found array X with dim %d. %s expected == 4." % (X.ndim, self.name))
+        y_prob = self._model.predict(X, verbose = self.verbose)
+        return y_prob
+    def predict(self, X):
+        probs = self.predict_proba(X)
+        y_pred = pd.get_dummies(np.argmax(probs, axis=1)).values
+        return y_pred
+    def score(self, X, y, scoring = 'accuracy', sample_weight=None):
+        """Returns the score using the `scoring` option on the given
+        test data and labels.
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            Test samples.
+        y : array-like of shape (n_samples,)
+            True labels for X.
+        scoring: str, please refer to: https://scikit-learn.org/stable/modules/model_evaluation.html#scoring-parameter
+        sample_weight : array-like of shape (n_samples,), default=None
+            Sample weights.
+        Returns
+        -------
+        score : float
+            Score of self.predict(X) wrt. y.
+        """
+        assert scoring in SCORERS.keys(), 'scoring is not in %s' % SCORERS.keys()
+        scoring = get_scorer(scoring)
+        return scoring(self, X, y, sample_weight=sample_weight)
+    def plot_model(self, to_file='model.png',
+                   show_shapes=True,
+                   show_layer_names=True,
+                   rankdir='TB',
+                   expand_nested=False,
+                   dpi=96):
+        if self.is_fit:
+            tf.keras.utils.plot_model(self._model,
+                       to_file=to_file,
+                       show_shapes=show_shapes,
+                       show_layer_names=show_layer_names,
+                       rankdir=rankdir,
+                       expand_nested=expand_nested,
+                       dpi=dpi)
+        else:
+            print('Please fit first!')
+    def save_model(self, model_path):
+        return save_model(self, model_path)
+    def load_model(self, model_path, gpuid=None):
+        return load_model(model_path, gpuid=gpuid)
+    def explain_model(self, mp, X, y,
+                      binary_task = False,
+                      explain_format = 'global',
+                      apply_logrithm = False,
+                      apply_smoothing = False,
+                      kernel_size = 3, sigma = 1.2):
+        '''
+        Feature importance calculation
+        Parameters
+        --------------
+        mp: aggmap object
+        X: trianing or test set X arrays
+        y: trianing or test set y arrays
+        binary_task: {True, False}
+            whether the task is binary, if True, the feature importance will be calculated for one class only
+        explain_format: {'local', 'global'}, default: 'global'
+            local or global feature importance, if local, then X must be one sample
+        apply_logrithm: {True, False}, default: False
+            whether apply a logarithm transformation on the importance values
+        apply_smoothing: {True, False}, default: False
+            whether apply a smoothing transformation on the importance values
+        kernel_size: odd number, the kernel size to perform the smoothing
+        sigma: float, sigma for gaussian smoothing
+        Returns
+        ------------
+        DataFrame of feature importance
+        '''
+        if explain_format == 'global':
+            explain_func = GlobalIMP
+        else:
+            explain_func = LocalIMP
+        dfe = explain_func(self, mp, X, y,
+                           binary_task = binary_task,
+                           task_type = 'classification',
+                           sigmoidy = False,
+                           apply_logrithm = apply_logrithm,
+                           apply_smoothing = apply_smoothing,
+                           kernel_size = kernel_size, sigma = sigma)
+        return dfe
+    @property
+    def clean(self):
+        clean(self)
+class MultiLabelEstimator(BaseEstimator, ClassifierMixin):
+    """ An AggMap CNN MultiLabel estimator (each sample belongs to only one class)
+    Parameters
+    ----------
+    epochs : int, default = 200
+        A parameter used for training epochs.
+    conv1_kernel_size: int, default = 13
+        A parameter used for the kernel size of first covolutional layers。
+    dense_layers: list, default = [128]
+        A parameter used for the dense layers.
+    batch_size: int, default: 128
+        A parameter used for the batch size.
+    lr: float, default: 1e-4
+        A parameter used for the learning rate.
+    loss: string or function, default: tf.nn.sigmoid_cross_entropy_with_logits。
+        A parameter used for the loss function
+    batch_norm: bool, default: False
+        batch normalization after convolution layers.
+    n_inception: int, default:2
+        Number of the inception layers.
+    dense_avf: str, default is 'relu'
+        activation fuction in the dense layers.
+    dropout: float, default: 0
+        A parameter used for the dropout of the dense layers, such as 0.1, 0.3, 0.5.
+    monitor: str, default: 'val_loss'
+        {'val_loss', 'val_metric'}, a monitor for model selection。
+    metric: str, default: 'ROC'
+        {'ROC', 'ACC', 'PRC'},  a matric parameter。
+    patience: int, default: 10000
+        A parameter used for early stopping。
+    gpuid: int, default: 0
+        A parameter used for specific gpu card。
+    verbose: int, default: 0
+        if positive, then the log infomation of AggMapNet will be print，
+        if negative, then the log infomation of orignal model will be print。
+    random_state: int, default: 32
+        Random seed
+    name: str
+        Model name
+    Examples
+    --------
+    >>> from aggmap import AggModel
+    >>> clf = AggModel.MultiLabelEstimator()
+    """
+    def __init__(self,
+                 epochs = 200,
+                 conv1_kernel_size = 13,
+                 dense_layers = [128],
+                 dense_avf = 'relu',
+                 batch_size = 128,
+                 lr = 1e-4,
+                 loss = tf.nn.sigmoid_cross_entropy_with_logits,
+                 batch_norm = False,
+                 n_inception = 2,
+                 dropout = 0.0,
+                 monitor = 'val_loss',
+                 metric = 'ROC',
+                 patience = 10000,
+                 verbose = 0,
+                 random_state = 32,
+                 gpuid = 0,
+                ):
+        self.epochs = epochs
+        self.dense_layers = dense_layers
+        self.conv1_kernel_size = conv1_kernel_size
+        self.dense_avf = dense_avf
+        self.batch_size = batch_size
+        self.lr = lr
+        self.loss = loss
+        self.batch_norm = batch_norm
+        self.n_inception = n_inception
+        self.dropout = dropout
+        self.monitor = monitor
+        self.metric = metric
+        self.patience = patience
+        self.gpuid = str(gpuid)
+        os.environ["CUDA_VISIBLE_DEVICES"]= self.gpuid
+        self.verbose = verbose
+        self.random_state = random_state
+        self.is_fit = False
+        self.name = "AggMap MultiLabels Estimator"
+        #print(self.get_params())
+        self.history = {}
+        self.history_model = {}
+        if self.verbose > 0:
+            self.verbose1 = self.verbose
+            self.verbose2 = 0
+        elif self.verbose ==0:
+            self.verbose1 = 0
+            self.verbose2 = 0
+        elif self.verbose < 0:
+            self.verbose1 = 0
+            self.verbose2 = abs(self.verbose)
+        print(self)
+    def get_params(self, deep=True):
+        model_paras =  {"epochs": self.epochs,
+                        "lr":self.lr,
+                        "loss":self.loss,
+                        "conv1_kernel_size": self.conv1_kernel_size,
+                        "dense_layers": self.dense_layers,
+                        "dense_avf":self.dense_avf,
+                        "batch_size":self.batch_size,
+                        "dropout":self.dropout,
+                        "batch_norm":self.batch_norm,
+                        "n_inception":self.n_inception,
+                        "monitor": self.monitor,
+                        "metric":self.metric,
+                        "patience":self.patience,
+                        "random_state":self.random_state,
+                        "verbose":self.verbose,
+                        "name":self.name,
+                        "gpuid": self.gpuid,
+                       }
+        return model_paras
+    def set_params(self, **parameters):
+        for parameter, value in parameters.items():
+            setattr(self, parameter, value)
+        return self
+    def fit(self, X, y,  X_valid = None, y_valid = None):
+        # Check that X and y have correct shape
+        if  X.ndim != 4:
+            raise ValueError("Found array X with dim %d. %s expected == 4." % (X.ndim, self.name))
+        if  y.ndim != 2:
+            raise ValueError("Found array y with dim %d. %s expected == 2." % (y.ndim, self.name))
+        # Store the classes seen during fit
+        self.classes_ = unique_labels(y)
+        self.X_ = X
+        self.y_ = y
+        if (X_valid is None) | (y_valid is None):
+            X_valid = X
+            y_valid = y
+        np.random.seed(self.random_state)
+        tf.compat.v1.set_random_seed(self.random_state)
+        model = aggmodel.net._AggMapNet(X.shape[1:],
+                                        n_outputs = y.shape[-1],
+                                        conv1_kernel_size = self.conv1_kernel_size,
+                                        batch_norm = self.batch_norm,
+                                        n_inception = self.n_inception,
+                                        dense_layers = self.dense_layers,
+                                        dense_avf = self.dense_avf,
+                                        dropout = self.dropout,
+                                        last_avf = None)
+        opt = tf.keras.optimizers.Adam(lr=self.lr, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) #
+        model.compile(optimizer = opt, loss = self.loss)
+        performance = aggmodel.cbks.CLA_EarlyStoppingAndPerformance((X, y),
+                                                                    (X_valid, y_valid),
+                                                                    patience = self.patience,
+                                                                    criteria = self.monitor,
+                                                                    metric = self.metric,
+                                                                    last_avf = None,
+                                                                    verbose = self.verbose1,)
+        history = model.fit(X, y,
+                  batch_size=self.batch_size,
+                  epochs= self.epochs, verbose= self.verbose2, shuffle = True,
+                  validation_data = (X_valid, y_valid),
+                  callbacks=[performance])
+        self._model = model
+        self._performance = performance
+        self.history = self._performance.history
+        self.history_model = history.history
+        self.is_fit = True
+        return self
+    def predict_proba(self, X):
+        """
+        Probability estimates.
+        The returned estimates for all classes are ordered by the
+        label of classes.
+        For a multi_class problem, if multi_class is set to be "multinomial"
+        the softmax function is used to find the predicted probability of
+        each class.
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            Vector to be scored, where `n_samples` is the number of samples and
+            `n_features` is the number of features.
+        Returns
+        -------
+        T : array-like of shape (n_samples, n_classes)
+            Returns the probability of the sample for each class in the model,
+            where classes are ordered as they are in ``self.classes_``.
+        """
+        # Check is fit had been called
+        check_is_fitted(self)
+        # Input validation
+        if  X.ndim != 4:
+            raise ValueError("Found array X with dim %d. %s expected == 4." % (X.ndim, self.name))
+        y_prob = self._performance.sigmoid(self._model.predict(X, verbose = self.verbose))
+        return y_prob
+    def predict(self, X):
+        y_pred = np.round(self.predict_proba(X))
+        return y_pred
+    def score(self, X, y, scoring = 'accuracy', sample_weight=None):
+        """Returns the score using the `scoring` option on the given
+        test data and labels.
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            Test samples.
+        y : array-like of shape (n_samples,)
+            True labels for X.
+        scoring: str, please refer to: https://scikit-learn.org/stable/modules/model_evaluation.html#scoring-parameter
+        sample_weight : array-like of shape (n_samples,), default=None
+            Sample weights.
+        Returns
+        -------
+        score : float
+            Score of self.predict(X) wrt. y.
+        """
+        assert scoring in SCORERS.keys(), 'scoring is not in %s' % SCORERS.keys()
+        scoring = get_scorer(scoring)
+        return scoring(self, X, y, sample_weight=sample_weight)
+    def plot_model(self, to_file='model.png',
+                   show_shapes=True,
+                   show_layer_names=True,
+                   rankdir='TB',
+                   expand_nested=False,
+                   dpi=96):
+        if self.is_fit:
+            tf.keras.utils.plot_model(self._model,
+                       to_file=to_file,
+                       show_shapes=show_shapes,
+                       show_layer_names=show_layer_names,
+                       rankdir=rankdir,
+                       expand_nested=expand_nested,
+                       dpi=dpi)
+        else:
+            print('Please fit first!')
+    def save_model(self, model_path):
+        return save_model(self, model_path)
+    def load_model(self, model_path, gpuid=None):
+        return load_model(model_path, gpuid=gpuid)
+    def explain_model(self, mp,
+                      X,
+                      y,
+                      explain_format = 'global',
+                      apply_logrithm = False,
+                      apply_smoothing = False,
+                      kernel_size = 3, sigma = 1.2):
+        '''
+        Feature importance calculation.
+        Parameters
+        --------------
+        mp: aggmap object
+        X: trianing or test set X arrays
+        y: trianing or test set y arrays
+            whether the task is binary, if True, the feature importance will be calculated for one class only
+        explain_format: {'local', 'global'}, default: 'global'
+            local or global feature importance, if local, then X must be one sample.
+        apply_logrithm: {True, False}, default: False.
+            whether apply a logarithm transformation on the importance values.
+        apply_smoothing: {True, False}, default: False.
+            whether apply a smoothing transformation on the importance values.
+        kernel_size: odd number, the kernel size to perform the smoothing.
+        sigma: float, sigma for gaussian smoothing.
+        Returns
+        ------------
+        DataFrame of feature importance
+        '''
+        if explain_format == 'global':
+            explain_func = GlobalIMP
+        else:
+            explain_func = LocalIMP
+        dfe = explain_func(self, mp, X, y,
+                           task_type = 'classification',
+                           binary_task = False,
+                           sigmoidy = True,
+                           apply_logrithm = apply_logrithm,
+                           apply_smoothing = apply_smoothing,
+                           kernel_size = kernel_size, sigma = sigma)
+        return dfe
+    @property
+    def clean(self):
+        clean(self)